What's new¶
Product updates to Anychat, newest first. Each entry links to the guide or reference page with the full details.
October 8, 2026 — Viewers can try agents in the emulator¶
A workspace member with the viewer role can now use the whole emulator: chat with an agent as one of its personas, start an Outreach to a persona in Outbound mode, and reset a persona to start over. All of it stays in the sandbox, so a viewer still can't send anything to a real person. Creating and editing personas still needs editor, so a viewer uses the personas an editor has set up.
October 6, 2026 — Tests read carousels card by card¶
A test's transcript now prints each card's chips under that card, so the judge can tell which card offered which action. Before, every chip in a carousel was listed in one line, and a case could fail when one card offered a button the others did not, such as a boat bookable at two marinas. The message's own chips are a separate suggested replies line, question options and time-picker slots are listed, and conversation reviews and persona transcripts use the same layout.
October 6, 2026 — Messages go to the number you name¶
A Person is now found only through a channel identity: a channel and the address that channel uses, such as the person's number on RCS. A send goes to the identity you name, never to a phone number typed on a profile, so a shared email address or a mistyped number can't send one person's message to someone else. Phone, email and names on a profile are information about the person, and two people may share them.
An identity keeps the same id on every provider and binding of its channel, so moving RCS to another provider keeps every conversation and every opt-out. A STOP covers all of that agent's senders on the channel, and never another channel. On a person's record, the new Channel identities card adds, moves and removes them, and Similar people suggests records to merge.
In the API, a send names its recipient as a channel and address, an
ASID, a Person or a persona, with a separate profile and
profileUpdate. People have identity routes and similar, imports take
channel and address columns, and the Adobe Journey Optimizer send
action needs profile.phone in E.164. externalId, one-person-per-phone
and duplicate flags are gone.
October 2, 2026 — Find the ids the API asks for¶
An agent's new Settings · API page lists its organization id, its agent id and the id of every Outreach, each with a copy button. An Outreach's page and a persona's page now show their own id too. These are the ids an Adobe Journey Optimizer action, or any call to the API, needs; before, they were visible only in the address bar.
Send from Adobe Journey Optimizer
October 2, 2026 — Choose who answers STOP, and word the replies yourself¶
When your messaging provider already answers STOP, START and HELP, as Vonage does with Opt-Out Assist, you can now say so on the channel's page. Anychat then stays quiet on those words, so people get one reply instead of two, and still sends your first-message notice. Every reply Anychat sends can now be reworded, with a preview filled in from your agent's details, and the notice has its own on/off switch.
Compliance messages are now on by default for new RCS and SMS channels, and JOIN opts someone back in like START.
October 2, 2026 — REVOKE and OPT OUT now opt people out¶
On every RCS and SMS channel, a person who replies REVOKE or OPT OUT is now opted out exactly as if they had replied STOP. These are the words the FCC's opt-out rule names alongside STOP, so a person who uses them is no longer treated as just chatting. A keyword with a trailing period or exclamation mark, like "Stop.", now counts too.
Messages the agent never sees, like a HELP that Anychat answers or a message from someone who has opted out, now appear in the conversation's transcript.
September 30, 2026 — See when a channel's messages aren't getting through¶
If your provider signs its webhooks with a different secret than the one saved on your channel, Anychat has to reject them, and your agent never sees the messages. The channel's page now says so in a red banner, names the credential to fix, and opens the form to fix it. It also shows when a webhook was last accepted, and before launch, which number was turned away because it isn't on the early-access list.
When messages don't reach the agent
September 30, 2026 — Early access keeps Vonage's test-device list¶
On a Vonage RCS channel, adding a number to early access now also adds it to the agent's test-device list at Vonage, and removing it takes it off. You no longer have to keep two lists in step. The Invite column shows when each person has accepted the tester invitation. On a provider whose list Anychat can't reach, like Webex Connect, the console marks each number as one to add in the provider's dashboard too.
RCS and the provider's tester list
September 28, 2026 — Map buttons and locations on Things¶
A Thing can now carry its location, and its card can offer a Map button that opens the place in the customer's maps app. Type an address or paste a Google Maps link, and Anychat pins it when you save, or fill it from your business address in one click. The agent also knows where each Thing is, so "where is the marina?" gets a real answer.
September 28, 2026 — Call and reply buttons on Thing cards¶
The buttons on a Thing's card can now do more than open a link. Add a Call button that dials the business, or a Reply button such as Book this that answers the agent right from the carousel. The agent knows which card the reply came from, so every card can carry the same button. On a channel that can't show buttons, a call shows as the number instead.
September 28, 2026 — AJO actions generated for each Outreach¶
Setting up an Adobe Journey Optimizer action no longer means copying payloads by hand. Anychat now generates everything to enter in AJO for an Outreach: the action names and URLs, the payloads to paste with each field marked constant or variable, where to map it from, and the setup and branching steps. It is generated from the Outreach, so it always matches the inputs the Outreach needs.
Send from Adobe Journey Optimizer
September 28, 2026 — People counts in one call¶
Ask the Copilot (or the API) for your People numbers and get them in one step, however large your book has grown: how many people outreach reached, how many were not reachable, had no consent, or declined, how many replied, and how many have a call on the books. The same filters and segments as the People list narrow the counts, and because the answer holds only numbers, anyone who can view People can pull it.
September 25, 2026 — Put a Thing in several collections¶
A Thing can now belong to more than one collection. Put a jet-ski rental in both Watercraft rentals and Wahweap Marina watercraft rentals, and the agent can show every rental or just the ones at the marina a visitor is near, from one Thing instead of two copies to keep in step. It makes location-specific recommendations much easier to set up.
A Thing in several collections
September 24, 2026 — Send from Adobe Journey Optimizer¶
An Adobe Journey Optimizer journey can now send an Outreach at any step and branch on what happened. Two custom actions do it: one sends, and one reads back whether the message went out and whether the person can receive RCS. Opt-outs and missing consent come back as outcomes the journey can branch on, not errors, and AJO's own retries never send twice.
Send from Adobe Journey Optimizer
September 24, 2026 — Choose an API key's scopes in the console¶
Creating an API key on Workspace · API keys now asks which scopes it carries, so each integration gets a key limited to what it does, and you can change a key's scopes later without replacing it. The new Adobe Journey Optimizer scope makes a key that can send and check outreach from AJO and do nothing else.
September 24, 2026 — Rate limits on sends through the API¶
Sends made with an API key now have a published limit: 600 a minute per
key, in bursts of up to 100. A journey platform or booking system that
goes faster gets 429 with a Retry-After header telling it how long to
wait, and nothing is sent until it retries. Sends from the console aren't
limited.
September 24, 2026 — Outreach never interrupts a conversation¶
Automated Outreach and Campaigns no longer send into the middle of an exchange. If someone wrote to your agent in the last 30 minutes, their Automated Outreach or Campaign message waits until 30 minutes after their last message, then goes out. Automated Outreach runs show these rows as waiting mid-conversation. A send you make by hand and an appointment reminder still go when they are due.
Sending guardrails · Run a campaign
September 24, 2026 — STOP stops every message on RCS and SMS; reminders and safe hours¶
A person who replies STOP on an RCS or SMS channel now receives nothing more from that agent until they reply START, with or without carrier compliance turned on. Before, a channel without carrier compliance did not record the STOP, so replies, follow-ups and appointment reminders could still reach the person. Carrier compliance now only adds the confirmation replies, the HELP answer and the disclaimer.
Appointment reminders get their own switch for safe hours. By default a reminder goes when it is due, whatever your safe hours, because a reminder just before a call is only useful then. Turn off Send reminders outside safe hours to hold a reminder until safe hours open; one that would land after the call has started is not sent.
