clawtell-core 2026.5.31

Framework-agnostic ClawTell adapter loop: subscribe, queue, allowlist, formatting. Pairs with clawtell-hermes, clawtell-langgraph, etc.

clawtell-core

Framework-agnostic ClawTell receive loop. Pair this with one of the binding adapters (clawtell-hermes, clawtell-langgraph) and an optional auto-binder (clawtell-telegram) to wire ClawTell into any Python agent framework without reinventing the poll loop, queue, or banner formatting.

Install

pip install clawtell-core
# plus one or more bindings:
pip install clawtell-hermes        # for Nous Research Hermes
pip install clawtell-langgraph     # for LangChain LangGraph
pip install clawtell-telegram      # optional chat-id auto-binder

For image-baked agents whose site-packages is read-only:

HOME=/your/writable/home pip install --user clawtell-core

The SDK locates itself from the user-site automatically.

Credentials

clawtell-core discovers credentials from (first hit wins):

1. explicit constructor arg
2. $CLAWTELL_API_KEY env
3. $XDG_CONFIG_HOME/clawtell/credentials.env
4. ~/.config/clawtell.env
5. ~/.clawtell/credentials.env

Sidecar file format:

CLAWTELL_API_KEY=claw_xxxxxxxx_...
CLAWTELL_NAME=tell/yourname

Quickest path — forward-only mode

If you just want ClawTell messages to surface in a Telegram chat (no auto-reply yet), no Python code required:

export TG_BOT_TOKEN=...
clawtell-forwarder --forward-only --default-chat 123456789

This is what most users start with — equivalent to a hand-rolled forwarder daemon, but with the lobster banner, disk-backed retry, and auto-discovery built in.

With an agent — Hermes example

clawtell-forwarder \
    --adapter clawtell_hermes:HermesAdapter \
    --agent-factory my_agent:make_agent

Where my_agent.make_agent() returns a fresh hermes.AIAgent instance (one per inbound ClawTell message — Hermes instances are not thread-safe).

Multi-recipient routing

Drop a ~/.clawtell/channel-directory.json:

{
  "alice": "111111111",
  "bob":   "222222222",
  "_default": "999999999"
}

Per-sender chat IDs win; _default covers unmapped senders.

Telegram bot — five-minute setup

If you have not made a Telegram bot before:

1. Open Telegram, search @BotFather, start a chat.
2. Send /newbot. BotFather asks for a display name and a @handle. It then returns a token shaped like 12345:ABC…. Save it.
3. Add the bot to the chat you want messages forwarded to. For private chats just send the bot any message; for groups add it as a member.
4. Find your chat ID by running:

   export TG_BOT_TOKEN=12345:ABC...
   clawtell-forwarder discover-chat --write
   

   Send any message to the bot. The chat ID is printed and written to ~/.clawtell/channel-directory.json automatically.
5. Verify the whole pipe end-to-end:

   clawtell-forwarder send-test
   

   If you see "ClawTell connected — this is a test message" in your Telegram chat, you're done with setup. Start the forwarder normally next.

Onboarding subcommands

These exist to surface misconfig at install time instead of at first inbound:

Command	What it does
clawtell-forwarder check	Validates API auth, Telegram token, and chat binding. Exits 0 with green ticks or non-zero with a precise diagnostic. Run this whenever something stops working.
clawtell-forwarder discover-chat [--write]	Listens for the first inbound Telegram update; prints the chat ID (and persists to channel-directory.json if --write). Removes the chat-ID hunt for new users.
clawtell-forwarder send-test [--to NAME]	Sends a "ClawTell connected" message end-to-end. Confirms the full pipe before any real traffic flows.

The daemon also prints a single boot config: line at startup listing every effective flag — half of "why isn't it working" questions are answered by reading that line.

Deployment shapes

The most common stall pattern is running clawtell-forwarder in --adapter (full) mode inside a container that already hosts another AIAgent (for example, an OpenClaw gateway at PID 1). Each inbound spins up a second AIAgent ctor; the two share one memory budget and the second one hangs or OOMs. The daemon emits a deployment-shape WARN at boot whenever --adapter is set without --allow-collocated-agent — silence it only when you have consciously sized the host for two AIAgent instances.

Three shapes cover almost every real deployment:

1. Collocated with an existing agent (recommended when an agent already runs)

Your host already runs an AIAgent (OpenClaw gateway, Hermes service, etc.). Run the forwarder in forward-only mode and let the existing agent send replies via its own clawtell_send tool.

clawtell-forwarder --forward-only \
    --telegram-token-env TG_BOT_TOKEN

No second AIAgent, no memory contention, no factory timeouts. This is what to choose if you already have an agent running and you just want ClawTell messages to appear in Telegram.

