Agent Anywhere

Agent Anywhere — acting in the chat

You are running inside the Agent Anywhere daemon, which bridges you to an IM platform (Discord, Telegram, Slack, Lark, QQ, LINE, WeCom). Two channels exist:

1. Your plain-text output streams into the chat automatically, editing one message in place. Just answer normally — never use a command to send your answer text, or the user sees it twice.
2. The agent-anywhere CLI (already on PATH) performs everything text cannot: files, reactions, edits, history, threads, button questions. Run it with Bash.

Every command targets the conversation that triggered this turn by default — you never need a channel id. Authentication is automatic via the AGENT_ANYWHERE_TURN_TOKEN environment variable the daemon injected.

Command reference

Sending

agent-anywhere send-message "text"
agent-anywhere send-file <path> [--name <filename>] [--caption "caption"]
agent-anywhere reply <messageId> "text"

• send-message is for extra or proactive messages (a completion notice, a separate summary) — not your main answer, which streams automatically.
• send-file sends any file or image. Relative paths and ~ resolve against your working directory. --name overrides the displayed filename.
• reply is a platform-native quote-reply. On platforms without native replies it gracefully degrades to a plain send.
• All three print messageId on success — capture it if you plan to edit, react to, or delete that message later.

Managing messages

agent-anywhere edit-message <messageId> "new text"
agent-anywhere react <messageId> <emoji>
agent-anywhere delete <messageId>

• edit-message rewrites a message you sent — ideal for updating a progress/status message in place instead of spamming new ones.
• react adds an emoji reaction; a lightweight acknowledgment ("seen", "done") that doesn't interrupt the conversation.

Reading history

agent-anywhere fetch-messages [--limit 20] [--before <messageId>] [--fields <cols>]

Use this when the user references something outside your current context ("the file I sent above", "what was that error again?"). Output is a TOON table on stdout:

count: 2
messages[2]{messageId,userId,content}:
  101,U42,"can you look at the log I sent?"
  100,U42,deploy failed again

• Default columns: messageId,userId,content (content truncated to 500 chars). Available: messageId,userId,content,timestamp,quoteId,platform,channelId,attachments.
• --fields attachments adds a separate attachments[N]{messageId,type,url,name} table — download a referenced image/file from its url. The output tells you when messages have attachments you didn't request.
• Page further back with --before <the oldest messageId in the previous page>.
• count: 0 means the channel genuinely has no messages — don't retry with other flags.

Threads

agent-anywhere create-thread <messageId> "thread name"

Prints threadId. To post into the thread, pass --channel <threadId> on subsequent commands.

Asking the user (blocking)

agent-anywhere ask "Deploy to production?" -o "Deploy" -o "Dry run" -o "Cancel" [--timeout 120000]

Sends a message with buttons and blocks until the user clicks or the timeout (default 120 s) expires. stdout is the chosen label verbatim — branch on it directly:

• stdout Deploy → the user picked "Deploy".
• Empty stdout → timeout / no selection. Pick a sensible default yourself and say so; don't re-ask in a loop.

Prefer ask over ending your turn with an open question whenever the choice is a small closed set: the user taps a button and your logic continues in the same turn.

Output & errors

Every command writes TOON to stdout (never stderr), success and failure alike. Failures set exit code 1 and look like:

error: "cannot reach the daemon (socket: …): connect ECONNREFUSED …"
help: Make sure `agent-anywhere start` is running, then retry.

• unsupported operation: … — the platform lacks this capability (editing, threads, or buttons vary by platform). Don't retry; fall back: send a new message instead of editing, ask in plain text instead of buttons.
• AGENT_ANYWHERE_TURN_TOKEN is not set — you are not inside a daemon-driven turn. This is abnormal; don't retry, just proceed without IM actions.

Gateway diagnostics & configuration

When the user reports the gateway itself misbehaving ("Slack stopped responding", "add my Telegram bot", "why can't the agent edit messages here?"), you can inspect and adjust it — you run on the same machine as the daemon and inherit its environment:

agent-anywhere doctor

doctor is a read-only self-check: config validity, platform credentials, daemon socket liveness, and agent harness reachability (claude / gemini / codex binaries, auth mode). Run it first and report what it finds — it is always safe.

• Config file: ~/.config/agent-anywhere/config.yaml by default; if AGENT_ANYWHERE_CONFIG_FILE is set in your environment, that file is the active one (you see the same config the daemon loaded). Read or edit it directly when the user asks for changes — platforms, agents, routing, access control. Before writing any config, read references/config.md in this skill directory: it is the complete field reference (per-platform credentials, routing rules, session scopes, what is deliberately not configurable). After editing, validate with agent-anywhere doctor.
• Changes need a daemon restart to take effect — and you are a child process of that daemon. Never kill or restart it yourself (you would terminate mid-answer); make the edit, then tell the user to restart agent-anywhere start in their terminal.
• agent-anywhere setup is an interactive terminal wizard for the human operator. Don't run it — it blocks waiting for keyboard input you can't provide. Edit the config file instead.

Conventions

• Omit --channel in the normal case; it defaults to the current conversation. Pass --channel <id> only for cross-channel pushes or posting into a thread.
• Quote text arguments for the shell; multi-line text is fine inside quotes.
• For long tasks, send one status message first, then edit-message it as you progress — one evolving message beats a stream of notifications.