Viatoris LangChain Adapter

A drop-in LangChain callback handler that emits a cryptographically signed Viatoris receipt for every tool call your agent makes.

No rewrite. No latency. Receipt submission is fire-and-forget on a background thread and never raises — Viatoris is designed to never crash your agent.

Install

pip install viatoris-langchain

This pulls in viatorisai (the core SDK) and works with any langchain-core>=0.1.0 agent.

Usage

Get an API key and create an agent passport at dashboard.viatoris.ai, then save the returned private_key_jwk somewhere your agent can read it. After that, one line of integration:

from viatoris_langchain import ViatorisCallbackHandler

handler = ViatorisCallbackHandler(
    org_api_key="at_live_...",
    agent_did="did:web:example.com:agents:scout",
    private_key_jwk=saved_key,
    key_id="did:web:example.com:agents:scout#key-1",
)

# Add to any LangChain agent — every tool call gets a signed receipt
agent = AgentExecutor(agent=..., tools=..., callbacks=[handler])

That's it. Run your agent normally. Every on_tool_end produces a signed ActionReceipt with the tool name, inputs, outputs, and a success/failure status. Browse them at dashboard.viatoris.ai/audit or query them through the Viatoris API.

Tracking LLM and chain calls too

By default the handler only emits receipts for tool calls — that's almost always what you want for an audit trail. If you also want to receipt every raw LLM call or every chain step, opt in:

handler = ViatorisCallbackHandler(
    org_api_key="at_live_...",
    agent_did="did:web:example.com:agents:scout",
    private_key_jwk=saved_key,
    key_id="did:web:example.com:agents:scout#key-1",
    track_llm=True,
    verbose=True,
)

Both track_llm and track_chains produce a lot of receipts on a busy agent. Tools-only is the recommended default for production.

How it works

LangChain emits callbacks at every stage of agent execution. This handler:

1. Stores tool inputs by run_id on on_tool_start.
2. On on_tool_end, looks up the stored input, builds an ActionReceipt, signs it with your agent's Ed25519 key, and submits it to the Viatoris receipt service on a background thread.
3. On on_tool_error, does the same but with result="failure" and the error in the outputs.

Every method body is wrapped in a try/except — if the SDK, the network, or anything else throws, your agent keeps running.

Documentation

Full Viatoris guide at viatoris.ai/quickstart.

Questions? Email jared@viatoris.ai.

License

MIT