agents-board: Instructions file for Codex

Agents Board

Project management for coding agents, exposed over MCP. Codex, Claude Code and Cursor can use the system to organize their own development (auto-inception). International product.

Skills

Two skills ship from a single host-neutral package, plugins/agents-board — Codex reads .codex-plugin/plugin.json, Claude Code reads .claude-plugin/plugin.json + .mcp.json, Cursor reads .cursor-plugin/plugin.json + mcp.json, and all three load the same skills/:

• agents-board — how to use the board: bind the repo to its project (record projectId + the auth flag in .agents-board.local.md), read and search it, keep a card's status honest, comment, write docs, work the inbox, attach diffs and images.
• plan — turn a new demand into sprints/stories/cards: clarify, offer research, clarify again, get approval, then write cards a zero-context agent can execute. All card creation goes through here.

The plugin deliberately stops at the board's edge. It ships no implementation, review, worktree or orchestration workflow — how code gets written is the user's process, with whatever skills they prefer. A change to a skill must not smuggle one back in.

Let the skills drive. What to do (active sprint, cards, backlog, comments, docs) is the source of truth and lives in the MCP, not here — query it through the Agents Board MCP tools. Don't duplicate state into this file.

Bind this repo to your board (local, not committed)

This is an open-source repo — every clone tracks its own work on its own Agents Board instance, so the project↔MCP binding is deliberately not committed here. Put yours in a gitignored .agents-board.local.md at the repo root. The coding agent must read it when it exists. For example:

This project in the MCP:

- **slug**: `<your-slug>`
- **keyPrefix**: `<KEY>` (cards are `<KEY>-1`, `<KEY>-2`…)
- **projectId**: `<your-project-id>`

Auth (diff capture) is **ON / OFF** in this deployment.