Citadel 🤖 Agent Open source

What it does

Citadel is an operating layer that sits around Claude Code or OpenAI Codex inside a single git repository. Instead of choosing between dozens of commands, you describe the outcome through one /do entry point, and Citadel routes the request to a matching workflow, keeps repository-local state between sessions, and records what happened so a fresh session knows the next step. It is aimed at work that spans more than one prompt: multi-step changes, parallel branches, and tasks that get interrupted.

What is inside

The release ships a catalogue of skills and hook scripts wired to the agent's lifecycle events. Work moves through five named states: request, run, evidence, needs you, and resume. Checks report passed, failed, blocked or unknown, and missing evidence is never promoted to success. Parallel work runs in isolated worktrees with ownership rules and shared discoveries. Operational state lives in a .planning/ folder and a few .claude/ and .citadel/ paths, separate from application code. A /unharness command produces a reviewable removal plan.

Works with

Claude Code and OpenAI Codex, both through their native plugin marketplaces. Node.js 22 or newer and a git repository are required.

How to install

For Claude Code, run these from the repository you want Citadel to manage:

claude plugin marketplace add SethGammon/[email protected] --scope local
claude plugin install citadel@citadel-local --scope local

For Codex:

codex plugin marketplace add SethGammon/Citadel --ref v1.4.3
codex plugin add citadel@citadel-local

Maintenance and safety

The project is MIT licensed, actively developed, and pins installs to tagged releases with checksum-verified archives for offline setups. The README is unusually candid: it states that its own experiments do not support a cost-savings claim and that it is not a sandbox, so it runs with the agent's normal permissions. Telemetry and state stay local by default. Read the security and privacy notes before using it on a sensitive repository.

Who should use it

Developers who run long or multi-session tasks with Claude Code or Codex and want durable handoffs, guarded workflows and a clear audit trail. For quick one-off edits it adds more structure than you need.