Coco
Integrate and operate

Choose an integration

Pick the guard, the gate, or the hook for your agent, and decide how independent the check needs to be.

Pick the guard, the gate, or the hook for your agent.

Coco checks tool calls wherever your agent runs. Find your agent below, then follow its guide.


Your agentUseWhat you getGuide
Python on coco-frameworkCocoGuardOne middleware line. An ESCALATE pauses the graph until a person decides.Python agents
Python on upstream LangChain or LangGraphcoco_sdk.langchain.CocoGuardThe same guard on upstream's middleware.LangChain and LangGraph
TypeScriptCocoEnforcerA check you call before your code runs the tool.TypeScript
Any other runtime or languagePOST /v1/checkThe same check over HTTP.HTTP API
Claude CodeThe trustcoco installerA PreToolUse hook, ahead of Claude Code's own permission prompt.Claude Code

Where the decision happens

PathWho decidesWhere the receipt lives
In processEngine, inside your application, against the contract files in the project.A SQLite ledger in the project, .coco/ledger.db by default.
Hosted gateThe gate, over HTTP, against the contracts registered for your project.The gate's ledger, shown in the dashboard's Audit Logs.
Claude Code hookThe engine, in the hook process, against the contracts in COCO_HOME.COCO_HOME, or the gate once you connect it.

The guard and the gate combine. A Python project decides in process until coco connect gives it a gate URL and a key. After that, load().engine() and load().guard() decide through the gate, and your agent code stays the same. The TypeScript client always calls the gate.

How independent the check needs to be

In process, the guard is part of your application, and whoever owns the code owns the guard. The gate is a separate service with its own keys, contracts, and record, so the agent's host can't change a rule or a receipt. Choose the gate when the record has to stand up outside the team running the agent.

Keep in mind: The gate's round trip hasn't been measured. An in-process rule decision takes 0.27 ms at the median.


On this page