Sending guardrails · Reminders · Carrier compliance
September 24, 2026 — Sends you make by hand or through the API follow your judgment¶
When you name the person, the send is yours. A send you make by hand,
through the API, or as a follow-up you send now goes out when you make
it, to anyone you have consent for, whether or not that consent is on
file in Anychat. It is refused only for an explicit no: your own
request saying the person did not opt in (consent_declined), a STOP
from the person (opted_out), or consent recorded on the Person as
denied (consent_denied). Consent you assert with the send is recorded
on the send as its evidence and never changes the person's record. Safe
hours don't delay it, and the frequency cap warns instead of blocking.
Sends Anychat makes on your behalf keep to your rules. Automated Outreach and Campaigns reach only people whose consent is granted, a scheduled follow-up stops once consent is recorded as denied, and all three wait for your safe hours. Safe hours now govern Automated Outreach and Campaigns as well as follow-ups. An Automated Outreach that answers leads the moment they sign up can be set to Send outside safe hours; consent still applies. Safe hours are the hours in which sending is allowed, read in the time zone you set at Settings · Sending.
Sending guardrails · Send to one person · Sending API
September 24, 2026 — Personas you can see, shape and reset¶
A persona is a Person record that lives in the sandbox, with a starting point. Test → Personas now lists them, and each opens on the same page as a Person. Edit its starting point with the People editors, including upcoming appointments and earlier messages, and saving puts the persona there at once.
Talk to one and watch it change: the emulator's new Record tab marks everything the agent changed since the starting point, and Captured keeps the whole trail. Reset goes back, appointments included, and the emulator now shows a persona's earlier messages and restores its conversation in any browser. When a customer conversation goes wrong, Make a persona from here on the message copies the person and the conversation up to that point so you can take it from there.
Manage personas · Personas API
September 23, 2026 — See which agents use a connection before you disconnect it¶
Workspace · Connections now lists the agents that use each account.
Disconnecting a Google account that agents mirror booked calls into
names those agents in the confirmation. Their bookings keep working but
stop reaching the calendar, and each one's Scheduling card shows
Needs attention until you choose a connected account again. The API
returns the same list as usedBy on each connection.
Turn on capabilities · Connections API
September 22, 2026 — Emulator links pick a persona, and the sandbox keeps its hands off your calendar¶
Someone opening a standalone emulator link now picks one of the agent's personas, or makes one by name, before talking — so two reviewers holding the same link no longer share a conversation, and what each of them said shows up under Sandbox conversations with a persona's name on it.
Three sandbox fixes ship with it. A persona's booking no longer takes a real slot: availability offered to your customers never counted emulator or Test bookings from today, and a sandbox conversation can only touch its own persona's appointments. What the agent writes to a persona's record during a sandbox conversation is listed in the captured-effects panel beside the booking, notification and email it would have made, so the panel is now the complete list of what the turn did. And the Testing concept page has been rewritten around the two worlds, Sandbox and Live, and the one rule for what is real and what is captured.
September 22, 2026 — Outreach, Automated Outreach and Campaigns side by side; Sends is one list¶
Everything that sends now lives in one place. Build · Outreach has three tabs, Outreach, Automated Outreach and Campaigns, and each one's controls sit on its own page: start or pause a Campaign, run or dry-run a rule, right where you set it up. Monitor · Sends is a single list of every send, filtered by Outreach, origin, Campaign, Automated Outreach, state and period. Filter it to one Campaign to see its progress and why rows did not send, or to one Automated Outreach to see its runs and whether it is keeping up.
See Monitor your agent and Outreach.
September 22, 2026 — Design messages and see them as you type¶
Message Templates used to be a text box of markup. To see the result you had to save the template and set a Preview Person first.
Templates now open in a composer. You add cards, carousels,
pictures and quick replies from a bar under the message. Typing {
inserts a field such as the person's first name as a pill, and clicking
the pill sets what happens when that field is empty. A preview beside
the message updates as you type, for a sample person or for a person
with no details on record, so you can check every fallback. The
templates list shows each template as the message it sends. The same
composer is used for an Outreach's opener on Build · Outreach.
Templates are stored exactly as before, and Source still shows and edits the markup.
September 22, 2026 — Write the whole Outreach in one place¶
Creating an Outreach used to send you somewhere else first: to Behavior · Message Templates for the opening message, to Behavior · Goals for the goal, and back to the Outreach afterwards to write the playbook that the create step had left blank.
Build · Outreach now authors all of it. The opening-message picker offers Write a new message…, which opens the composer in place; the goal picker offers Define a new Goal…; the playbook is written on the same screen as the rest, including while the Outreach is still being created. One Create Outreach saves the lot.
Nothing moved: a message written here is an ordinary Message Template on the agent, a goal an ordinary entry in its goal library, and both stay editable on Behavior and reusable anywhere. Choosing an existing one still works, and Edit opens it without leaving the page.
See Write an Outreach.
September 22, 2026 — One Automated Outreach per lead source¶
An agent used to watch one sheet. If leads arrived from paid social in one place and the website form in another, you had to merge them upstream and accept a single filter, consent column and Outreach for both.
Automated Outreach is now a named collection: the Automated Outreach tab of Build · Outreach holds one rule per lead source, each with its own sheet tab, filter, consent column, Outreach and hourly send cap. Each keeps its own run history, and Monitor · Sends filtered to a rule shows whether it is keeping up.
Nothing to do on an existing agent: what you had is now a rule called "Automated Outreach", running exactly as before.
See Automate outreach from a sheet.
September 22, 2026 — The agent menu is Monitor, Build, Test, Settings¶
The Run tier is now Monitor, and Overview sits inside it as the agent's home once it is built. Four labels changed with it, each to the word the documentation already used: Test · Links is Test links, Settings · Access is Test · Emulator links (so "access" means who on your team can do what, and nothing else), Settings · Concepts is Concept Studio, and Capabilities · Integrations is Custom integrations. Old links redirect.
See Monitor your agent for the daily view and Manage access for workspace roles.
September 22, 2026 — Campaigns send an Outreach¶
A campaign now picks one Outreach and sends it to the list you choose. The opener, the Playbook, the goal, the class and the follow-up all come from the Outreach, so a campaign has nothing of its own to author. Every recipient is an ordinary send you can read under Monitor · Sends, with a campaign filter.
Also in the Agent API.
September 22, 2026 — Automated Outreach sends an Outreach, with a shadow run first¶
Pick the Outreach your linked sheet sends. Before you switch, a Shadow run decides every row under the new Outreach alongside the old one, reports how many agree each run, and sends nothing itself. Automated Outreach never sends the same Outreach to a person twice.
Also in the Agent API and
the Anychat MCP server
(replace_automated_outreach_config).
September 22, 2026 — One reusable reason to reach out¶
Outreach is now the one place to write a message that starts a conversation. Under Build · Outreach, write it once: an opener from your Message Templates, its own playbook for what happens after the person replies, a goal, a class, and, if it needs one, a follow-up and the inputs the opener requires. Send it to someone from their record on the People page and see the message rendered for that person before it goes.
Every send is one row under Monitor · Sends. A send held for safe hours or the frequency cap shows as queued, and you can cancel it from there before it goes out. The frequency cap and safe hours themselves now live in one place, Settings · Sending.
Also in the Agent API and the
Anychat MCP server (list_outreach, create_outreach,
send_outreach_to_person, cancel_outreach_send, list_outreach_sends).
September 22, 2026 — Messages are Message Templates again¶
The Behavior section and every label in the console now say Message Templates. Nothing about the templates themselves has changed.
