SpecManager – a full agile team for founders, as a Claude Code plugin

SpecManager, a Claude Code plugin that replaces ad-hoc vibe coding with a gated spec-driven development lifecycle, has been released as an MIT-licensed open-source tool. The plugin enforces a PRD-to-walkthrough workflow with a localhost kanban board backed by plain markdown, targeting single-user, fully local use for founders. It is available via GitHub and requires Node 20+ and Claude Code.

Spec-driven development for Claude Code. A Claude Code plugin that replaces ad-hoc vibe coding with a gated lifecycle — PRD → Architecture → Design → Plan + tasks → Build → Walkthroughs — rendered as a localhost kanban board backed by plain markdown in your repo. Claude drafts each stage from the previous approved one and your existing codebase ; you edit and approve in the board. Every artifact is git-tracked markdown, so it diffs, reviews, and travels with the code. Single-user, fully local, no auth, bound to 127.0.0.1 . specmanager.org https://specmanager.org · MIT licensed 1. Add the marketplace from GitHub /plugin marketplace add joanseg/specmanager 2. Install /plugin install specmanager@specmanager 3. Reload plugins /reload-plugins 4. Reconnect MCP server /mcp If plugin:specmanager:specmanager ✘ failed select and click enter to reconnect 4. In your project, scaffold SpecManager and open the board /specmanager-init /specmanager-board Example of a Specmanager board: That's it — no build step. The compiled server and UI are committed, and a SessionStart hook installs runtime dependencies into the plugin's data dir on first launch. Requirements: Node 20+ and Claude Code. The board runs at http://127.0.0.1:4317 change the port in the plugin's user config . Each feature is a row on the board that flows left to right through the lifecycle. You drive it with slash commands; Claude does the drafting via dedicated subagents, and you approve each stage in the board before the next unlocks. /specmanager-interview "Checkout corridor" OPTIONAL pre-PRD interview — extracts & stress-tests the idea /specmanager-prd "Checkout corridor" create the feature + draft its PRD → approve in board /specmanager-architecture <feature draft the Architecture → approve in board /specmanager-design <feature OPTIONAL high-fi HTML mockups → approve /specmanager-plan <feature plan.md + phased tasks → approve in board /specmanager-build <feature next build one phase at a time /specmanager-walkthrough <feature <phase per-phase walkthrough /specmanager-walkthrough <feature final feature-level roll-up /specmanager-board open the kanban board anytime <feature is the feature's slug or id reported by /specmanager-prd when it creates the feature . /specmanager-interview is optional: a multi-turn interview in your Claude session that pulls the idea out of your head and challenges it using the gstack office-hours https://github.com/garrytan/gstack/tree/main/office-hours forcing-question method before any PRD exists. The result can be stored as interview.md in the feature's PRD folder — shown as a chip on the board, never gating anything — and the PRD drafter grounds itself in it when present. | Stage | Produces | Unlocks when | |---|---|---| PRD | press-release.md , prd.md | always the entry point | Architecture | architecture.md | PRD approved | Design optional | high-fidelity HTML mockups | PRD approved runs in parallel with Architecture | Plan | plan.md + task records tasks.json , rollup | Architecture approved and Design, if one exists, approved | Build | no doc — it's execution | Plan approved | Walkthroughs | <slug .md per phase + a final roll-up | the phase's tasks are all done | Gates are enforced in shared code, not in prompts — Claude cannot draft a stage whose gate is closed. Plans are organised into phases , each a testable, runnable increment; tasks carry a Fibonacci complexity score and anything over 3 must be split. /specmanager-build builds one phase and stops at its boundary. approved is a commitment, not a freeze. Reopen an approved doc and every downstream doc that depended on it gets a non-blocking stale badge — a "the basis changed, review me" signal. Regenerate or re-approve to clear it. /specmanager-board opens a grid: one row per feature, one column per stage PRD · Architecture · Design · Plan · Build · Walkthroughs . It updates live over websockets as Claude writes docs or you edit them. - Each cell is a card showing status draft / approved , a stale badge, and whether it was generated by agent or human. - Clicking a card opens a doc panel drawer over the dimmed board to read, edit, and approve. Design docs get a CodeMirror HTML editor with a live sandboxed <iframe preview. - The Build cell groups tasks under collapsible phase headers with per-phase progress bars instead of a document. - The Walkthrough column shows one card per phase plus a final roll-up card that stays locked until every phase walkthrough is approved. Edits you make in the board and writes Claude makes both flow through the same shared logic layer, so the two views never drift. AI writes use optimistic concurrency, so a stale agent write can't clobber your manual edits. Everything is markdown with YAML frontmatter under your project, version-controlled by git: <your-project / CLAUDE.md carries a managed SpecManager block between markers docs/DESIGN.md managed design-system spec inferred from your UI .claude/specs/ manifest.json rebuildable board index derived from frontmatter features/<slug / feature.json prd/ press-release.md, prd.md architecture/ architecture.md design/ <mockups .html plan/ plan.md, tasks.md, tasks.json walkthroughs/ <slug .md Frontmatter is authoritative; manifest.json is a cache you can delete and rebuild. The managed regions in CLAUDE.md and DESIGN.md live strictly between < -- specmanager:start -- / < -- specmanager:end -- markers — nothing outside them is ever touched. The plugin ships its compiled server/dist and ui/dist , so users install with no build step. Rebuild before committing source changes: cd plugins/specmanager/server npm install npm run build npm run selftest core flow against a tmp dir npm run selftest-board boots board: REST + WS + file watcher npm run selftest-phases phase rollup + Fibonacci ≤3 validation npm run selftest-build per-phase gates + walkthrough storage npm run smoke-mcp MCP wire protocol + tools registered cd ../ui npm install npm run build → ui/dist, served by the board server Then reinstall in a test repo: /plugin marketplace update specmanager /plugin install specmanager@specmanager /reload-plugins CLose all Claude Code sessions. Then open CLaude Code sesion and reconnect mcp: /mcp select specmanager:specmanager, click enter and reconnect Tech stack: Node 20+, TypeScript, MCP stdio transport, Fastify + ws , chokidar , gray-matter , zod , React 18 + Vite, CodeMirror 6. One shared @specmanager/core library backs both the MCP server Claude's interface and the board server the UI's interface ; one MCP process boots the board. Repo layout: the marketplace manifest is at the repo root .claude-plugin/marketplace.json ; the plugin itself lives in plugins/specmanager/ . Design docs are under docs/ the original full spec is archived at docs/temp/original-specs/architecture-and-spec.md . — select it and reconnect. If it persists, fully restart: quit Claude /mcp shows specmanager failed Ctrl-C twice , claude daemon stop , kill stragglers ps aux | grep mcp.js , confirm the port is free lsof -nP -iTCP:4317 -sTCP:LISTEN , then relaunch claude from your project. Board won't open — the MCP process boots the board server on startup; if it isn't running, restart your Claude session. /specmanager-board reports the URL so you can open it manually.— do the full restart above rather than retrying the reload. /reload-plugins reports a load error - Website: https://specmanager.org https://specmanager.org