SethGammon/Citadel
The operating layer for Claude Code + OpenAI Codex: persistent project memory, intent routing, safety hooks, cost telemetry, and parallel agent fleets.
What it solves
Citadel provides an operating layer for coding agents like Claude Code and OpenAI Codex, ensuring they work reliably across real-world repositories. It addresses the problem of lost context between sessions, unclear workflow choices, and the risks associated with multi-step changes in large codebases.
How it works
Citadel acts as a coordination layer that sits on top of the agent's runtime. It uses a natural-language entry point (/do) to route requests to specific skills or workflows. It implements a lifecycle that includes routing, execution with repository safeguards (hooks), and recording results and handoffs to the repository itself, which remains the source of truth. This allows work to survive interruptions and be resumed across different sessions.
Who it’s for
Developers using AI coding agents (specifically Claude Code and OpenAI Codex) who are managing complex, multi-step tasks that extend beyond a single prompt.
Highlights
- Durable State: Preserves project state, decisions, and discoveries in repo-local files so work can be resumed across sessions.
- Unified Entry Point: Uses
/doto select and run appropriate workflows based on natural language requests. - Operation Fork: Allows running the same objective through isolated branches for comparison between different AI runtimes (e.g., Claude Code vs. Codex).
- Repository Safeguards: Employs lifecycle hooks to apply repository rules and gate consequential actions.
- Mission Control: A dashboard for managing campaigns, agents, operations, and handoffs.
- Local-First: Project state and telemetry stay local by default.