Kyvvu

Integration · Local engine

Kyvvu for any language

For agents in JavaScript, Go, Rust or anything else: run the engine as a local HTTP sidecar and ask it before every step.

Read the docs

Set up

One command, then every step is checked.

Same engine
The decisions are the Python SDK's: no Python needed in your agent.
Two calls per step
POST /evaluate before the step, POST /record after it. Read the verdict from the body: allow, warn or block.
Local or platform policies
Load manifests from disk with no network, or fetch them from your workspace.
$ kyvvu serve --baseline                       # the bundled 15-policy baseline$ kyvvu serve --policy-file my-manifest.yaml   # your own manifest (repeatable)

The verdict

Read the action, not the status code.

The server answers HTTP 200 for every decision. A 422 means the request was malformed: it isn't a verdict.

actionblockedWhat your agent does
allowfalseRun the step.
warnfalseRun the step. A high, medium or low policy was violated, and the response says which.
blocktrueDon't run the step. A critical policy was violated.

Example

Ask before every step.

POST /evaluate before the step, run it if it's allowed, then POST /record with the result. Here against kyvvu serve --baseline.

Five endpoints
/health, /evaluate, /record, /end_task and /register_agent.
Never ungoverned by accident
With no policy source, an invalid manifest or zero policies in effect, kyvvu serve won't start, unless you pass --allow-ungoverned.
Platform policies
With policies from your workspace, send the agent_id the server prints at start. Any other id gets a 409.
curl -X POST http://127.0.0.1:8080/evaluate \  -H "Content-Type: application/json" \  -d '{    "intended": {      "agent_id": "agent_3f9c2a7b1d4e6f80",      "task_id": "task-abc",      "action_id": "act-1",      "kind": "step.model",      "step_name": "chat_gpt-4o",      "input": {"user_message": "Hello"}    },    "context": {      "agent_id": "agent_3f9c2a7b1d4e6f80",      "environment": "production",      "risk_classification": "limited",      "agent_allowed_tools": ["chat_gpt-4o"]    }  }'

Good to know

  • Listens on 127.0.0.1:8080 by default.
  • Task state is kept in memory: a task can't continue on another server or after a restart.
  • The server has no authentication of its own: keep it on localhost as a sidecar, or put it behind something that authenticates.

Next

Try it before you install it.

In your browser

The playground

Build an agent's path, pick a policy and watch the engine decide.

On your machine

Five minutes, no account

kyvvu try runs a demo agent against the bundled baseline.

For your team

Request a demo

See it on the agents you actually run.