The industry has gotten very good at watching AI agents: more telemetry, more monitoring, more behavioral analysis. But watching an agent take a destructive action isn't the same as stopping it.
Shrike is the enforcement layer. You declare what each agent is allowed to do — its scope, its policy, the actions that need a human. When an agent takes an action — a tool call, a shell command, a SQL query, a message to another agent — Shrike evaluates it against your policy at runtime, before it executes, and returns a verdict: allow, warn, require human approval, or block. Out of the box it enforces safe defaults, so an agent is governed before you write a single line of policy.
Shrike doesn't decide on its own authority. It enforces yours — at a point the agent can't reason its way around.
The model isn't the risk. The action is.
- You set the policy; Shrike enforces it. The decision criteria come from your declared agent scope and policy — with safe defaults on out of the box. Shrike is the mechanism that makes that decision run at runtime, not an authority of its own.
- Governs actions, not just inputs. Every tool call, shell command, database write, and agent-to-agent message is evaluated across the full multi-turn session — the prompt, the response, and the action it triggers.
- Runs where the agent can't reason around it. Enforcement happens out of band, independent of which model is driving.
- Session-aware. Decisions account for what the agent has already done this session, not just the request in front of it.
- 9-layer cognitive pipeline. Prompt injection, jailbreak, data exfiltration, PII exposure, and unsafe actions — caught before execution.
- MCP-native, SDK-first. Drop it in at the enforcement point in your own code, or run it as a gateway. Free tier to start.
| MCP server | shrike-mcp · npx shrike-mcp |
| TypeScript / JS SDK | shrike-guard · npm i shrike-guard |
| Python SDK | shrike-guard · pip install shrike-guard |
| Docs | shrikesecurity.com/docs |
| Website | shrikesecurity.com |
Apache-2.0 · Built for teams putting AI agents into real operations.