September 22, 2026 — Test links: preview an opening before anyone taps it¶
A test link is a shareable URL under Test · Links: a tester enters their own number and the real agent opens the conversation on their phone, exactly as it would for a new contact. Redeem the same link in the emulator as a persona to see the opening it sends before sharing it with anyone.
September 21, 2026 — A default card image, so a card is never dropped to plain text¶
Some channels refuse to show a card without a picture. Until now, a card you wrote without one arrived as a couple of lines of text — and on RCS, where only the newest message's buttons stay on screen, its link could scroll out of reach entirely.
Set a default card image on Behavior · Identity and those cards render as cards, using your picture wherever one is missing. It is a field of its own, separate from your logo and banner — a room photo belongs on a card where a wordmark does not. Leave it empty and the card still renders, on a plain grey panel.
Cards you already give a picture to are untouched.
September 19, 2026 — The Emulator: talk to your agent as anyone, from anywhere in the console¶
The phone in the console is now the Emulator, and it has a cast. Pick a persona to talk as, pick the channel look (RCS, Apple Messages, WhatsApp or SMS), and see what a handset would see. A Captured panel beside the phone lists everything the agent did instead of doing for real: the booking it would have written, the notification it would have sent, the integration it would have called.
You no longer have to start by typing. Start from an Outreach, a program trigger or an outreach link, and the real opening is sent to your persona and appears in the phone, ready to reply to. It is the fastest way to see what an Outreach will actually say before you send a real audience.
Everything the phone does is kept, under Test → Sandbox conversations, where it reads and replays like any other conversation, and counts toward nothing.
September 19, 2026 — Let a colleague try the agent without a console account¶
A client, a reviewer or a copywriter can now talk to one of your agents through a standalone emulator link. They open the link, pick a channel look, and type. It is the same sandbox, so nothing they do reaches a real customer and nothing they say lands in your statistics.
Create links under Settings → Access: give one a label, set how long it lasts (14 days by default, 90 at most), and copy the url once. Revoke kills it immediately for everyone holding it.
Share a standalone emulator link
September 19, 2026 — Your Playbook's Examples are Tests, and any conversation is one click from being one¶
The examples you write on a Playbook to show it how a reply should look now run as Tests. Each Playbook with Examples gets a suite of its own, kept in step with the Examples as you edit them, and the Playbooks page shows that Playbook's pass rate beside its name (Tests: 3 of 4 passed) so you can see at a glance which parts of the agent are holding.
When a real conversation goes wrong, you no longer retype it. Open the turn inspector on the turn that went wrong and choose Make a Test: the person's messages become the script and the checks start from what the agent actually did, ready for you to edit into what should have happened. It works on live and sandbox conversations alike.
Suites can also run on their own. Set Run on a schedule to every day or every week, with the time and time zone, and a failing run tells your team through the agent's notifications.
September 19, 2026 — Real phones before the public: early access and launch¶
Every channel is now connected first and launched second. Until you launch it, only the phone numbers on its early-access list reach the agent there, and everyone else's messages are dropped. Your own staff and the people who have to approve the work use the real agent, on a real phone, on the real channel, while the rest of the world still cannot find it.
Add numbers on the channel's detail page, launch when you are ready, and pause later with Disable if you need to. On RCS that Anychat runs for you, each number is also invited through Google's tester program and the list shows who has accepted.
September 19, 2026 — Test a channel's connection¶
A channel can look perfectly healthy and still be unable to send or receive: the credentials are stored, the channel reads Active, and nothing tells you otherwise until a real message fails to arrive. The emulator can't help, because it uses a simulated transport and never touches your provider.
Test connection, on a channel's page under Channels, closes that gap. It asks your provider to prove the credentials work, and sends nothing to anyone. You get a short list of checks — are the stored credentials well-formed, does the provider accept them, and (if you give a phone number) can that handset actually receive messages from this agent — each one passed, failed, or skipped, with the provider's own words when something is wrong.
Worth running right after you connect a channel, and first whenever an agent has gone quiet on one. See Channels.
September 19, 2026 — Personas, and a sandbox where nothing is ever sent¶
Every agent now has personas: test people with a saved starting point — name, details, messaging consent, even prior conversation — that the agent treats exactly like a real customer. Three come ready (a first-time visitor, a returning guest with a booking on Friday, a lead who opted out); add your own, or make one from a real conversation to reproduce a problem and pick it up from that point. Reset a persona and it starts over from its preset.
Everything a persona does happens in the sandbox: the agent reads your real calendar, knowledge and people, but every booking, notification, email, integration call and outreach send is captured and shown back to you instead of performed. Tests run each case against a fresh copy of a persona, and any outbound send — a per-person send, an Outreach, a follow-up — can be pointed at a persona to rehearse it end to end.
The same door is open to scripts and assistants: the sandbox turn API
(and the matching @anychat-ai/mcp tools) sends a persona a message and
returns the reply plus everything the agent captured.
September 18, 2026 — Lead sheets are picked up in minutes, not hours¶
A connected lead sheet is now monitored for changes instead of checking it once an hour. Add a lead and the agent usually reaches them within about ten minutes. It also waits for your edits to settle, so a sheet you're still filling in — a pasted block of leads, a row typed a field at a time — is picked up once you've finished, never halfway through.
September 18, 2026 — Behavior replaces Definition, and Media joins Knowledge¶
The Definition page is now Behavior: the same document, under the name that says what it is for. What the agent knows is Knowledge; how it acts is Behavior.
Two things moved with it. Media is now part of Knowledge, beside Sources, Facts, Things and Business info, because the pictures your agent may show are things it knows rather than words you wrote. And outbound setup (the opener, the follow-up, and Automated Outreach from a Google Sheet) now sits under Outreach instead of under Capabilities, next to the sending you already do there.
Nothing about your agents changed, and every old link redirects. See Agent Behavior and Knowledge.
September 18, 2026 — Let another system start the conversation¶
Your agent can now be told to send a specific message to a specific person by an outside system — a journey platform, a CRM, your own scheduler. You name the programs it is allowed to send (each one a playbook plus the message that opens it), and nothing outside that list can be triggered. Consent is checked every time, and a person who has opted out stays opted out no matter what the calling system believes.
September 18, 2026 — See what happened to every message you send¶
Every outreach send now keeps a record: whether it went out, whether the phone received it, whether it was read — and when it didn't go out, the reason in plain terms (no phone on file, not RCS-capable, no active channel). Send one from the People page or the API and you get a send id you can look up any time in the next 180 days.
September 17, 2026 — Your agent reaches every capability you've set up¶
A Playbook's Needs are what the agent has at hand while it's active, and the agent also knows every other capability you've set up. When a guest in a question-answering Playbook asks to book, the agent loads Scheduling and offers times. Large Things catalogs work the same way: the agent sees the Things that matter to the conversation, and a search finds the rest.
September 17, 2026 — Your agent remembers its follow-ups¶
Every follow-up your agent sends is now part of the conversation. When someone answers a follow-up, the agent knows what it asked and replies in context. This used to be a per-agent setting and is now how every agent works; nothing needs to be turned on.
Following up with people who don't reply
September 16, 2026 — Keep the Playground open while you work¶
The Playground now docks beside any page in the console. Press ⌘⇧E to open it, then keep talking to your agent while you edit Playbooks, Knowledge or settings. It remembers the conversation, agent, channel and persona when you close it, move to another page, or reload the console. It works alongside Copilot, and it can pop out into its own window that you can place anywhere.
September 16, 2026 — Describe how a reply should look, in a Playbook¶
A Playbook can now say how a reply should look using the same markup you
write in a Message. [[chips: Boat rentals | Dining | Activities]] in a
Playbook body is read as a description of the turn, and the agent renders
the chips itself and adapts them to each channel. Plain English works just
as well, and you no longer need a Message for chips alone. Definition lint
no longer reports markup in a Playbook, with one exception: a time picker
written into instructions, whose slot ids only exist once the agent checks
availability.
