Snapshot of official Anthropic Channels docs (channels.md, channels-reference.md, routines summary) fetched 2026-05-06 from code.claude.com. Bob needs these to implement the MCP Watcher in Layer 2; the channels-reference.md is the primary implementation reference. docs/channels/README.md cross-references the snapshot back to spec §4 and notes the key facts confirmed: capability key, notification method, lifecycle, the research-preview --dangerously-load-development-channels caveat, and version minimums (v2.1.80 for channels, v2.1.81 for permission relay which we are not using in v1). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| channels-reference.md | ||
| channels.md | ||
| README.md | ||
| routines-summary.md | ||
Claude Code Channels — Reference Snapshot
Snapshot of the official Anthropic docs on Channels, the MCP-over-stdio push primitive that the agent-watcher MCP Watcher (Layer 2) targets.
Fetched 2026-05-06 from https://code.claude.com/docs/en/.
| File | Source | What's in it |
|---|---|---|
channels.md |
/en/channels.md |
Overview: what channels are, supported plugins (Telegram/Discord/iMessage/fakechat), enterprise controls, comparison to other Claude Code surfaces. |
channels-reference.md |
/en/channels-reference.md |
The implementation reference. Server options, notification format, full webhook-receiver walkthrough, two-way reply tools, sender gating, permission relay. This is the primary doc for Layer 2. |
routines-summary.md |
/en/routines.md (summarized) |
Brief summary of Routines, the adjacent Anthropic managed-cloud cron primitive. We are NOT building on Routines — included for context only. |
Why this matters for agent-watcher
Layer 2 (MCP Watcher) is implemented as a Channel — an MCP server declared in each agent's mcp.json, spawned by Claude Code over stdio. It emits notifications/claude/channel events when new pings arrive in the inbox, surfacing them mid-session. See agent-watcher.md §4.
Key facts confirmed by the reference docs (relevant to our spec):
- Capability key:
experimental: { 'claude/channel': {} }registers the channel listener. - Notification method:
notifications/claude/channelwithparams.content(string body) andparams.meta(string→string map of attributes). - Events arrive as:
<channel source="agent-watcher" ...>BODY</channel>injected into Claude's context. - Reply tools: standard MCP tools registered via
ListToolsRequestSchema/CallToolRequestSchema. Maps cleanly to ourack/respond/mark_handledplan. - Research preview gotcha: custom channels (anything not on Anthropic's allowlist) require launching with
--dangerously-load-development-channels server:agent-watcher. We'll document this in INSTALL.md. - Lifecycle: Claude Code spawns the MCP server as a stdio subprocess on session start. Server runs while session is open. No keep-alive responsibility on the server side — stdio close = shutdown.
- Min version: Claude Code v2.1.80 (channels) / v2.1.81 (permission relay, which we are NOT using in v1).
Caveat
This is a snapshot, not a live mirror. If you find behavior that doesn't match these docs, trust live behavior over this snapshot and update via:
cd /home/bob/agent-watcher && /home/bob/.local/bin/refresh-channels-docs # TODO if this becomes painful
For now, manual re-fetch when something feels off.