Signal Messaging (via signal-cli)

signal-mcp

[Release] [License: MIT]

An unofficial MCP integration for Signal via signal-cli. It lets an MCP client send you Signal messages (task-done pings, blocking questions, permission requests) and read your replies.

Requirements: a running signal-cli JSON-RPC daemon (signal-cli daemon --http ..., plus --tcp ... if you want replies) and a Signal phone number already registered with signal-cli. This project does not register numbers, run the daemon, or talk to Signal's servers itself; it is only a client of your daemon. It is not affiliated with or endorsed by Signal Messenger or the Signal Foundation.

It consists of an MCP server (app/server.py) and an optional reply subscriber (app/subscriber.py) that fronts the existing signal-cli daemon without re-registering or bind-mounting its data directory.

Install

The image is published at ghcr.io/joshuawowk/signal-mcp and listed in the MCP Registry as io.github.joshuawowk/signal-mcp. Run with no arguments, the container starts the reply subscriber (the long-running mode used by the Compose deployment below); the MCP server is started explicitly.

MCP client over stdio (send-only)

docker run -i --rm \  --add-host host.docker.internal:host-gateway \  -e SIGNAL_SENDER=+15555550100 \  -e SIGNAL_DEFAULT_RECIPIENT=+15555550101 \  -e SIGNAL_RPC_URL=http://host.docker.internal:8080/api/v1/rpc \  ghcr.io/joshuawowk/signal-mcp:latest python3 /app/server.py --stdio

Notes:

• On Linux you can instead use --network host and drop the --add-host and SIGNAL_RPC_URL settings (the default 127.0.0.1:8080 then reaches your daemon).
• With a bridge network the daemon must be reachable from the container. A daemon bound only to 127.0.0.1 is not. The signal-cli HTTP endpoint is unauthenticated: do not expose it beyond a trusted interface.
• A standalone stdio container has no subscriber, so replies are not collected: signal_check_replies returns nothing and signal_health reports the receive path as down. For replies, use the Compose deployment below.

Compose deployment (send and receive replies)

See Setup.

Architecture

MCP client ──stdio──> docker exec ──> python3 /app/server.py ─┐                                                                  │ HTTP JSON-RPChost: signal-cli-daemon.service  ◀── 127.0.0.1:8080 (synchronous)─┤  (send, version)                                  ◀── 127.0.0.1:7583 (TCP stream)─┤  (receive notifications)                                                                  │       docker container (signal-mcp, restart=unless-stopped)      │       PID 1: subscriber.py ── persistent TCP socket ─────────────┘                            └─> append /data/inbox.jsonl  ◀── signal_check_replies reads

• No bind-mount of signal-cli data. The container is purely a client of the host daemon. Signal identity keys never enter the container.
• No re-registration. The Signal identity stays on the host.
• Two-channel daemon. Synchronous calls (send, version, health) go over HTTP on :8080. Incoming-message notifications stream over a persistent TCP socket on :7583. The HTTP daemon's receive method is permanently locked by signal-cli's internal auto-receive loop, so TCP is the only path that works for replies.
• One daemon flag added. The --tcp 127.0.0.1:7583 arg was added to the existing signal-cli-daemon.service ExecStart (see "Daemon configuration" below). HTTP behavior for any other consumer is unchanged.

Setup

git clone https://github.com/joshuawowk/signal-mcp && cd signal-mcpcp .env.example .env       # set SIGNAL_SENDER and SIGNAL_DEFAULT_RECIPIENTdocker compose builddocker compose up -ddocker compose logs -f signal-mcp   # watch the subscriber start up

Verify the daemon is reachable from the container:

docker exec signal-mcp python3 -c \  "import rpc; print(rpc.ping())"# expected: 0.14.3

Register the MCP server with Claude Desktop

Merge claude_desktop_config.snippet.json into ~/.config/Claude/claude_desktop_config.json under mcpServers:

"signal": {  "command": "sg",  "args": ["docker", "-c", "exec docker exec -i signal-mcp python3 /app/server.py"]}

Restart Claude Desktop.