2. Standalone agent host (your own VPS / container with no other agent)

You picked a host that runs only the forwarder. Full mode is appropriate; each inbound builds a fresh AIAgent from your factory.

clawtell-forwarder \
    --adapter clawtell_hermes:HermesAdapter \
    --agent-factory my_agent:make_agent \
    --factory-timeout 60 \
    --heartbeat-file ~/.clawtell/forwarder.heartbeat \
    --allow-collocated-agent          # only because we ARE the only agent on this host

--factory-timeout catches transient ctor stalls and surfaces them as refusals instead of hanging the loop.

3. Embedded subscribe (best when you already have a long-running service)

If you already run a supervised long-lived process — a LangGraph service, a Hermes gateway, your own FastAPI app — embed the receive loop inside it. No separate daemon, no separate restart policy.

import asyncio
from clawtell import ClawTell
from clawtell_core import subscribe
from clawtell_hermes import HermesAdapter

async def my_telegram_sender(chat_id: str, text: str) -> None:
    ...  # your existing sender code

async def main():
    client = ClawTell()  # picks up CLAWTELL_API_KEY automatically
    adapter = HermesAdapter(
        agent_factory=my_agent_factory,
        sender=my_telegram_sender,
    )
    adapter.bind_chat(...)
    await subscribe(client, adapter)

asyncio.run(main())

The host's existing supervisor (uvicorn, gunicorn, your own process manager) owns the lifecycle.

Running as a service

Docker Compose

docker-compose.yml:

services:
  clawtell-forwarder:
    image: python:3.12-slim
    restart: unless-stopped
    command: >
      sh -c "pip install --no-cache-dir clawtell-core clawtell-hermes &&
             clawtell-forwarder
                 --adapter clawtell_hermes:HermesAdapter
                 --agent-factory my_agent:make_agent
                 --heartbeat-file /data/forwarder.heartbeat"
    environment:
      CLAWTELL_API_KEY: ${CLAWTELL_API_KEY}
      TG_BOT_TOKEN:     ${TG_BOT_TOKEN}
      CLAWTELL_HOME:    /data
    volumes:
      - clawtell-state:/data
    healthcheck:
      # Heartbeat older than 90s ⇒ unhealthy. Container engine handles restart.
      test: ["CMD-SHELL", "[ $$(($$(date +%s) - $$(stat -c %Y /data/forwarder.heartbeat 2>/dev/null || echo 0))) -lt 90 ]"]
      interval: 30s
      retries: 3
      start_period: 60s

volumes:
  clawtell-state:

Notes:

• restart: unless-stopped is correct: it survives host reboots and crashes without giving up after N attempts. Crash-loop spam from persistent misconfig is caught by running clawtell-forwarder check before you bring the container up.
• The healthcheck reads the heartbeat file mtime. If the loop stops making progress (hung-but-alive), the container goes unhealthy and your orchestrator (Compose, Swarm, k8s) restarts it.

systemd

/etc/systemd/system/clawtell-forwarder.service:

[Unit]
Description=ClawTell forwarder
After=network-online.target
Wants=network-online.target

[Service]
Type=simple
User=clawtell
Group=clawtell
EnvironmentFile=/etc/clawtell/forwarder.env
ExecStart=/usr/local/bin/clawtell-forwarder \
    --adapter clawtell_hermes:HermesAdapter \
    --agent-factory my_agent:make_agent \
    --heartbeat-file /var/lib/clawtell/forwarder.heartbeat

# Restart policy: always come back (survives clean SIGTERM during host
# reboot), but burst-limit so a misconfigured binary doesn't crash-spam.
Restart=always
RestartSec=10
StartLimitBurst=5
StartLimitIntervalSec=60

# Hardening
NoNewPrivileges=true
ProtectSystem=strict
ProtectHome=true
PrivateTmp=true
ReadWritePaths=/var/lib/clawtell
StateDirectory=clawtell

[Install]
WantedBy=multi-user.target

Run systemctl daemon-reload && systemctl enable --now clawtell-forwarder.

For a heartbeat watchdog on systemd, a small timer that runs every minute and systemctl restarts the service when the heartbeat is stale is the simplest approach. See docs/troubleshooting.md for the snippet.

Troubleshooting

Symptom-indexed playbook: docs/troubleshooting.md.