Guardrail checks the data before the agent acts on it

What Guardrail is

Check structured data before an agent takes an action on it. Small scope on purpose: repeatable checks, a bounded local implementation, and a policy an operator can read.

What changed

Pull request #2 revives the project as a v2 alpha with secure policy evaluation, a repository CLI and MCP tools, and a MetaHarness harness. The decision rule is explicit: a missing field denies, while a present but falsy value remains valid.

Protections include bearer token authentication, a 64 KiB input cap and bounded provider calls.

Get started

Python for the dependencies, Node for the harness CLI:

git clone https://github.com/ruvnet/guardrail.git
cd guardrail
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt

Start the MCP surface from the repository harness:

node .harness/runtime/cli.mjs mcp

Expected result: policy tools and the policy resource available to a local MCP host.

Use it today

Practical case: an agent about to submit a form or call a paid API. Input is the structured payload; the workflow is evaluate against policy first; the output is a decision with a reason.

Acceptance test: submit a payload with a required field set to zero or false. It must pass. Then remove the field entirely. It must deny.

Push it further

Experimental commentary. A policy layer is only useful if it sits on the path that actually executes. The limitation is coverage: a check an agent can route around is theatre.

Falsifiable test: try to reach the action without passing through Guardrail. If you can, the integration is wrong regardless of how good the policy is.