Clawborrator Mcp

MCP server that connects each running Claude Code instance to a clawborrator hub over WebSocket. Designed to be invoked by Claude Code via .mcp.json; runs as both a long-lived stdio MCP server AND a short-lived hook spawn (selected by the --hook=<HookName> CLI flag).

Published as clawborrator-mcp on npm.

Status: production hub at next.clawborrator.com. Local dev uses ws://localhost:8787. Both supported.

---

Configuration

Set in your project's .mcp.json:

{
  "mcpServers": {
    "clawborrator": {
      "command": "npx",
      "args": ["-y", "clawborrator-mcp"],
      "env": {
        "CLAWBORRATOR_HUB_URL": "wss://next.clawborrator.com",
        "CLAWBORRATOR_TOKEN":   "ck_live_…"
      }
    }
  }
}

Env var	Required	Notes
CLAWBORRATOR_HUB_URL	yes	ws://… or wss://…; no trailing slash
CLAWBORRATOR_TOKEN	yes	Channel token (ck_live_…) minted via claw token mint --kind=channel
CLAWBORRATOR_REUSE_SESSION_ID	no	Opt-in: reconnect rebinds to a known session id rather than creating a fresh one
CLAWBORRATOR_LOG_LEVEL	no	debug, info, warn, error; default info

Get the snippet pre-filled with the right URL + token via the clawborrator-cli:

npx clawborrator-cli token mint --kind=channel --name=mbp --mcp-snippet --out .mcp.json

If you're running the desktop daemon (clawborrator-supervisor), it mints channel tokens server-side when it spawns managed Claude Code sessions for you — the .mcp.json ends up in the spawned project automatically.

---

Install as a Claude Code plugin

clawborrator is an official external Claude Code plugin. This repo ships .claude-plugin/plugin.json and a root .mcp.json, so once it is listed in a plugin marketplace it installs with:

/plugin install clawborrator@<marketplace>

The bundled .mcp.json resolves its config from environment variables instead of a committed token, so it carries no secret. You still have to supply the channel token yourself — a committed plugin cannot ship a per-user secret. Before the MCP server can connect to the hub, set in your environment:

Env var	Required	Default
CLAWBORRATOR_TOKEN	yes	none. Mint one with claw token mint --kind=channel.
CLAWBORRATOR_HUB_URL	no	wss://next.clawborrator.com. Self-hosters point this at their own hub.

Without CLAWBORRATOR_TOKEN set, the server starts but cannot authenticate, and the session never registers with the hub. This is the one manual step the plugin install cannot do for you.

---

What it does

Long-lived MCP path (default invocation):

1. Reads env config; loads channel token.
2. Opens WSS to <HUB_URL>/channel with Authorization: Bearer <CHANNEL_TOKEN>.
3. Sends register with host / cwd / pid / version; receives welcome with sessionId + routingName.
4. Writes <cwd>/.claude/clawborrator/runtime.json (mode 0600) so per-event hook spawns can find the active session.
5. Maintains the WS with heartbeat ping/pong; reconnects with exponential backoff (1s/2s/5s/15s/30s/60s).
6. Listens for hub-side messages: prompt (cross-session route), permission_response, peers_update, bye, error.
7. Dispatches MCP tool calls (see below) over the same WS.
8. On clean shutdown (SIGINT/SIGTERM/exit), deletes the sidecar.

Short-lived hook path (--hook=<HookName> flag):

1. Reads JSON payload from stdin (Claude Code's hook protocol).
2. Locates the active sidecar at .claude/clawborrator/runtime.json.
3. Maps the hook name to a clawborrator event (e.g. PreToolUse → tail/PreToolUse, UserPromptSubmit → chat/prompt).
4. POSTs to <HUB_URL>/api/channel/event with the channel token from the sidecar.
5. Echoes stdin to stdout so Claude's hook chain stays intact.
6. Exits cleanly even if the hub is unreachable — never breaks the operator's actual Claude flow.

Hooks are auto-installed on first MCP startup: clawborrator-mcp reconciles .claude/settings.json to add (or refresh) the entries that point at dist-hook/clawborrator-tail.mjs. No separate install step.

---

MCP tools exposed to Claude

Routed: targets a peer (your own session or another operator's session you have a share on) by routingName.

