tryratify 0.1.1

Local authority enforcement client for Ratify

Ratify

Ratify is an authority control plane for AI agents. You define which tool calls an agent may execute autonomously, which require another principal's approval, and which remain reserved.

Ratify Cloud owns Authority Grants, approvals, client identity, and uploaded audit records. Ratify Client runs beside the agent, synchronizes versioned rules, evaluates tool calls locally, and uploads audit batches asynchronously. Normal tool evaluation does not require a per-call Cloud round trip.

See Ratify v1 architecture for the lifecycle and stale rules behavior.

Authority decisions use three terms:

• Delegated: the agent may execute the action.
• Requires Approval: execution waits for a Cloud reviewer.
• Reserved: the agent may not execute the action.

Install

Ratify uses three related names:

• PyPI package: tryratify
• CLI command: ratify
• Python import: ratify_sdk

Install the published package from PyPI:

pip install tryratify

Create an API key

Sign up at tryratify.com/signup. Signup creates an account, a default agent, and an API key. Copy the key when it appears; its secret is shown only once.

Additional keys can be created or revoked from API Keys.

Start Ratify Client

Use the Cloud URL, API key, and agent ID shown during onboarding:

ratify client start \
  --cloud-url https://tryratify.com \
  --api-key '<copied-key>' \
  --agent-id '<agent-id>'

Ratify detects the platform, downloads and verifies the matching client binary, caches it under ~/.cache/ratify/sidecar/<version>/ratify-sidecar, and starts it. The client listens at http://127.0.0.1:8088 by default. The Cloud Clients page shows its heartbeat, current ruleset version, latest published version, account state, and quota state.

Publish an authority rule

Open Authority Grants, publish YAML using the exact fully qualified tool names used by your agent, and wait for the client version to match the published version:

agent: <agent-id>

rules:
- name: Read public data
  tool: custom.read_data
  condition: true
  action: delegated

- name: Publish changes with review
  tool: custom.publish_changes
  condition: true
  action: requires_approval

- name: Keep destructive operations reserved
  tool: custom.destroy_system
  condition: true
  action: reserved

Rule actions are exactly delegated, requires_approval, and reserved.

Run the first governed action

Wrap actual tool entrypoints with Guard. Use stable namespace.action names; the same name must appear in the tool call and Authority Grants.

from ratify_sdk import Guard


def read_data(item: str) -> dict[str, str]:
    return {"item": item, "value": "public"}


guard = Guard(
    agent_id="<agent-id>",
    sidecar_url="http://127.0.0.1:8088",
)
tools = guard.wrap({"custom.read_data": read_data})

result = tools.dispatch({
    "id": "call-1",
    "name": "custom.read_data",
    "input": {"item": "launch-status"},
})

A Delegated call executes locally. A Requires Approval call appears in Cloud and resumes after review. A Reserved call is not executed. Each decision is queued locally and uploaded to Cloud Audit.

Development

Release maintainers should follow the release guide.

go build ./...
go test ./...
python -m pytest sdk/python/tests
python -m build

The optional customer-support demonstration is isolated under examples/demo and is not part of the Ratify product runtime or Python distribution.

Ratify is licensed under the Apache License 2.0.