Project releases

Agentic DevOps v2 plans deployments instead of running them

The v2 rewrite replaces model driven shell execution with a deterministic local planner. It turns a small deployment specification into Docker, Kubernetes and CI files, shows exactly what changes, and never touches your cluster.

GitHub activity: · Published:

What it is

Give it a small deployment specification and it produces a Dockerfile, a Deployment, a Service, a NetworkPolicy and a GitHub CI template. You review the diff, then hand the files to whatever actually deploys.

It never runs deployment commands, reads cloud credentials or calls a model. That is a deliberate reduction from the historical version.

What changed

Pull request #4 is the v2 revival with validated artifacts, a shared CLI and MCP surface, and a full repository harness. A follow up, #5, removed duplicate rollback validation while proving equivalent output with benchmark evidence.

Policy validation rejects mutable image tags, unknown fields, privileged ports, excessive resources and any edited bundle.

Get started

Node.js. Clone the repository first — the MCP entry point is a local file:

git clone https://github.com/ruvnet/agentic-devops.git
cd agentic-devops
npm install

Then register the local server with your host:

{
  "mcpServers": {
    "agentic-devops": {
      "command": "node",
      "args": ["/absolute/path/agentic-devops/src/mcp.mjs"]
    }
  }
}

Expected result: seven actions available through both the CLI and MCP, sharing one policy.

Running tests through the CLI or MCP requires the local operator to set AGENTIC_DEVOPS_ALLOW_TESTS=1. The subprocess has a 30 second deadline, a stripped environment, and callers cannot choose commands or paths.

Use it today

Practical case: a service that needs consistent deployment files across ten repositories. Input is the specification; the workflow is generate, review, commit; the output is a bundle that policy has already checked.

Acceptance test: point the image tag at a mutable tag such as latest. Generation should be rejected.

Push it further

Experimental commentary. The interesting frontier is reversibility — restoring a previous artifact bundle as a first class operation rather than a git revert. The dependency is that the bundle, not the cluster, is treated as the source of truth.

Falsifiable test: restore a prior bundle and diff it against what you generated originally. Any drift means something outside the bundle is mutating your deployment.

Read the original on GitHub

Revive v2: validated deployment artifacts, CLI/MCP and full repository harness (#4)

Agentic DevOps repository

Back to the newsroom