Skip to content

Yi Guard · Docs

Every agent action, within bounds

Guard checks each tool call, MCP call and outbound request against your policy before it runs, inside your own process, and keeps a redacted record of every decision. We set it up together with your team.

Getting access

To start, ask us for access. We set up the hosted control plane, the private SDK download and the first policy together with your team. Get access on our contact page.

What Guard does ​

  • Checks actions before they run. A tool call, an MCP tool call or an HTTP request goes through Guard first. Guard returns one of five decisions: allow, deny, allow with warning, require approval or degrade. The tool runs only when the decision lets it.
  • Fails closed. When Guard can't decide (no valid policy, revoked credentials, the audit queue full, the policy offline for too long), the tool does not run.
  • Keeps your data at home. With the SDK, the decision is made in your process from a cached copy of the policy. The control plane receives redacted events: names, decisions, policy versions and digests of arguments and results, never the values.
  • Turns behaviour into policy. Watch the agent in Observe mode, generate a least-privilege candidate from what it did successfully, preview the change, have a second person review it, then publish.
  • Records who changed what. Policy versions, reviews, publishes, rollbacks and approval decisions are all kept, with a control audit log you can export.

What it does not do ​

  • Guard protects the calls that go through it: wrapped tools, the MCP proxy, the guarded HTTP client. An agent that can open its own sockets or run shell commands needs a separate network or sandbox boundary as well.
  • It does not read or filter the content of HTTPS traffic. There is no general data loss prevention.
  • It cannot promise to stop every prompt injection. It limits what an injected instruction can make the agent do, within the rules you publish.
  • The demo scenarios in our workbench are fixed replays of recorded model output, not live attack tests against a model.

Start here ​

I want to…Go to
Connect an agent to our hosted control planeQuickstart: hosted control plane
Run the control plane myselfQuickstart: self-hosted, Self-hosted control plane
Understand the decisions and what happens offlineHow Guard works
Write rulesPolicies, Policy schema
Go from observing to enforcingPublishing a policy
Protect MCP toolsMCP proxy
See which frameworks are coveredFramework adapters
Answer a security reviewSecurity
Fix a failing connection checkTroubleshooting