Anychat Copilot¶
Anychat Copilot is an assistant panel built into the console. It answers questions about the product from these docs, reads your agents, People, and conversations to answer questions about your workspace, drafts configuration changes that apply only after you approve them, and opens the page you need. It works with your own permissions: nothing it does is outside what you could do yourself in the console.
Copilot is enabled per workspace. If you do not see the button described below, see Availability.
Opening Copilot¶
Any of these opens the panel, docked on the right of the console:
| From | Action |
|---|---|
| The header | Click the sparkle button (tooltip Copilot (⌘J)). |
| The keyboard | Press ⌘J on Mac or Ctrl+J elsewhere. If the panel is already open, the shortcut focuses the composer; if the composer already has focus, it closes the panel. |
| The command palette | Run Open Copilot. |
The panel stays open as you move between pages and reopens on your next visit if you left it open. Close it with the X in its header or with the shortcut.
Type in the composer (Ask Anychat Copilot…) and press Enter to send.
Asking questions and making changes¶
Copilot handles two kinds of request differently.
Questions run immediately. Copilot searches and reads these docs
(search_docs, read_doc) and reads your workspace through the
Agent API: agents and their Definitions,
capabilities and channels, People and segments, conversations, logs, and
activity. Each lookup appears in the transcript as a small labeled row so
you can see what it consulted. Examples:
- "How does scheduling handle same-day bookings, and what is this agent set to?"
- "Which of my agents have no channel connected?"
- "Summarize yesterday's conversations for the Acme agent where the person asked about pricing."
Changes are proposals. When Copilot wants to change something (add a playbook, edit guardrails, update a person, connect a channel), the panel shows an approval card titled Approve this change? with a plain-language summary and, for edits to something that already exists, a field-by-field diff — each changed field with its current value struck through and the new value beside it (long lists collapse behind Show all changes). Show the exact change expands the raw payload that will be sent. Nothing is applied until you click Approve. Deny stops it; Copilot then explains what it intended and asks how to proceed rather than trying again. A card that receives no decision within ten minutes expires and applies nothing.
For edits to one agent, the card offers Also auto-approve edits to this agent this session. Tick it when you are iterating on a Definition and would rather not confirm each save; the setting lasts for the current chat only.
Some actions always ask, whether or not auto-approve is on: deleting anything (an agent, playbook, template, person, segment, channel, or webhook), requesting a managed channel, importing a definition or applying an agent template over an existing agent, disconnecting a connected integration, and rotating a webhook secret. These are marked as destructive on the card.
Guided jobs. For common multi-step jobs Copilot loads a built-in
procedure first (you'll see a load_skill row) and then works through it
with you — asking before each change as usual. Try: "Set up WhatsApp for
this agent", "Configure outreach for the Acme agent", "Diagnose what went
wrong in yesterday's conversation with Sam", "Review this agent against
best practice", or "Create an agent for acme.com". Where a step needs the
console (for example importing a website), Copilot says so and offers to
open the page.
Stopping a turn. While Copilot is working, the send button becomes Stop (Escape in the message box also stops). The turn ends with a muted Stopped line; anything it was about to change is left untouched, and a pending approval card is withdrawn.
Page awareness and navigation¶
Each message you send carries the page you have open: the current agent, the section (Playbooks, Channels, People…), and the record when one is open (a person, a channel, a concept). "This agent", "the person I'm looking at", or "the playbook called Pricing on this agent" resolve without ids or searching.
Copilot can also move you around the console. When an answer lives on another page it offers to open it, and it can point at a navigation item so you can find it yourself. After you approve a change to something you are viewing, it refreshes that page so the result appears without a reload. Each of these appears in the transcript as its own row.
Sessions and modes¶
Each conversation with Copilot is a chat session. Click the chat title in the panel header to open the sessions menu: start a New chat, resume an earlier one, or delete one. Rename under the title names the current chat. Sessions belong to you within the current workspace (other members do not see them) and are kept for 30 days after their last use, then removed.
Every chat runs in one of two modes, shown under the title and switched by clicking it:
| Mode | Model | Use for |
|---|---|---|
| Fast (default) | Claude Sonnet | Everyday questions and edits. |
| Deep | Claude Opus | Slower and more thorough: reviewing an agent end to end, diagnosing a conversation, larger multi-step changes. |
The footer shows the chat's turn count, tokens, and cost so far.
Files¶
Each chat has a private Files area (the strip above the message box; you can also drop files onto the panel). Upload a CSV, JSON, Markdown, text or YAML file (up to 10 MB) and Copilot can read it — for example a list of people to review before importing, or notes to turn into a playbook. Files Copilot writes (a report, an export, a cleaned-up CSV) appear in the same strip with a download button. Files live with the chat: other chats and other members don't see them, and they're removed with the chat (or after 30 days). Copilot's file tools can only touch this folder — nothing else on the server.
What Copilot remembers about you¶
Copilot can keep short notes about your preferences — "prefers short answers", "mostly works on the Acme agent" — when you state one or ask it to remember. Notes are private to you within the workspace and are used at the start of every chat; they are never customer data or platform state (the API is the source of truth for that). Review or clear them from the sessions menu → What Copilot remembers about me… → Forget everything.
From the command line¶
If you use the Anychat CLI (an internal tool for operators who work in a
terminal), the same assistant is there: anychat copilot opens an
interactive session (approvals answered with y/N, Ctrl-C stops a turn)
and anychat copilot ask "…" runs a single question for scripting. Both use
your CLI sign-in and the same permissions as the console panel.
Safety¶
- Your permissions. Copilot acts as you. Every read and write goes to the Agent API with your own session token and is checked against your role and scopes in this workspace, exactly as if you had made the change in the console. A viewer's Copilot cannot write; no Copilot can reach another workspace.
- Approvals. Reads run; writes wait for your approval; destructive actions ask every time. Only the payload shown on the card is sent.
- Data. Copilot reads and changes your workspace only through Anychat's own services (the Agent API and docs.anychat.ai). Anything it reads from your workspace (agent configuration, People records, conversation logs) is treated as data, not as instructions to follow.
- Metering. Usage is metered per workspace against a daily token budget. When the budget is reached, the panel footer shows the usage and when it resets; Copilot resumes after the reset.
Availability¶
Copilot is enabled per workspace by Anychat. When it is not enabled for your workspace, the header button and the command-palette entry are absent. Ask your Anychat contact to enable it.