September 15, 2026 — Show a whole collection as one carousel¶
When your Things are grouped into collections, your agent can now send an entire collection, such as every rental boat, as a single carousel. It names the collection instead of listing each Thing, so cards reliably appear when someone asks what is available, and Things you add later show up without editing anything. There is nothing to turn on: grouping Things into a collection is enough.
September 15, 2026 — Objectives retired in favor of Goals¶
Goals are now the one way to say what a conversation is working toward: a single outcome, assigned per Playbook or Outreach with one default, that Anychat can score. Objectives are gone. Agents that already had a default Goal work exactly as before; for every other agent, its objectives text was added to the end of its default Playbook's instructions, so it reads the same guidance. The Sales Development, Franchise Development, Med Spa Concierge, Local Business and Vacation Resort templates now come with Goals.
September 15, 2026 — Additional rules move into Standing instructions¶
Guardrails are now just the three hard-limit lists: Topics to avoid, Claims never to make, and Required disclaimers. The free-form Additional rules box is gone; any text you had there was moved into Standing instructions automatically, one instruction per line. Scheduling also warns you when it is set up without available hours, so an agent that can book a call at any hour is something you chose.
Guardrails · Scheduling availability
September 15, 2026 — Procedures merged into Playbook instructions¶
A Playbook's procedures are now part of its instructions, so everything the agent reads while a Playbook is active is edited in one place. Your agents behave exactly as before — each procedure's text was added to the end of its Playbook's body, word for word. Scheduling also gains a Reply playbook setting: choose which Playbook handles a reply to a pre-call reminder when there is no conversation in progress.
Playbooks · Scheduling reminders
September 15, 2026 — Messages can carry live Things¶
A Message body can now pull cards straight from your Things catalog instead of hand-building a carousel: name one or more Things by id or name, or embed a whole collection, and the agent sends the current picture, description, and links every time.
September 15, 2026 — Needs are explicit¶
Every Playbook now lists Needs: exactly which Capabilities it acts through (Scheduling, People, Notifications, custom integrations), chosen from what the agent has. Knowledge is available in every Playbook. Nothing changed for an existing Playbook — each one was set to the Capabilities it already used — but from here on, what a Playbook can reach for is a choice you make, not an implicit "everything the agent has."
September 15, 2026 — Standing instructions¶
A new addition sits alongside Personality and Guardrails. Standing instructions are short house rules the agent follows on every turn, for the things that don't fit either — like never using bullet lists.
September 15, 2026 — Goals: one outcome per conversation¶
Objectives rank several aims at once; a Goal names a single outcome for a conversation to work toward, and Anychat scores it automatically. Set one on a Playbook, on an Outreach, or as the agent's default, with an event to watch for (a call booked, contact info captured, a reply, a link tapped) or a criterion a reviewer judges at the end. Assign one and the agent works toward it in place of its Objectives.
September 15, 2026 — Rules the agent can be judged against¶
Some rules aren't a word list. "Never promise a delivery date" or "never discuss a competitor's pricing" depend on what the reply means, not what it says word for word. Set a Rule's mode to Judged and a fast model reads the reply against your sentence before it sends, catching what patterns miss, with roughly a second added to a reply that trips it and no conversation ever blocked by a slow or failed check.
September 15, 2026 — Messages the agent never volunteers¶
Every message template still shows up in the agent's index, but now you can mark one as only when named: set it once and the agent sends that template only when a Playbook calls for it by id or name, never on its own judgment. Use it for a consent notice, a legal disclaimer, or an Outreach opener that should go out only where you've placed it.
September 15, 2026 — Rules you can rely on¶
Rules give you an authoring surface for the output checks Anychat enforces against a reply, not just what the agent is instructed to do. Write a Rule with the phrases or patterns to catch, choose whether a violation gets rewritten once and falls back or is blocked outright, and see it listed alongside the platform's own rules, like opt-out and quiet hours, so you can see everything that governs the agent in one place.
September 15, 2026 — Name a template and the agent knows it¶
Mention a message template's id or name anywhere in a Playbook and the agent picks it up automatically. Every turn already carries a short index of every template on the agent, so there's nothing else to wire up before the agent can send it.
September 14, 2026 — Export documents say facts, things and exact¶
An agent's export document now writes capabilities.knowledge.facts
and capabilities.knowledge.things, matching the console's own names
for Facts and Things, and a Fact's word-for-word setting is now called
exact. Files exported before this change, and anything written
against the older faqs, products, or verbatim spellings, still
import without changes.
September 14, 2026 — Refresh a Source; Sources over the API¶
A URL Source can now be refreshed on demand, re-fetching and re-indexing the page in place instead of deleting and re-adding it. Sources can also be added, refreshed, and removed through the API and MCP server, not only the console — useful for pasting text straight in or keeping a Source in sync from your own tooling. Upload stays a console-only path.
September 14, 2026 — Facts you can pin and scope¶
Two new controls on each Fact. Always include keeps a Fact in every conversation no matter how long the list grows or how the question is ranked — the hours, the address, the policy that must never fall out of view. Only in Playbooks ties a Fact to the situations where it applies, so an answer that only makes sense during booking doesn't surface while the agent is doing something else. Every Fact you save today is confirmed and in use, as before.
September 14, 2026 — Things and collections: a catalog with more to say¶
Your catalog can now hold more than a name, a picture, and a link. A Thing (the console's former Products entry) can carry up to four labelled action links — Book a table, View menu, See availability — instead of one generic "View details" button. New collections group related Things under a heading, such as Dining or Rooms, with their own picture, name, description, and action fields per card, so the agent shows the right group instead of your whole catalog. Existing entries keep working exactly as they did; none of this is required to use Things the way you already do.
September 14, 2026 — Knowledge now speaks of Sources and Facts¶
The two halves of an agent's knowledge have names that say what they are. Sources are the material the agent reads from: uploaded files and imported website pages. Facts are questions with the answer you have confirmed; the agent treats each one as the answer, in any wording, and a Fact marked Exact goes out word for word. Nothing about how the agent answers has changed, and your content is exactly where it was; the console pages, the API, and the MCP tools now use the same words, so what you see in the Knowledge area matches what you ask the Copilot for.
If you call the API or use the MCP server: the routes moved to
…/knowledge/sources and …/knowledge/facts, and the tools are
get_facts, replace_facts, and list_knowledge_sources. The old
paths are gone.
Guide: Knowledge. Reference: Facts and Knowledge sources.
September 12, 2026 — Rich formatting in FAQ answers, and templates built from templates¶
An FAQ answer can now carry the same rich formatting a message template does: quick-reply chips, a card, a carousel, a picture from your media library. A verified answer is no longer limited to prose, so "where are you?" can come back as a card with a map and a Get directions button.
Message templates can also pull in another template by id, so copy you reuse in several places lives in one template and updates everywhere at once.
The console also checks your ids as you write. A [[template: …]] or
image:… reference that names nothing is flagged in Definition lint,
wherever you wrote it — a template, a Playbook body, a procedure, or an
FAQ answer.
Guides: Knowledge, Message Templates, AML.
September 12, 2026 — Connect an account once, not once per feature¶
A connected account is now a connection: you link Google or Slack once for your business, and any capability that needs it points at that connection. Before, a grant belonged to a single tool, so the same Google account had to be connected separately for each feature that touched it, and disconnecting in one place left the other quietly broken.
Connections live in account settings with their real identity on show — "Google, connected as ops@yourcompany.com" — plus whether each one is healthy, expiring, or needs reconnecting. Scheduling names the connection it mirrors booked calls into, which also means a calendar can no longer be marked connected when nothing is actually linked.
