MCP server de kitbash

Kitbash

Kitbash is an open-source MCP server that tells your AI agent which parts of your idea already exist, which to adapt, and which to just write. It deletes the repos that don't exist before you ever see them. Built openly, for everyone, forever free.

claude mcp add kitbash -- npx -y github:Open-Dev-Society/kitbash

No API key. No sign-up. Works in Claude Code, Cursor, Claude Desktop, VS Code, Codex, Windsurf, and any other agent that speaks MCP.

📋 Table of Contents

1. ✨ Introduction
2. 🌍 Open Dev Society Manifesto
3. 🔍 What It Looks Like
4. 🔋 Features
5. 🤸 Quick Start
6. 💬 Usage
7. ⚙️ How It Works
8. 🔐 GitHub Token
9. ☁️ Self-Hosting
10. 🧱 Project Structure
11. 🧪 Scripts & Tooling
12. 🤝 Contributing
13. 🛡️ Security
14. 📜 License
15. 🙏 Acknowledgements

✨ Introduction

Your coding agent will happily rewrite a PDF parser for you. It'll take twenty minutes and it'll look completely correct.

It won't be. PDF parsing is years of edge cases that appear in no spec: broken xref tables, CID fonts with no ToUnicode map. The same goes for timezone math, OAuth, Unicode normalization, container muxing, and rate limiting under contention. You find out in production.

The opposite trap is just as common: pulling in a dependency to save twenty lines of glue.

Search can't help. It ranks by popularity, not fit, so the right small library stays buried on page four. And when a model will write anything in twenty minutes, rebuilding feels free, so nobody looks first.

Kitbash is the ten-second judgment a senior developer makes before writing anything, turned into a tool. It doesn't make you ship faster. It makes you ship less. Four components you never build beat four components you build quickly.

Every part of your idea gets one verdict:

Verdict	Meaning
🟢 BORROW	A maintained library exists and the correctness was hard-won. Rewriting it is the mistake. Take the dependency.
🟡 KITBASH	A good reference exists but isn't a clean fit. Read it, adapt the approach, leave the dependency.
🔵 WRITE	Generic enough that your agent should just write it. A component you never take on is one you never maintain.

🌍 Open Dev Society Manifesto

We live in a world where knowledge is hidden behind paywalls. Where tools are locked in subscriptions. Where information is twisted by bias. Where newcomers are told they're not "good enough" to build.

We believe there's a better way.

• Our Belief: Technology should belong to everyone. Knowledge should be open, free, and accessible. Communities should welcome newcomers with trust, not gatekeeping.
• Our Mission: Build free, open-source projects that make a real difference:
  • Tools that professionals and students can use without barriers.
  • Knowledge platforms where learning is free, forever.
  • Communities where every beginner is guided, not judged.
  • Resources that run on trust, not profit.
• Our Promise: We will never lock knowledge. We will never charge for access. We will never trade trust for money. We run on transparency, donations, and the strength of our community.
• Our Call: If you've ever felt you didn't belong, struggled to find free resources, or wanted to build something meaningful — you belong here.

Because the future belongs to those who build it openly.

🔍 What It Looks Like

Real output for "a CLI that ingests podcast RSS, transcribes episodes, and makes them searchable." Every repo returned 200 from the GitHub API seconds before it was printed.

#	Component	Verdict	Part	Stars	License
01	Podcast feed parsing	🟢 BORROW	gpodder/podcastparser	144	ISC
02	Speech-to-text	🟢 BORROW	SYSTRAN/faster-whisper	24,000	MIT
03	Transcript search index	🟡 KITBASH	simonw/sqlite-utils	2,100	Apache-2.0
04	Episode pipeline and CLI	🔵 WRITE	yours: about a hundred lines	—	—

Two things to notice:

• A 144-star repo beat a 2,400-star one for the lead slot. podcastparser is the parser the gPodder client actually uses. It streams instead of building the XML tree, and it already normalizes the itunes:duration mess. feedparser is the popular fallback, ranked second. Surfacing the exact fit GitHub search buries is the whole product.
• It told you to write part 04 yourself, and argued for it: "The interesting state here is domain state, not queue state. Transcription is minutes per episode, so the only checkpoint that matters is a row per episode with a status column. Any task-queue dependency would still leave you writing that row yourself." A tool that recommends a repo for everything is a tool nobody believes, so WRITE is a required outcome.

🔋 Features

• Decomposition with judgment
  • Splits an idea into 3–6 functionally distinct components, refusing to split one library's job three ways
  • A built-in hazard list (PDF, crypto, timezones, OAuth, codecs, Unicode, rate limiting, …) steers hard-won domains toward BORROW
  • WRITE is mandatory. A slate with zero WRITE components gets sent back for review
• Fit over popularity
  • The under-starred exact fit ranks ahead of the popular general-purpose library
  • The agent's own ordering leads, and repo health can only demote. Stars are a weak signal on purpose
  • Target stack and license constraints shape every recommendation
• Live verification
  • Every owner/name is resolved against the GitHub API. Repos that don't exist are dropped and counted in the report
  • Archived, disabled, unlicensed, forked, renamed, and abandoned repos are flagged
  • BORROW on a repo nobody has touched in two years gets a loud warning
  • If every candidate for a component dies, the agent is sent back for replacements before it can answer
  • About 950 ms to verify 10 repos, in parallel
• Honest under failure
  • Rate limits and outages are reported as "couldn't check", never as "doesn't exist"
  • An offline snapshot fills network gaps, and every cached record is labelled cached
• Install it your way
  • Claude Code plugin, local MCP server, remote connector, or plain agent skill with no MCP at all
  • Zero config: picks up your existing gh login

🤸 Quick Start

Prerequisites: Node.js 18+. A GitHub token is optional but recommended (see GitHub Token).

