Decided before the side effect
The SDK evaluates the policy locally, right before the tool runs. Raw arguments, results and model text stay in your process.
YI GUARD
Yi 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.
DEMO
The agent is serving customer C-1001.
Agent wants to look up order 1042
Checkinglookup_order(customer_id="C-1001", order="1042")It is this customer’s own order.
Agent wants to look up order 2077
Checkinglookup_order(customer_id="C-2077", order="2077")It is someone else’s order, and no rule allows it.
Agent wants to export every customer record
Checkingexport_customers(fields=["name", "phone", "address"])It gets the masked export, with no phones or addresses.
Agent wants to refund ¥199 on order 1042
Checkingrefund(order="1042", amount=199)The agent stops and waits for you.
Rules
Decided locally. Arguments are never uploaded.
WHAT IT DOES
The SDK evaluates the policy locally, right before the tool runs. Raw arguments, results and model text stay in your process.
Allow, deny, allow with warning, require approval or degrade. When Guard can’t decide, the tool does not run.
Adapters for LangChain, LangGraph, Claude Agent SDK, OpenAI Agents, Pydantic AI, Google ADK and more, plus the Vercel AI SDK in TypeScript.
Put Guard between the agent and its MCP servers. Tools are admitted one by one, their metadata is pinned, and every call is checked.
Start in Observe mode, see what the agent really does, generate a least-privilege policy and preview it before anyone publishes.
Each decision carries the policy version, rule and reason. Pull events and the audit log as NDJSON into your own log system.
HOW TO START
Wrap its tools with the SDK, or put the MCP proxy in front of its MCP servers.
Run in Observe mode and see what the agent really does.
Generate a least-privilege policy from what the agent did, have a second person review it, then enforce it.
CODE → EVAL → GUARD · CLOUD
01 · BUILD
Tell it what to change. It edits the code in an isolated sandbox, runs the tests and opens a draft PR for you to review.
Explore Yi Code02 · BEFORE LAUNCH
Runs your agent for real in a sandbox, through normal tasks, tool faults and attacks. A version that fails the release gate does not ship.
Explore Yi EvalMODELS · COMPUTE
Works with the OpenAI and Anthropic SDKs. Change the base URL and call models. Each sandbox and VM is a separate virtual machine, created and deleted through the API.
Explore Yi CloudNext step