If your business needs two accounts of the same kind, connect both and name them; capabilities choose between them.
Guide: Capabilities in the console.
September 12, 2026 — A custom integration is one thing again¶
Defining a custom integration used to take two screens: you described the ability on one tab and typed the URL and credential on another, and nothing checked that the two halves agreed. Now it is one record on Capabilities · Integrations — what the agent calls it, when to use it, the inputs it gathers, and the endpoint behind it.
Two things follow from that. An exported agent now carries the endpoint, so an integration survives a move between agents instead of arriving half-defined. And deleting one removes its credential with it, rather than leaving a stored secret behind for you to remember.
The authentication header stays write-only, as before: stored encrypted, shown as set, never echoed back.
Guide: Capabilities in the console.
September 12, 2026 — Notifications is a capability you configure¶
The agent can tell your team when something needs attention, and now there is a page where you say where those alerts land: Capabilities · Notifications. Pick Slack or email, set the sender the agent uses for any email it sends, and swap the service later without touching a Playbook.
The important half is what happens when you leave it off. Until now the agent was always handed the ability to notify, even with nowhere to send, so the attempt failed silently mid-conversation. Now the capability is off until you configure it: the agent cannot notify anyone, knows it cannot, and will not tell a customer that a teammate has been alerted.
Your webhook URL and API key are write-only, as connection secrets have always been: stored encrypted, shown as set with a Replace action, and never included in an exported agent. Changing the from-address no longer means re-entering the key.
Guide: Capabilities in the console.
September 11, 2026 — Review: have a model read your whole agent¶
Ask for a review and a frontier model reads the whole agent at once, the way a colleague would: the Definition, which capabilities are actually configured, the channels it is bound to, and your content. It reports what the automatic checks cannot see. A Playbook that offers to book while Scheduling is off. Two Playbooks that answer the same question, so routing is a coin flip. A Playbook that contradicts one of your guardrails. An instruction to send a carousel on an agent that only has SMS. A hand-off to a phone number your Identity never set.
Every finding quotes the words it is about and says where to fix them, many come with replacement wording to compare against what you have, and some name a conversation worth adding to an evaluation suite. Dismiss the ones you disagree with and they stay gone. Reviews run only when you ask, from the Definition page, the Copilot, the API, or the MCP server, so nothing is spent while you edit. Available to orgs with the feature enabled.
September 11, 2026 — FAQs: verified answers the agent always uses¶
Write the questions people ask most, with answers you have checked, under Knowledge · FAQs, and the agent answers them your way on every turn: the same question in any wording counts, and every fact, number, and price stays exactly as you wrote it, whatever the agent has learned from documents. Mark an entry Verbatim to have the answer sent word for word (prices, policies, legal or medical text), and add other wordings under Also asked as to sharpen matching, with Suggest wordings to draft them for you. Website import now turns an FAQ page into FAQs instead of a document, lint flags placeholder answers and duplicate questions, and FAQs travel with the agent's export and templates. Usage counts show which FAQs people ask and whether verbatim answers went out as written, an FAQ document left by an earlier import converts to FAQs in one click, and a generated FAQ coverage suite under Test checks every FAQ against the live agent. All of it is in the API and the MCP server too.
September 11, 2026 — Copilot: a resizable panel, and chats that tidy themselves¶
The Copilot panel is no longer a fixed strip. Drag its left edge to give a long diff or a wide table the room it needs, or narrow it to keep more of the page in view; double-click the edge to reset. The message box has a grip to make it taller for longer prompts. Both sizes are remembered in your browser. Chats now clear themselves out: a chat and its files are deleted 7 days after its last turn, so the sessions menu only ever lists what you have used lately. There is no setting to extend this, and you can still delete a chat sooner.
September 10, 2026 — Evaluations: describe tests in English, generate suites¶
You no longer need to write the structure of an evaluation case by
hand. Every case now opens on two text areas, Scenario and
Expected outcome: say who the person is, what the agent knows
about them, what they say, and what must be true afterwards, then click
Update from description and the platform compiles it into the
person fixture, the turns, the checks, and the judge criteria, using
your agent's own Playbooks, representatives, and People fields so every
reference is real. Hand edits under Advanced survive a recompile.
To start a suite, choose Generate cases in the New suite
dialog, describe the situation in a sentence, pick roughly how many,
and the platform writes distinct cases from different angles, checks
each against the agent, and lists them ready to run; Generate more
does the same for an existing suite. The judge now grades your criteria
only; the standard dimension scores and per-suite judge model are
retired. In the Agent API the case document gains brief, and two
routes, cases/generate and cases/compile, are matched by
generate_evaluation_cases and compile_evaluation_case in the
Anychat MCP server 0.19.0.
Evaluations guide · How tests are written · API reference
September 10, 2026 — Evaluations: test your agent like code¶
Write down the conversations your agent must handle well and run them as a suite whenever you change something. A case describes the person and what the agent knows about them, how the conversation opens, what the person says or taps, and what must be true afterwards: exact checks on what went out and what was recorded (a time picker was shown, an appointment was booked, a People field was written, no price was quoted) plus plain-language criteria an AI judge grades with the line it quotes as evidence. Cases run against the real agent through the production pipeline; the person is synthetic, and bookings, notifications, and integration calls are recorded instead of sent. Each run reports its pass rate, which cases are newly failing since the last run, and whether the Definition, the runtime, or the judge changed in between. Find it under Evaluations in the agent's Test tier; the Copilot can write and repair cases for you. Runs are metered in evaluation units against a daily budget per workspace. Also in the Agent API and the Anychat MCP server.
Evaluations guide · How an evaluation runs · API reference
September 9, 2026 — One runtime: the Generic (v1) runtime is retired¶
Every agent now runs on the same framework; the last agents on the
earlier Generic (v1) runtime were moved today. Every agent therefore has
capabilities declared per Playbook, deterministic paths for step-by-step
Playbooks, Policies the platform enforces, and a record of every turn
you can open from a conversation with Inspect turns. The Switch to
v2 control and the Generic (v1) badge are gone from the Overview
page, and the framework-v2 org feature no longer exists. In the
Agent API an agent's framework always
reads v2, and sending it on PATCH changes nothing. The
Anychat MCP server 0.17.0 drops the framework
input on create_agent and update_agent. The documentation now
describes the one runtime; start at How agents work.
September 9, 2026 — The agent menu is organized by what you are doing¶
Every agent's navigation now has three tiers plus Settings. Run is for watching and acting day to day: Conversations, People, Sends, Calls, and Activity. Build is the agent itself: Definition, Capabilities, Knowledge, and Channels. Test is for checking behavior: Playground and Testers. The editing pages that lived under Settings now sit under Definition (Identity, Personality, Guardrails, Objectives, Playbooks, Message templates, Media) and Capabilities (Scheduling with its Representatives tab, People fields, Outreach with its Automated tab, and Integrations for connections and custom integrations). Outreach links moved to the Outreach page and logs to Activity. The chat emulator has left the Overview page for its own Playground page. Settings keeps Model, Concepts, and Export & import. Old links redirect to the new pages.
September 9, 2026 — RCS via your own Vonage account¶
The RCS channel gains a third provider: connect an agent to an RCS agent you already run on Vonage, alongside the existing Google RBM and Webex Connect options. Choose Use your own account → Your own Vonage RCS agent when connecting the channel, paste the Application credentials, and point the Application's webhooks at the receiver URL the console shows you. Text, media, rich cards, carousels, suggested replies and actions, delivery and read receipts, and outreach links all work the same as on the other RCS providers, and SMS fallback and opt-out stay consistent across providers because Anychat handles them, not Vonage. Guide: RCS via Vonage.
