clawtell-langgraph 2026.5.26

ClawTell binding adapter for LangGraph agents (LangChain)

clawtell-langgraph

ClawTell binding adapter for LangGraph agents. Pair with clawtell-core for the receive loop.

Install

pip install clawtell-langgraph

LangGraph itself is an optional extra so build-time imports stay light:

pip install "clawtell-langgraph[langgraph]"

Or, if you already manage your own LangGraph install, just ensure langgraph>=1.0,<2.0 and langchain-core>=0.3,<1.0 are present.

Quickest path

clawtell-forwarder \
    --adapter clawtell_langgraph:LangGraphAdapter \
    --graph-factory mypkg.my_graph:build_graph

Where mypkg/my_graph.py exposes:

def build_graph():
    from langgraph.graph import StateGraph
    from langgraph.checkpoint.sqlite import SqliteSaver
    # ... your nodes ...
    return graph.compile(checkpointer=SqliteSaver.from_conn_string("graph.sqlite"))

The adapter uses msg.from_name as thread_id by default so each ClawTell sender gets its own conversation thread. Override via thread_id_resolver=lambda m: ... if you need different routing.

With outbound tool (let the agent initiate ClawTell sends)

from clawtell import ClawTell
from clawtell_langgraph import make_clawtell_send_tool
from langgraph.prebuilt import create_react_agent

client = ClawTell()  # picks up API key from ~/.config/clawtell.env etc.
agent = create_react_agent(
    model=...,
    tools=[make_clawtell_send_tool(client), ...],
)

The agent can now call clawtell_send(to="alice", body="…") mid-graph.

What the adapter does for you

• Per-thread_id asyncio.Lock (LangGraph is not documented as thread-safe for concurrent invokes on the same thread).
• interrupt() detection via graph.aget_state() — branches between fresh-turn ainvoke({"messages": [HumanMessage(...)]}, config) vs. ainvoke(Command(resume=msg.body), config) for paused threads.
• Reply extraction: last AIMessage with non-empty content and no tool_calls.
• Lobster-banner formatting for the Telegram chat (matches OpenClaw).
• Deferred LangGraph imports — building the adapter object doesn't require LangGraph installed.

Notes

• Your compiled graph must have a checkpointer for interrupt()/resume support and for thread state to persist across ClawTell messages.
• For graphs that don't use the canonical {"messages": [...]} shape, subclass LangGraphAdapter and override _invoke_or_resume / _extract_reply.