Why the sg wrapper

sg docker -c "…" runs the command under the docker supplementary group. This is required when the Claude Desktop process was started before your user was added to the docker group — existing sessions inherit a stale supplementary-group list and docker exec returns permission denied while trying to connect to the docker API at unix:///var/run/docker.sock, which surfaces in Claude Desktop's logs as write EPIPE / Server transport closed unexpectedly.

The exec inside the -c string replaces the temporary bash spawned by sg with the docker exec process, so SIGTERM/SIGPIPE from Claude Desktop reach the MCP server cleanly when the connection closes.

If you log out + back in (or reboot), all new processes inherit the docker group and the wrapper isn't strictly needed — but it's harmless to keep, and it makes the config portable across hosts where the group situation may vary. The signal server will appear in the MCP server list and its tools (signal_send_message, signal_send_question, signal_send_permission_request, signal_check_replies, signal_health) become available to any session.

Install the companion skill

An example agent skill lives in skill/SKILL.md (written for one person's setup; adapt the name and wording). Copy or symlink it into your agent's skills directory, for example:

mkdir -p <your-skills-dir>/signalln -sf "$PWD/skill/SKILL.md" <your-skills-dir>/signal/SKILL.md

End-to-end test from a Claude session

Use signal_health to confirm the daemon is up, then signal_send_messagewith text="signal-mcp end-to-end test from session XYZ" andsession_id="signal-mcp-test-001". Then wait 30 seconds and callsignal_check_replies with session_id="signal-mcp-test-001".

Reply with anything from your phone; the next signal_check_replies call should return your message.

Tools

Reliability notes (Jun 2026)

• The inbox file (/data/inbox.jsonl) is created at subscriber startup and is never deleted by retention pruning — an empty inbox stays as a 0-byte file. Previously, pruning removed the file entirely once all records aged out (14-day retention), which made signal_check_replies / signal_health look like the subscriber was down. signal_check_replies and signal_health also ensure the file exists on every call.
• signal_health and signal_check_replies report subscriber_running (scans /proc for subscriber.py) and receive_channel_ok (TCP connect to the daemon's :7583), so a fresh session can distinguish "no replies yet" from "receive path broken".
• signal_send_question tolerates clients that serialize the options list into a string (see tools table).

Config (.env)

Operational

docker compose ps                       # statusdocker compose logs -f signal-mcp       # subscriber outputdocker exec signal-mcp tail -f /data/inbox.jsonl   # raw repliesdocker compose restart signal-mcp       # bounce the containerdocker compose down                     # stop

If the host daemon is restarted, the subscriber will reconnect on its next poll (errors back off up to 60 s).

Notes & limitations

• Linux host only. network_mode: host is a no-op on Docker Desktop for macOS/Windows. To move it: switch to a bridge network, set SIGNAL_RPC_URL=http://host.docker.internal:8080/api/v1/rpc, and add extra_hosts: ["host.docker.internal:host-gateway"].
• HTTP daemon polling. signal-cli's HTTP mode does not stream receive events; we poll. A TCP-mode daemon with WebSocket subscription would be lower-latency, but would require changing your existing daemon service.
• Attachments are skipped. The subscriber only persists text bodies. Reactions, typing indicators, and read receipts are ignored.
• The phone number is not memorized. It lives only in .env and the registration snippet, both inside this project.

Daemon configuration

This project requires the host's signal-cli-daemon.service to expose both HTTP (for synchronous calls) and TCP (for streaming receive notifications). The unit file at ~/.config/systemd/user/signal-cli-daemon.service should have:

ExecStart=/usr/local/bin/signal-cli -a +15555550100 daemon --http 127.0.0.1:8080 --tcp 127.0.0.1:7583 --receive-mode on-start

To apply changes:

systemctl --user daemon-reloadsystemctl --user restart signal-cli-daemon.service

A backup of the original unit file is written next to the live one with a .bak.YYYYMMDD-HHMMSS suffix the first time signal-mcp is set up.

Why not just use HTTP for receive?

signal-cli's HTTP daemon mode is synchronous-only. receive over HTTP always errors with "Receive command cannot be used if messages are already being received" because the daemon's internal receive loop owns the queue and HTTP has no way to deliver push events. TCP mode opens a persistent newline-delimited JSON-RPC stream and pushes receive notifications as they arrive — that's the path subscriber.py uses.