September 7, 2026 — Campaigns: message a list of people at once¶
A Campaigns page joins People in every agent. Pick an audience (a saved segment, a filter, or an import), the Playbook the agent follows when people reply, and the opening message; stage to see exactly who is sendable, who is excluded and why, a rendered preview, and how long the send will take — nothing goes out yet. Start sends at the channel's governed rate while the progress bar updates live; pause, resume, or cancel at any time. Every recipient ends in exactly one state, and the report always adds up to the audience you staged. Each person is sent to at most once, consent and opt-out are enforced, and marketing campaigns respect the frequency cap. Campaigns on real RCS channels are enabled: they ride the channel's rate-governed send queue, while the agent's live replies keep their own path.
September 3, 2026 — Marketing messages are spaced out automatically¶
An agent that sends promotions by hand, from a lead sheet, and through
follow-ups could reach the same person several times in a day. Every
Outreach now has a class, Marketing or Service,
and the agent keeps a frequency cap between marketing messages to
the same person: 48 hours by default, adjustable or off under
Settings · Outreach. Inside the cap, a manual send goes out with a
warning, Automated Outreach skips the person until a later run ("too
soon since last message"), and follow-ups keep their configured timing
while still counting as marketing. Service messages, such as
reservation reminders, pre-check-in details, and appointment
confirmations, are never held back. Nothing changes for people you
have already contacted. Also in the
Agent API (minTouchIntervalHours,
playbook outreach.kind, the send result's warning) and the
Anychat MCP server.
How often a person can be messaged
September 3, 2026 — One person per phone number¶
Within an agent, a phone number now belongs to exactly one live person.
Importing a list keyed by phone updates the people already on file
instead of creating flagged duplicates, so a lead who fills in your form
twice (or re-read from a connected sheet) is one record — and is
not messaged again. Adding or editing a person by hand with a number
someone else already has is refused, and the console offers to open the
existing record so you can update or merge it. Numbers the agent merely
heard in conversation never move a phone between records: those rows
are reported for review instead. Also in the
Agent API: phoneE164 on
Person, 409 agent.people.phone_conflict, and per-row rowOutcomes on
import.
People — Matching and update modes
September 2, 2026 — Imports match emails regardless of capitalization¶
A lead sheet that spells an email Jane@Acme.com used to create a
second Jane next to the jane@acme.com already on file, and every
hourly automated-outreach scan could do the same. Emails now match
regardless of capitalization or stray spaces, and are stored in lower
case; existing records that turned out to be the same person are
flagged as possible duplicates for you to merge rather than merged
silently. External IDs remain exact keys, and the ids automated
outreach assigns to sheet rows now have a documented, reserved
gsheet: prefix. Follow-ups to people who have talked to your agent on
more than one channel go to the RCS conversation they most recently
wrote from.
People — Matching and update modes
September 2, 2026 — Framework v2 is now the runtime for every new agent¶
Every agent you create from today runs on Framework v2, in every org,
with nothing to enable: capabilities declared per Playbook, deterministic
paths for step-by-step Playbooks, Policies the platform enforces, and a
record of every turn you can open from a conversation. The create dialog
no longer asks which runtime to use. Agents created earlier keep running
on the Generic (v1) runtime exactly as before and now show a
Generic (v1) badge on their Overview page so you can see which ones
are left to move; moving one stays a per-agent, reversible switch for
orgs in the rollout. Also in the Agent API
(framework defaults to v2 on create; generic is no longer accepted
for new agents), the Anychat MCP server
(create_agent), and the CLI.
September 2, 2026 — See who has no messaging consent¶
Automated Outreach imports every qualified lead but only messages the
ones with recorded consent, and until now the ones it skipped looked
like any other person who had never been contacted. The People list now
says so: a person with no consent on file shows No consent in the
Outreach column, and one who said no shows Declined, each with a
tooltip explaining what happened. The manual Outreach button still
works in both cases, but it carries the same warning, in the button and
again in the send dialog, so nobody messages someone who never opted
in by accident. A new Consent filter (and a matching field in saved
segments) lists everyone who still needs consent, and the person page
has a Messaging consent card where you can record consent you
obtained elsewhere, with a note that stays on the record as evidence.
Also in the Agent API (consent on
Person, the consent filter, and consent on create and PATCH) and
the Anychat MCP server (list_people,
update_person).
August 27, 2026 — Reach every new lead from your spreadsheet, automatically¶
Point an agent at the Google Sheet where your leads already land and it
takes the pipeline from there: every hour it picks up new qualified
rows, creates the People records, and sends your outreach opening to
every lead with recorded consent and a working number — no more
downloading, renaming columns, and sending one at a time. Leads that
aren't ready to message (no consent, bad phone, missing details) are
imported anyway and flagged for you, and the agent can note each row's
outcome right back in the sheet for the team that lives there. Also in
the Agent API and the
Anychat MCP server
(get_automated_outreach_config, start_automated_outreach_run, …).
August 26, 2026 — Local Business: an agent that knows your hours¶
The new Local Business agent template gives a storefront,
restaurant, or service business an agent that answers everyday
customer questions out of the box. Behind it is structured
business info — address,
weekly and holiday hours, amenities, price range — that website import
fills straight from your site. The agent answers hours, location, and
amenity questions from these facts, and "are you open right now?" is
answered from a computed open-or-closed status in your business's own
timezone — never guessed. Also in the
Agent API and the
Anychat MCP server (get_business_info,
replace_business_info).
August 25, 2026 — Website import picks up your business hours¶
When you import from a website, the importer now reads machine-readable business hours published on the site and fills the agent's contact-availability line with them — so the agent can tell customers when you're reachable without you typing a schedule. As with everything the import touches, it never overwrites a value you've already set.
Knowledge — importing from a website
August 25, 2026 — Conversations: whole threads, not 30-minute slices¶
The agent's Conversations page now reads the way your phone's
Messages app does. The sidebar lists contacts — one row per person,
named when a Person record exists — and picking one opens their whole
thread across all time, newest first, with subtle dividers marking each
sitting (a session). Replay a single session or the entire
conversation in the channel emulator ("Entire history" toggle), and
inspect turns from any session divider on Framework v2 agents. The
overview card's Recent conversations list follows the same model and
deep-links into the thread. For API clients, sessions now have their own
endpoint: GET …/sessions in the Agent API
(the old …/conversations route remains as a deprecated alias) and
list_sessions in the Anychat MCP server 0.7.0.
Delivery receipts no longer open sessions of their own, so
status-only rows stopped appearing in the message journal.
August 19, 2026 — Media library: rich cards and carousels from your own pictures¶
Every agent has a media library: the pictures
(and, as rich channels allow, videos) it may show in rich cards,
swipeable carousels and media messages. Upload files or add by URL at
Settings · Media; each item gets a stable id, a name and a usage hint
for people, and a description written by a vision model for the agent.
Message templates and playbooks refer to items by id (media:
image:<id>), so a menu-as-a-carousel or a "here's what we do" showcase
early in the conversation is plain authored copy. Drop a new file on an
item to update the picture everywhere it is used; removing an item
deletes the hosted file, and any message that still mentions it sends
without the picture. In AML: [[carousel]] (one
card per line) and card buttons (link: / reply: / dial:). Also in
the Agent API and the
Anychat MCP server 0.6.0 (list_media,
add_media, update_media, replace_media_content, describe_media,
delete_media).
