Cordum Documentation
Cordum is the Agent Control Plane — a deterministic governance layer between autonomous AI agents and the real systems they act on. Every tool call is evaluated by a policy engine, every decision is auditable, and the whole stack runs on your infra.
:::tip First time here? Skip the plumbing tests. Start with Guardrails Demo → — 2 minutes to watch the Safety Kernel catch a dangerous request, suggest a remediation, and route an approval. :::
Read this first
Concepts
Orient yourself on the architecture.
- Concepts Overview — the map of what Cordum governs
- Glossary — one-line definitions for every term
- System Overview — the service graph
- Agent Protocol (CAP) — the wire contract
Getting Started
Get the stack running locally.
- Guardrails Demo — recommended first run
- Minimal Stack Smoke Test — confirm every service is reachable
- Docker Setup — local compose walkthrough
- Local E2E Testing — integration test flow
Governance deep-dives
How the policy engine actually works.
- Safety Kernel — decision flow, rule types
- Output Policy — post-execution redaction
- Output Safety — scanner pipeline
- Audit Chain — SIEM export, integrity verification
Build on Cordum
Integrate your agents and tools.
- Framework Integrations — LangChain, CrewAI, AutoGen, OpenAI Agents, LangGraph
- Packs — reusable policy + workflow bundles
- MCP Server — Model Context Protocol support
- CAP Repo — protocol spec + SDKs (Go, Python, Node, C++)
Deploy and operate
Everything you need to run Cordum in production — install, Helm, Kubernetes, config reference, horizontal scaling, performance tuning — lives under Operations.
API Reference
- API Overview — REST endpoints
- gRPC Services — Safety Kernel, Context Engine
- CLI Reference —
cordumctl - WebSocket Streaming — real-time events
Three surfaces for the protocol
The Cordum Agent Protocol (CAP) is the wire contract between the platform and your agents. It lives in three places, each for a different audience:
| Surface | What you'll find |
|---|---|
| cordum.io/protocol | Positioning — "MCP for agent behavior", business case, adoption notes |
| This page → /concepts/agent-protocol | How the control plane consumes CAP (subjects, pointers, state) |
| github.com/cordum-io/cap | Canonical spec, conformance tests, SDKs for Go / Python / Node / C++ |