Cordum v1.1.0: Open Agent Control Plane for AI Governance

What changed

Cordum has released version 1.1.0 of its open-source agent control plane, a system designed to provide governance, safety, and trust for autonomous AI agents. This release focuses on enabling pre-execution policy enforcement, implementing approval gates, and maintaining detailed audit trails for agent actions. The platform is built to be framework-agnostic, supporting integrations with popular tools like LangChain, CrewAI, and MCP, as well as any other framework.

Key components of the Cordum platform include an API gateway, scheduler, safety kernel, workflow engine, context engine, and a dashboard. The latest release offers multiple installation paths, including a one-command script for installing from source (./tools/scripts/quickstart.sh) and running pre-built Docker images (docker compose pull && docker compose up -d). Both methods require Docker Desktop v4+ or Docker Engine v20.10+ with Compose v2, and Go 1.26.3+ for initial certificate generation. The source installation sets up the full stack with auto-generated secrets and certificates, culminating in a smoke test that verifies an approval workflow. The Docker image approach allows users to pull and run services directly, with options to customize API keys and passwords. For Kubernetes deployments, a Helm chart is available.

Cordum also provides container images for its various services, including api-gateway, scheduler, safety-kernel, workflow-engine, context-engine, mcp, and dashboard, available on GHCR and Docker Hub. These images are multi-arch, supporting both linux/amd64 and linux/arm64. Image signatures are available via cosign for verification. The release details specific ports used by each service (e.g., 8082 for Dashboard, 8081 for API Gateway) and provides troubleshooting guidance for common issues like port conflicts, Docker memory allocation, and authentication problems.

A notable feature highlighted is Cordum Edge, described as a compliance firewall for local AI-agent actions, including Claude Code tool calls. This component aims to deny risky actions before they execute, require approval for edits, and export evidence bundles. A quickstart demo is available, showcasing a three-verdict workflow that demonstrates allowance, denial, and required approval for different agent tasks.

Why it matters for builders

For AI builders, Cordum 1.1.0 offers a robust framework to implement governance and safety measures for their autonomous agents. The ability to enforce policies *before* an agent acts, coupled with approval mechanisms, provides a critical layer of control that can prevent unintended consequences or costly mistakes. This is particularly relevant for agents performing sensitive operations or interacting with external systems.

Practical impact

Developers can integrate Cordum into their existing agent workflows by cloning the repository and running the quickstart script or Docker Compose commands. The platform's compatibility with frameworks like LangChain and CrewAI means minimal disruption for many users. Testing the provided demo workflows, such as the three-verdict scenario, can help builders understand how Cordum's safety kernel and approval gates function in practice. For those deploying to production, exploring the Helm chart and image verification processes is recommended.

Caveats and source limits

The source material indicates a "fresh release" for version 1.1.0 but does not provide specific performance benchmarks or independent validation of its governance capabilities. While the platform is described as open-source, the exact license details beyond "Other" are not specified in the provided excerpt, which could impact commercial use. The documentation for enabling user authentication with a password requires specific character complexity, but the exact enforcement rules beyond the general description are not detailed. The Cordum Edge component's integration depth with specific local AI models like Claude Code is mentioned but not fully elaborated upon in terms of all supported models or specific configuration nuances.