August 18, 2026 — Framework v2 preview: capability manifests, deterministic paths, turn traces¶
A new agent runtime is available as an opt-in preview. It reads the
same Definition you already author, but with fewer moving parts: the
model sees exactly what you wrote (nothing injected behind your back),
each playbook declares the capabilities it uses,
a playbook can carry a deterministic path
(a Flow — generated from the playbook text, then edited) for the parts
that must not vary, and every turn leaves a
trace you can open from the Messages
page to see which playbook, tools and rules were in play. A
Definition lint on the Overview
points out gaps with one-click fixes, and works on every agent, v1 or v2.
Orgs in the preview see Switch to v2 preview on an agent's Overview;
switching back is one click and changes nothing else. Also in the
Agent API and the
Anychat MCP server 0.4.0 (lint_agent,
list_turns, generate_playbook_flow, update_agent framework).
The agent framework (v2 preview)
August 17, 2026 — Copilot: files, memory, and the command line¶
Anychat Copilot can now work with files: upload a
CSV or notes to a chat's Files strip (or drop them on the panel), let
Copilot read them, and download whatever it writes back — reports,
cleaned-up lists, exports. It also keeps short notes about your
preferences ("What Copilot remembers about me…" in the sessions menu, with
a one-click Forget everything), confirms only what the console actually
did when it opens a page for you, and shows a field-by-field diff even for
whole-document changes such as importing a definition or applying a
template update. And it's in the CLI: anychat copilot for an interactive
session, anychat copilot ask "…" for one-shot questions.
August 16, 2026 — Agent API: manage testers via /v1¶
Retired September 19, 2026: the tester app and Testers were replaced by standalone emulator links.
Inviting people to try an agent in the tester app before it launches no
longer needs the console. The new
Testers resource listed an agent's
testers, invites one by email (the response carries their one-time
sign-in link for you to share), disables or re-enables them,
regenerates a lost link, and removes them: the same actions as
Settings · Testers, with the same guarantees (a link is returned
once, on invite or regenerate, and never on list). The
Anychat MCP server 0.3.3 adds the matching
tools (list_testers, invite_tester, set_tester_disabled,
regenerate_tester_link, remove_tester), so "invite jane@example.com
to test the medspa agent" works from Claude Code or Cursor too.
August 16, 2026 — Agent API: configure scheduling, People fields, products and custom tools; send outreach¶
Four capabilities that you could previously configure only in the
console, or by round-tripping a whole export document, are now their own
resources in the Agent API: the
Scheduling configuration
(hours, reminders, representatives), your
People fields, the
product catalog, and the
contracts of your custom integrations.
Each is a plain GET / PUT (custom tools also get per-tool PUT and
DELETE), returns exactly what an export shows for that slice, and
reports validation problems by field path. And the People page's Send
and Follow up buttons have an API equivalent:
POST .../people/{personId}/outreach
opens the outreach conversation (with an optional per-send Playbook or
greeting), and .../outreach/follow-up sends a nudge now, with the same
checks and outcomes as the console. Playbook POST and PUT also carry
procedures now, so a full-playbook save keeps its procedures.
Anychat Copilot can carry these steps out without handing you off to a console page, and the Anychat MCP server 0.3.0 exposes them as tools, so "add a representative", "declare a budget field", or "send outreach to Jane" work from Claude Code or Cursor as well.
August 16, 2026 — Copilot: see exactly what will change, stop a turn, guided jobs¶
Approval cards in Anychat Copilot now show a field-by-field diff of what an edit will change (current value → new value) before you approve. A Stop button (or Escape) ends a turn mid-way and withdraws any pending change. And Copilot has built-in procedures for common jobs — setting up a channel, configuring outreach, diagnosing a conversation, reviewing an agent against best practice, creating an agent from a website — so "Set up WhatsApp for this agent" walks the whole job with you, one approval at a time.
August 15, 2026 — Meet Anychat Copilot, your assistant inside the console¶
Anychat Copilot is a new panel in the console (the sparkle button in the header, or ⌘J / Ctrl+J) that knows the docs, your workspace, and the page you're on. Ask it how scheduling handles same-day bookings and what this agent is set to; ask which agents have no channel connected; or ask it to add a pricing playbook to the agent you're editing. Reads happen on the spot. Anything that changes your agent arrives as an approval card first, with the exact change one click away, and deleting or provisioning always asks. It can open pages for you and refreshes what you're looking at after a change. Copilot acts with your own permissions and is enabled per workspace; ask your Anychat contact to turn it on.
August 15, 2026 — Template updates become one reviewed merge¶
When the template your agent was created from publishes a new version, you now review the whole update in one place instead of piece by piece. The agent's Definition page shows a quiet "new version available" note; opening it lists every difference between your agent and the new template — changes the template made, changes you made, and the few places where both changed — and you decide, item by item, whether to take the template's version or keep yours. Nothing changes until you apply, and your own additions are never touched. The per-piece update pills and the Settings · Updates page are retired.
Template updates in the Agent API
August 15, 2026 — Drive Anychat from Claude Code, Claude Desktop, or Cursor¶
Your own AI agent can now operate your Anychat workspace. The new
@anychat-ai/mcp server connects any MCP host (Claude Code, Claude
Desktop, Cursor, and others) to your account with an API key, so you can
ask it to review an agent's guardrails, add a Playbook, set up a WhatsApp
channel, or dig through yesterday's conversations, in plain language,
from the tool you already work in. It reads and writes through the same
Agent API the console uses, with your key's scopes; anything that changes
your agent asks for your approval first, and it can consult these docs on
its own before acting.
Use Anychat from Claude Code, Claude Desktop & Cursor
August 15, 2026 — A sharper Sales Development template, and an honest agent preview¶
New agents created from the Sales Development Agent template now start with a tighter, industry-neutral setup: three ranked objectives instead of six (book the call, understand the lead, answer product questions — greeting and confirmation live in the Playbooks where they belong), guardrails and lead fields that fit any business rather than one vertical, and a discovery Playbook that adapts to whether the customer messaged first or your agent reached out. Existing agents are unchanged — as always, template updates only apply when you choose them on Settings · Updates.
The agent Chat Preview now says what it's simulating. A new Outreach / Inbound toggle lets you watch both kinds of conversation start — the outreach opening with sample (or real) lead intake, or a visitor messaging first — with a caption naming the active scenario. And when you Build Agent from a concept, the concept's opening line only becomes your Outreach opener when the concept actually demonstrates outreach; inbound demo openers no longer leak into outreach conversations.
August 13, 2026 — Your branding on concept pages, with a whitelabel add-on¶
Concept pages you publish can now carry your brand by default. The new Workspace · Branding page holds your company name, brand color, logo, favicon, and footer; every concept you generate picks them up, and publishing fills any gaps a concept left blank. Already-published pages list right on the Branding page with a one-click Republish to apply a change everywhere. With the new whitelabel add-on, Anychat attribution disappears from your published pages entirely — your favicon, your footer, your brand only — so the demo you send a prospect reads as yours.
August 10, 2026 — Pick the Playbook for each outreach send, and preview the exact message¶
The Send outreach dialog now decides the whole conversation, not only the greeting. Pick the Playbook the conversation follows (confirm an appointment, qualify and book a call, whatever the moment calls for), and the dialog renders the exact opening message against the recipient's real record before you send, so you can sanity-check what is going out. A template that cannot render for this person disables sending instead of surprising you. On Settings · Playbooks you can now mark any Playbook available for outreach and attach its own opener; an Outreach's Playbook stays out of regular conversations unless you choose otherwise.
August 9, 2026 — Claude Sonnet 5 and Claude Opus 5 in the Model picker¶
Settings · Model now offers Claude Sonnet 5 and Claude Opus 5
alongside the existing options, so you can move an agent to the newest
models with one selection. The page and the agent header now show the
model your agent actually runs, including agents that predate the
picker, and newly created agents start on the platform default (Claude
Sonnet 4.6) until you choose otherwise. Your selection is saved in the
agent's llm block and travels with export and
import.
