Adapters

Adapters

Experimental (v1): these adapters record through the v1 laptop signer. They work and are tested, but may change and get less hardening than the stable core; new agents use the v2 integrations.

An adapter routes a framework's real tool calls through the SDK, so each one is policy-checked before it runs and its result is recorded. Anything the framework does outside the adapter is not seen.

Any function

from tracekit.agent_sdk import Tracer
from tracekit.adapters import traced

with Tracer(agent="my-agent") as tracer:
    @traced(tracer, name="Bash")          # sync or async
    def shell(command): ...

A denied call raises PermissionError and the function body never runs.

Policy rules match on the tool name. The default rules know Bash, Write, Edit, WebFetch and so on. If your shell tool is called run_command, the shell rules will not apply to it: name it Bash for tracing, or add rules for your own tool names in a policy file (extends: default).

LangChain and LangGraph

pip install "tracekit-ai[langchain]"

from tracekit.adapters.langchain import TracekitCallbackHandler
with Tracer(agent="my-graph") as tracer:
    graph.invoke(inputs, config={"callbacks": [TracekitCallbackHandler(tracer)]})

The handler uses LangChain's callback system, which LangGraph shares. It raises on a denied call, and LangChain propagates the exception before the tool body executes. It is tested against langchain-core and a LangGraph ToolNode. Model calls are not recorded by this adapter.

MCP client sessions

from tracekit.adapters.mcp import traced_session
session = traced_session(session, tracer, server="github")
await session.call_tool("create_issue", {"title": "..."})    # gated and recorded as mcp__github__create_issue

Names follow Claude Code's mcp__<server>__<tool> convention, so the default and strict policies' MCP rules apply. A denied call raises PermissionError and never reaches the server. MCP tool errors (isError) stay results for the caller and are recorded as failed calls.

Model calls, onchain transactions, other harnesses

• Model calls: tracekit_sdk.init() (OpenAI, Anthropic, Google Gen AI) or OpenTelemetry (otel).
• Onchain transactions behind a guard: tracekit_onchain.guarded_tx (contrib/onchain).
• Codex CLI, Cursor and Gemini CLI: hooks, tracekit init --dev --agent ... (coding agents).

To add an adapter for something else: translate each tool call into tracer.tool(name, args) and write a test that proves a denied call never executes.