colber-langchain

LangChain integration for the Colber platform — observability callbacks, semantic memory backed by colber-memory, and a 6-service toolkit. Apache-2.0.

This is the first public plugin in the "Lego" GTM lever. It depends only on langchain-core (not the full langchain package) and the published colber-sdk PyPI release, so it stays lightweight and is independently versionable.

Install

pip install colber-langchain

For local development inside the Colber monorepo:

pip install -e apps/colber-langchain

Components

ColberCallbackHandler

Captures LangChain run events as Colber observability spans + structured logs. Attach to any chain / agent via the standard callbacks=[...] argument.

from colber_langchain import ColberCallbackHandler

callback = ColberCallbackHandler(
    agent_did="did:key:z6Mk...",
    operator_id="op-demo",
    service_name="my-langchain-agent",
)
# Then plug into any LangChain runnable:
result = my_chain.invoke({"input": "hello"}, config={"callbacks": [callback]})

Hooks covered: on_chain_start/end/error, on_llm_start/end/error, on_chat_model_start, on_tool_start/end/error, on_agent_action, on_agent_finish. Each hook produces a span; *_error hooks also emit a structured log. Trace correlation follows LangChain's parent_run_id so every nested step lives under a single trace id.

Network failures to the observability service are caught and logged at WARN; the chain is never aborted because the backend is sick.

ColberMemory

Backs LangChain's BaseMemory with the colber-memory service (Qdrant + ACL + chiffrement). Cross-agent share semantics are wired in: pass share_with=[did1, did2, ...] and every saved memory is automatically shared.

from colber_langchain import ColberMemory

memory = ColberMemory(
    agent_did="did:key:z6Mk...",
    top_k=5,
    share_with=["did:key:z6MkPeer1", "did:key:z6MkPeer2"],
)

A chat-history flavour (ColberChatMessageHistory) is also exported for use with RunnableWithMessageHistory.

ColberToolkit

Exposes the 6 Colber services as LangChain tools (one per operation, 14 in total).

from colber_langchain import ColberToolkit

toolkit = ColberToolkit()
tools = toolkit.get_tools()  # list[BaseTool]

Pass services=["negotiation", "insurance"] to scope down the surface for a deal-only agent.

The observability service is not exposed as a tool — ColberCallbackHandler already gives the agent first-class observability without LLM-driven calls.

Configuration

The plugin reads three environment variables when no explicit ColberClient is passed:

If none are set, the plugin falls back to ColberClient.local() (β-VM ports on localhost).

You can always pass a pre-built client explicitly:

from colber_sdk import ColberClient
from colber_langchain import ColberToolkit

client = ColberClient.from_base_url("https://api.colber.dev", auth_token="...")
toolkit = ColberToolkit(client=client)

Out of scope (Wave 2.1+ follow-up)

• LangGraph integration (custom graph nodes wrapping Colber services).
• Custom prompt templates pre-baked with reputation/memory context.
• Attestation flow helpers (auto-sign + verify Colber payloads inside an agent loop).

These are noted as Wave 2.1+ follow-ups in the Colber ROADMAP.

License

Apache-2.0 — same license as colber-sdk and @colber/mcp.