August 9, 2026 — Playbooks carry their procedures¶
A Playbook now owns its step-by-step instruction blocks — its procedures — instead of referencing them from a separate Skills library. Everything about one situation lives in one place: open a Playbook and its exact mechanics (like the full call-booking procedure) are right there, editable inline, with catalog lineage per procedure. Existing agents were converted with a per-agent verification that the instructions the model receives are byte-for-byte identical, so live behavior is unchanged. The Settings · Skills page retires; its route lands on Playbooks.
August 9, 2026 — Capabilities: what your agent can do, in one place¶
The console's new Capabilities page shows everything your agent can
do beyond conversing — scheduling, knowledge, people memory, outreach,
notifications, custom integrations — as one roster, each card
describing the ability in plain terms with its configuration status and
links. Representatives now sit with Scheduling, where they belong:
they're who the booked calls are with, and the page tells you when none
are configured. Custom HTTP abilities are now called Custom
integrations. The exported agent file was reorganized to match — a
definition layer (what you author) and a capabilities layer (what
you configure) — and the Capabilities
concept page joins The Agent Definition in the docs.
August 8, 2026 — Your agent as one document: the Definition view¶
An agent is defined by a single document, and the console now shows it that way. The new Definition page presents everything you've authored — identity, personality, guardrails, objectives, playbooks, tools, knowledge, content, people — as one outline-navigable view, with each section marked by when it applies (every turn, the active Playbook, or on demand) and linked to its editor. The documentation was rebuilt around the same model: The Agent Definition walks the document section by section, How a Turn Is Composed shows what the runtime does with it, and seven new authoring guides cover Playbooks, Goals, Guardrails, Personality, Capabilities, Knowledge & Things, and Message Message Templates.
August 6, 2026 — Create and publish concepts from the CLI¶
Concept pages can now be managed end-to-end from the anychat CLI:
author a concept as a YAML or JSON document — branding, page copy, and
the scripted conversation flows — then anychat concepts create
--input concept.yaml --publish takes it straight to a live page on
concepts.anychat.ai. List, update, republish, and unpublish work the
same way, so a batch of demo pages can be built and kept in version
control instead of assembled by hand in Concept Studio.
Concepts — From the command line
August 3, 2026 — Filter People, save segments, export the view¶
The People list now filters by tag, source, whether a phone number is on file, and RCS reachability, alongside the existing search and status filter. A combination you use often can be saved as a named segment and reapplied from a picker in one step; segments stay live, matching whoever fits the filter today. Export follows the active filters, so the CSV you download is exactly the list on your screen. The same filters and segments are available through the Agent API.
People — Filtering and segments
July 30, 2026 — Turn off same-day bookings¶
If your team needs lead time before a call, the Scheduling page now has an Allow same-day bookings toggle. Turn it off and the earliest bookable day becomes tomorrow: the agent stops offering today's times and steers anyone asking for "today" to times from tomorrow onward. The cutoff follows your agent's timezone, so a person a few timezones ahead can't book into what is still today for your team.
July 29, 2026 — A Calls page, and calls that don't disappear¶
Your agent now has a Calls page in the sidebar: every call it has ever booked — upcoming, completed, and cancelled — with the person, the rep, and when it was booked. Calls stay on the list after they happen, so a same-day booking you weren't watching for is still there when you check in. The People list keeps the most recent past call visible in its Scheduled Call column too, instead of going blank once the call time passes. Agents also got better at booking in the right timezone: the platform now infers a lead's likely timezone from their phone number so "9 AM" means 9 AM where they are.
Scheduling — Where calls show up
July 22, 2026 — Bigger phones on concept pages¶
Published concept pages now give the phone nearly the full height of the browser window, so the demo reads clearly across a desk or on a projector. A new Phone size control in the concept editor adjusts it per concept, from 0.85× to 1.3× of the phone's natural size; every page still scales down to fit small screens. Already-published pages pick up the roomier layout automatically.
July 21, 2026 — People: search, export, and a scalable list¶
The People page now searches by name, email, phone, or company, filters by funnel status, and loads large lists incrementally. An Export button downloads the whole CRM as a spreadsheet-ready CSV whose columns match the import template, so a round trip through your own tools is straightforward.
July 21, 2026 — Smarter imports and duplicate merging¶
Re-importing an updated list now updates matched records instead of skipping them: choose fill empty fields only (the default) or update matched records at import time. Possible duplicates are flagged as records arrive, and a side-by-side merge dialog lets you combine a pair field by field, or mark them "not duplicates" for good.
July 21, 2026 — Provenance, field locks, and tags on People¶
Every People field now shows who set its current value (your team, an import, or the agent) and when. Contact details the agent learned in conversation read as unverified until a person on your team confirms them, and any field can be locked so the agent can never change it. Records also carry free-form tags, editable by your team and addable by the agent.
July 20, 2026 — Choose a concept's playback speed¶
Concepts can now play at the pace you choose — from 0.5× for an audience reading along to 2× for a quick teaser. Pick the speed in the concept editor and it applies everywhere the concept plays: the live preview, the published page, and exported videos.
July 16, 2026 — One-tap replies on call reminders¶
Pre-call reminders now offer three quick replies — Confirm, Cancel, and Reschedule — so the person can act the moment the reminder lands. Cancel double-checks before cancelling; Reschedule offers new times right away. On channels without tappable replies, typing the same words works just as well.
July 16, 2026 — Export and import agents¶
You can now export an agent as a portable definition file and import it into another agent — or another account — to duplicate a proven setup, move an agent between environments, or keep a versioned backup. Secrets and people data never leave with the file.
July 14, 2026 — Concept videos and branding¶
Concepts picked up two upgrades: export a concept as a video (MP4/MOV) to drop into a pitch deck or send to a stakeholder who won't click a link, and upload a logo to have the concept page match the brand's colours automatically.
July 6, 2026 — Export a conversation history for a call¶
Prepping to call a lead? Open their record on the People page and click Export for call to download a clean, printable summary — their key details, what the agent has learned about them, and the full conversation transcript — so you can walk into the call fully briefed.
July 6, 2026 — Automatic follow-ups¶
Anychat can now automatically follow up with people who don't reply to your initial outreach. Turn it on for an agent — "if no reply after 2 days, send a check-in" — and Anychat sends one follow-up to quiet contacts on schedule and cancels it the moment they respond, so no one who's already talking to the agent gets pestered.
You stay in control: a status chip on the People page shows where each person is, quiet hours keep the follow-up out of the middle of the night, and you can send follow-ups by hand (as many as you like), or pause / resume / stop the automatic one for any person. Follow-ups are RCS-only for now, off until you turn them on, and only ever apply to people you contact after enabling them.
Following up with people who don't reply
June 25, 2026 — Concept Studio¶
Show a prospect what an Anychat agent would feel like for their business — before building anything. Point Concept Studio at a website and it generates a polished, shareable concept page with a scripted agent conversation playing in a channel emulator; refine it, publish it at its own link, and when the prospect says yes, turn the concept into a real agent.
June 10, 2026 — WhatsApp conversational messaging¶
Agents can now hold two-way conversations on WhatsApp through the Meta Cloud API, alongside RCS, Apple Messages for Business, and SMS — so you can meet customers on the channel they already use.
June 1, 2026 — Outreach links: let people text your agent first¶
Not every conversation starts from a number you already have. An outreach link is a shareable URL — for an email signature, a QR code, or an ad — where a person enters their own mobile number and your agent opens the RCS conversation with them. Links are use-limited and time-limited, and you can pause one at any moment without losing its remaining uses.