Skip to main content
Glama

Demo: the agent proposes an action, you apply it on Plan, Claude walks to the plan it works on, and the plan is checked with evidence

Why

Your coding agent (Claude Code, Codex) plans and builds fast. The cost is a plan you never reviewed and a "done" you cannot check. The Blueprint is a small desktop app that holds your whole project as one sheet, and your own agent talks to it over MCP:

  • It proposes, you apply. Everything the agent sends is a numbered proposal. Nothing changes in your data until you apply it on the Planejar (Plan) screen.

  • Done needs proof. A plan is only marked done with the command or test that was run and its result, which becomes a comment on the plan.

  • You see the work. Claude appears on the sheet, on the plan it is working on, with its subagents.

  • Small and private. One ~2.5 MB executable, a SQLite file on your machine, no account, no telemetry.

Related MCP server: e-worker-mcp

Quick start (about 5 minutes, Windows)

  1. Download Blueprint-<version>-instalador.exe from Releases and run it. It installs for your user only (no administrator) and, when you keep the connect task checked, adds the Claude Code plugin and registers the MCP in Codex.

  2. Open The Blueprint from the Start menu. A window and a tray icon appear.

  3. Open Claude Code (or Codex) in your project folder as you always do.

  4. Ask: "Plan this project in The Blueprint." The proposal arrives in the app; review it on Planejar and apply it.

  5. Ask: "Start the first action." As the agent works, the Claude walks to the plan on the sheet, and each plan is checked with its evidence.

Not on Windows, or no installer? Build it with Rust: cargo build --release, then blueprint serve runs the same app as a local web server at http://127.0.0.1:3000 (details).

How it works

Workspace
└─ Blueprint            the scope
   └─ Project           groups actions (optional)
      └─ Action         status, instructions
         ├─ Plan        title, description, done + evidence
         └─ Checkpoint  name, done

The agent reads the structure, proposes changes, reports progress and marks plans done through MCP tools (listar_blueprints, ler_blueprint, propor_blueprint, registrar_progresso, marcar_concluido and more; full list). You start an action yourself, with a confirmed click or the blueprint claude command, so the app never sends work to an agent on its own. Plans are sized for AI pace (0.5 to 4 hours) and the app measures how long each one really took. The model in detail.

Documentation

The guides are English first; each has a Portuguese version in docs/pt-BR/.

Install from source, performance

installer details, data folder, measured numbers

Claude Code plugin

skill, agent, hooks, status line, /blueprint:status, the terminal panel

MCP and other agents

tools, Codex, offline queue

Codex and other agents

connect, link the folder, what Codex sees without the plugin

Updates

background download, SHA-256 check, install only on your click

Server and commands

blueprint serve, environment variables, every subcommand

Development

cargo dev, tests, end-to-end checks

How it compares

Backlog.md, Task Master, Vibe Kanban, Plane

Troubleshooting

two processes, app closed, plugin missing, SmartScreen, update refused

FAQ

Does the app run my agent or spend tokens by itself? No. You start an action, and only then Claude Code runs, as usual.

Does it send my data anywhere? No account and no telemetry. The only network call is a GET to GitHub Releases to look for updates, which you can turn off (how).

Where is my data? In %APPDATA%\Blueprint\blueprint.db. blueprint exportar writes it to JSON.

Windows shows a SmartScreen or Defender warning. The installer is not code-signed yet, so that can happen on the first run. Compare its SHA-256 with the SHA256SUMS file in the same release; signed releases are on the roadmap.

macOS and Linux? Not yet; see the roadmap. The server mode works anywhere you can build it.

Codex? The MCP works today; a native integration (hooks, skills, status) is planned.

Roadmap, contributing, security

  • ROADMAP.md: what exists, what is being built, what comes next.

  • CONTRIBUTING.md: ideas and good first issues are welcome in Issues.

  • SECURITY.md: the MCP routes answer only on loopback; prototypes run in an opaque-origin sandbox.

License

MIT. The isometric kit descends from Anatomy (MIT); see NOTICE.md.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      A
      quality
      C
      maintenance
      A local-first MCP server that handles daily work tasks through your AI assistant: converts meeting notes into todos, manages todo lifecycle, tracks work hours, generates daily/weekly reports, organizes files via move-only operations, and diagnoses dev environments, all guarded by a human-maintained preview/apply safety model.
      25
      -
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI coding agents to take and progress tasks, hand in completed work with evidence, and enables users to publish, search, and retrieve guides, file bugs, manage inbox/board/activity/logs, and plan tasks over MCP.
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Connects Claude Code, Codex, Cursor, and other coding agents to a project's idea, feature topology, and ordered task queue over MCP, so they can answer planning jobs, claim, build, and tick off implementation tasks live in your repository, or assemble a shareable PRD.
      3 npm
      MIT