Awesome AI Agents › AI Agent Gateway

cordum-io/cordum

⭐ 504 Go repository created 2026-01-11

Cordum is an agent control plane that puts a policy checkpoint in front of agent actions rather than inside the agent. Its safety kernel evaluates every submitted job and returns one of three verdict classes, allowing it, blocking it outright, or holding it for human approval, and a bundled demo fans a single workflow out to three topics to exercise all three decisions in under a minute. The stack ships as seven Go services, an API gateway with mutual TLS, a scheduler, the safety kernel, a workflow engine, a context engine, an MCP service and a web dashboard, backed by NATS and a TLS-secured Redis. Jobs are submitted over an HTTP API with an API key and a tenant header, so the design assumes multiple tenants sharing one deployment. Deployment is packaged for operators: a quickstart script stands up the whole stack from source with generated secrets, auto-provisioned certificates and a post-deploy smoke test that runs a real approval workflow, published multi-architecture images cover the docker compose path, and a Helm chart on Artifact Hub covers Kubernetes. Every release image is signed with keyless cosign, and the documentation includes verification commands and guidance on keeping secrets out of Helm values. A separate component, Cordum Edge, applies the same governance to a developer terminal by hooking Claude Code tool calls, denying risky actions before they run and requiring approval for edits. The project implements its own agent control protocol, published as a separate repository. It is licensed under BUSL-1.1, so it is source-available rather than open source. The audience is platform and compliance teams that need an audit trail and enforceable limits before letting autonomous agents touch production systems.

https://github.com/cordum-io/cordum

ai-orchestrationai-safetyautonomous-agentsgovernancellm-agentsworkflow-enginegolangkubernetespolicy-enforcementhuman-in-the-loop