PlanForge
emiperez95/planforge
Co-planejador com diagramas Mermaid editáveis em loop fechado
Ver no GitHubAgent para co-planejamento humano-IA com loop fechado em navegador. Gera planos HTML com diagramas Mermaid editáveis em tempo real, você edita e envia de volta, o agent captura as mudanças e continua iterando. Elimina o atrito entre edição visual e refinamento automático.
Instalação
npx convez-ai emiperez95/planforge1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Especificação de arquitetura com diagramas interativos
- Planejamento de projetos técnicos com iteração visual
- Documentação viva em HTML com refinamento colaborativo
- Prototipagem ágil de sistemas com feedback contínuo
Veredito

Solução bem concebida para fechar o gap entre edição visual e iteração assistida. Requer Python 3.10+ local e Claude Code; ainda em v0.1.0, sem distribuição em marketplace. Valor real está em workflows que mistem prosa + diagramas estruturados.
README
A tradução para o português deste README ainda está na fila.
planforge
Closed-loop human-AI coplanner for spec + diagram authoring. The agent generates an HTML plan with editable Mermaid diagrams, you edit and send, the agent picks up the changes and continues. Built for Claude Code.
Status
✅ v0.1.0 — first functional release. The loop runs end-to-end. See plan.md for the roadmap and what's next.
What it is
A Claude Code plugin that provides a forge skill (invoked as /planforge:forge). It runs a closed feedback loop between the agent and you:
- The agent generates an HTML plan — prose sections + Mermaid diagrams — and writes it to disk.
- The skill launches a local stdlib HTTP server in the background and opens the page in your browser.
- You read, edit the prose, and edit the Mermaid diagrams (they re-render live). Meanwhile you can keep chatting with the agent in parallel.
- You click Send to agent (or Approve & finalize). The browser POSTs the full edited plan state to the local server.
- The server writes the response to disk and exits — the background task completing notifies the agent automatically (no polling).
- The agent picks up your edits, refines the plan, and serves the next version (your tab auto-advances). The loop continues until you Approve.
No API keys to manage — the plugin reuses the agent's existing Claude Code session.
Why
Most diagram-as-spec workflows force you into either pure-text iteration (no visual editing) or out-of-band tooling (Figma, then paste back into chat). planforge closes the loop: rendered diagrams you can edit, with automatic round-trip back to the agent.
Built originally to support empirical work on diagram-driven AI coding pipelines. It generalizes to any human-AI co-creation task whose artifact mixes structured parts (diagrams, code, JSON) with prose.
Requirements
- Claude Code
- Python 3.10+ (the local server is stdlib-only — no pip installs)
- Internet access for the first render (Mermaid loads from a CDN; local vendoring is planned)
Install
planforge isn't on a marketplace yet. For v0.1.0, load it from a local clone:
git clone https://github.com/emiperez95/planforge.git
claude --plugin-dir ./planforge
--plugin-dir loads the plugin for that Claude Code session. Marketplace distribution is planned for a later release.
Usage
In a Claude Code session with the plugin loaded:
/planforge:forge plan a URL shortener service
The agent drafts a plan and opens it in your browser. Edit the prose and Mermaid diagrams, then:
- Send to agent — hand your edits back; the agent refines and serves the next turn (your tab auto-advances).
- Approve & finalize — end the loop; the agent saves the final
converged.html.
Per-run artifacts (each turn's HTML + your responses, plus the final converged.html) are written under /tmp/planforge-<run_id>/.
Development
See plan.md for the phased roadmap. Phase 2 (functional v0.1.0) is complete; remaining work and deferred features are tracked there.
Dogfooding it globally while you develop
To use planforge across your other projects without a per-session --plugin-dir
or a copied marketplace install, symlink the skill into your personal skills
directory:
ln -s "$(pwd)/skills/forge" ~/.claude/skills/planforge
It then loads in every session as /planforge, and edits to the repo are live —
no reinstall. The skill resolves server.py / template.html relative to its own
base directory, so the same SKILL.md works both as this personal-skill symlink
and as a namespaced plugin (/planforge:forge). Skills load at session start, so
open a fresh session to pick it up.
Contributions follow Conventional Commits. Enable the commit-msg hook:
git config core.hooksPath .githooks
The CHANGELOG is regenerated from commit history via git-cliff at each release.