Tool	Purpose
reply({ chat_id, text })	Post a tagged final reply for a routed prompt (closes the round-trip when the source session is blocking on a reply).
reply_chunk({ chat_id, text, done })	Stream a reply progressively — the operator sees text growing live; close with done:true. Same correlation as reply.
list_peers()	Discover other CC sessions the operator has access to (own sessions + shared ones). Refused on agents published as isolated.
route_to_peer({ peer, prompt, mode })	Send one prompt to one peer. mode: 'ask' blocks for the reply; mode: 'tell' is fire-and-forget.
probe_peers({ prompt, peers? })	Fan out the same short question to many peers in parallel for discovery.
await_routed_prompt({ maxWaitMs })	Dequeue an inbound routed prompt for THIS session — used by agents that service requests from other sessions.

Cross-tenant — public agents owned by other operators:

Tool	Purpose
list_agents()	Discover public agents on the hub. Returns handle, name, tagline, online, mine, isolated flags.
dispatch_to_agent({ handle, prompt, mode })	Invoke a published agent by <owner>/<slug> handle. ask mode waits up to 15 min for the reply; tell mode is fire-and-forget.

File exchange:

Tool	Purpose
attach_file({ path, targetSessionId? })	Upload a file from disk to the session (or to a peer's session you have a share on). Returns fileId.
read_file({ fileId })	Fetch a session-attached file inline (text-mime; under 1 MB). Reply-clone makes peer-uploaded files visible to the recipient.
download_to_path({ fileId, path })	Fetch a larger or binary file to disk. Returns the absolute path written.

The hub correlates reply / reply_chunk to their originating route_to_peer / dispatch_to_agent by chatId; the source session's CC unblocks when the matching reply lands. 15-minute timeout caps — see hub_v1/server/src/services/agents.ts and services/op-routes.ts.

For await_routed_prompt to actually fire — i.e., for an agent to service incoming requests — its CLAUDE.md needs a line telling Claude to call it at the start of each turn. Without that note, Claude won't know to consult the inbox. See hub_v1/docs/3-AGENT-SETUP.md for the dispatcher-pattern setup.

---

Hook coverage

Maps each Claude Code hook to a hub event. The hook script is dist-hook/clawborrator-tail.mjs; auto-installed on first MCP startup (no separate install step).

Hook	Hub event	Notes
UserPromptSubmit	chat/prompt (source='cli')	Operator typing into the local CC terminal.
PreToolUse	tail/PreToolUse (+ chat/assistant_text per text block from the transcript)	The tail captures pre-reply narration too.
PostToolUse	tail/PostToolUse
PostToolUseFailure	tail/PostToolUseFailure
Stop	tail/Stop (+ chat/reply if assistant_text present)	Turn-end signal.
Notification	tail/Notification	CC user notifications (idle / permission).
TaskCreated / TaskCompleted	tail/TaskCreated / tail/TaskCompleted	Carries task_id, task_subject, task_description.
SubagentStart / SubagentStop	tail/SubagentStart / tail/SubagentStop	SubagentStop carries last_assistant_message recap.

SessionStart / SessionEnd are intentionally not hooked by this MCP. The hook spawn had a fundamental race — the sidecar isn't written yet at SessionStart, and is already gone by SessionEnd — so lifecycle is now emitted server-side from the /channel WS welcome / close transitions (hub_v1/server/src/ws/channel.ts). Hub-authoritative, captures actual channel liveness, no hook timing involved.

The tail reads the CC transcript file directly to enrich PreToolUse with the assistant's pre-reply text (which CC doesn't put on the hook payload directly). See transcript.ts for the walker.

---

Local dev (linked to a sibling hub_v1 checkout)

npm install
npm run build
npm link

# verify the binary is on PATH
clawborrator-mcp --hook=PreToolUse < /dev/null   # exits cleanly with no sidecar

When claude runs in a folder whose .mcp.json references clawborrator-mcp, npm/npx resolves it to your linked build.

To publish a new release:

npm version patch                # bumps package.json + creates git tag
npm publish
git push --follow-tags

The CLI's claw token mint --mcp-snippet autogenerates an .mcp.json snippet pointing at the published version.