riskgate¶
Certify, reject, probe or ask: calibrated, consequence-aware checks for the actions of LLM agents.
riskgate sits between an AI agent and its tools. Before a consequential action runs, it predicts what the action would do under every explanation of the situation that the evidence cannot rule out, and lets it through only if the worst plausible outcome clears a threshold calibrated on labelled outcomes, so that unsafe actions get through at most a chosen fraction α of the time, whatever the quality of your simulator.
| Verdict | When | What happens |
|---|---|---|
| CERTIFY | safe under every plausible explanation, by a calibrated margin | the tool runs |
| REJECT | unsafe under every plausible explanation (or blocked by a hard rule) | the agent gets the reason and re-plans |
| PROBE | the explanations disagree and a read-only check can tell them apart | run the check, then decide again |
| ASK | anything else: ambiguity, unfamiliar evidence, stale calibration, errors | a human decides, with the gate's explanation |
For everyone: the careful night-shift assistant¶
A careless assistant presses the button that fits its first guess. A careful one asks "what could be going on, what would this button do in each case, and is it safe in all of them?"; if it cannot tell, it first looks at a second gauge or calls the supervisor. riskgate is the careful assistant, and it keeps score: tested on hundreds of past or simulated situations, it can say "when I say safe, I am wrong at most 5% of the time", and it says "I don't know" when a situation looks unlike anything it was tested on.
Where to start¶
| If you want to… | Read |
|---|---|
| understand the idea | Concepts |
| run something in five minutes | Quickstart |
| see a realistic lifecycle with real output | End-to-end tutorial |
| start from your agent's traces and outcome labels (no simulator) | Start from your traces |
| plug in your own simulator and diagnoser | Bring your own domain |
| see the use cases of the RiskGate site | Use cases |
| know exactly what is guaranteed | Calibration & guarantees |
| run it in production | Operations · Configuration · CLI |
| use it with LangGraph, ADK, OpenAI Agents SDK, the Claude Agent SDK or MCP | Integrations |
| try it from Claude, Gemini CLI or Codex with agent skills | Agent skills |
| look up a class or parameter | API reference |
| see the evidence (simulated domains) | Results |
Product site, use-case stories and the interactive simulator: riskgate-phi.vercel.app.
Install¶
Python ≥ 3.10; numpy is the only runtime dependency. Optional extras: yaml, otel, mcp, langgraph, adk,
openai-agents, docs.
Status¶
Release candidate 1.0.0rc1 (Apache-2.0). All results in these docs come from simulated domains with scripted agents; real-LLM evaluations are planned. riskgate is a supervisory layer: it does not replace certified safety systems, access control or human oversight.