Claude Code plugin

/plugin marketplace add Open-Dev-Society/kitbash
/plugin install kitbash@kitbash

MCP server

claude mcp add kitbash -- npx -y github:Open-Dev-Society/kitbash     # Claude Code
codex mcp add kitbash -- npx -y github:Open-Dev-Society/kitbash      # Codex CLI

Add to ~/.cursor/mcp.json, claude_desktop_config.json, or your client's MCP config:

{
  "mcpServers": {
    "kitbash": { "command": "npx", "args": ["-y", "github:Open-Dev-Society/kitbash"] }
  }
}

Add to .vscode/mcp.json:

{
  "servers": {
    "kitbash": { "type": "stdio", "command": "npx", "args": ["-y", "github:Open-Dev-Society/kitbash"] }
  }
}

Native Windows needs npx wrapped in cmd:

{ "command": "cmd", "args": ["/c", "npx", "-y", "github:Open-Dev-Society/kitbash"] }

The first launch installs and builds the package, which takes about 25 seconds. After that it starts in about 3.

Remote connector (claude.ai, ChatGPT, any client that takes a URL)

Kitbash speaks Streamable HTTP with no auth. Run your own endpoint (see Self-Hosting) and add it as a custom connector:

https://YOUR-DEPLOYMENT/mcp

Agent skill (no MCP required)

npx skills add Open-Dev-Society/kitbash

This installs skills/kitbash/SKILL.md into Claude Code, Cursor, Codex, and every other agent skills.sh supports. The skill carries the same rubric and shells out to the same fact-checker — pinned to an exact commit, see Security — so it works with no MCP connection at all.

💬 Usage

Describe something you were about to build:

Use kitbash: a desktop app that watches a folder and makes scanned PDFs searchable. TypeScript, MIT.

Saying "use kitbash" helps. Otherwise the agent sometimes answers from its own knowledge and skips the tool. A run takes about a minute. Naming your stack and license is optional, but it sharpens every recommendation.

⚙️ How It Works

There is no model inside this server. That's the design, not a shortcut.

Recall is the product: knowing that a 144-star parser exists at all. The best recall available is in the model already running your session, which is far larger than anything this server could afford to host. So Kitbash doesn't try to out-remember it. It aims that recall, then stops it from lying.

sequenceDiagram
    participant U as You
    participant A as Your agent
    participant K as Kitbash
    participant G as GitHub API
    U->>A: "Build me X"
    A->>K: kitbash(idea)
    K-->>A: rubric (never an answer)
    Note over A: decomposes the idea, assigns verdicts,<br/>recalls candidate repos
    A->>K: kitbash_verify(slate)
    K->>G: GET /repos/{owner}/{name}, in parallel
    G-->>K: exists? archived? license? last commit?
    K-->>A: verified report + next_action
    A-->>U: the report, and only the report

1. kitbash returns a rubric and nothing else. It never returns anything that could pass for an answer. If it did, the agent would show it and skip verification.
2. kitbash_verify fact-checks the slate. It drops the repos that don't exist, attaches health signals, ranks the survivors, and renders the report the user sees.

A model asked for repo names produces plausible ones, and invented names look completely real in a chat window. Step 2 is the only reason to trust step 1: every repo in the final report was confirmed to exist seconds before you saw it.

🔐 GitHub Token

Kitbash works without a token, but GitHub limits unauthenticated use to 60 requests an hour. It picks up a token from these sources, in order:

1. GITHUB_TOKEN
2. GH_TOKEN
3. gh auth token (if the GitHub CLI is logged in, there's nothing to configure)

A token needs no scopes. Without one, private repos look the same as nonexistent ones (both 404), so they'll be reported as not found.

☁️ Self-Hosting

The same package serves Streamable HTTP for remote connectors:

git clone https://github.com/Open-Dev-Society/kitbash.git && cd kitbash
npm ci
GITHUB_TOKEN=ghp_… PORT=3000 npm run start:http    # → http://localhost:3000/mcp

Deploy it anywhere that runs Node 18+ (Render, Railway, Fly.io, a VPS) with build command npm ci and start command npm run start:http.

• Stateless: a fresh server per request. Nothing from one caller reaches the next.
• Bounded: at most 12 components and 5 repos per component, so no single caller can drain the shared token.
• Set GITHUB_TOKEN: every caller shares its 5,000 requests an hour, and verification makes one request per candidate repo.

🧱 Project Structure

src/
├── index.ts            entrypoint: stdout guard, mode dispatch (stdio · --http · verify)
├── server.ts           MCP server assembly; stdio and Streamable HTTP transports
├── rubric.ts           the rubric the agent follows — this file is the product
├── tools/
│   ├── plan.ts         `kitbash` tool: hands out the rubric
│   └── verify.ts       `kitbash_verify` tool and CLI: fetch, assess, rank, report
├── github/
│   ├── client.ts       REST transport and error classification (404 vs rate limit vs auth)
│   ├── repos.ts        batched, deduplicated repo fetches
│   └── token.ts        token resolution ladder
├── score.ts            health assessment, disqualification, ranking
├── render.ts           the markdown report
├── cache.ts            offline snapshot
├── types.ts            the frozen contract every module depends on
└── scripts/            snapshot pre-warm, rubric example check, skill generator
skills/kitbash/         the agent skill (generated from rubric.ts)
plugin/                 the Claude Code plugin
.claude-plugin/         the plugin marketplace manifest
server.json             MCP Registry manifest
site/                   landing page

🧪 Scripts & Tooling

Command	What it does
npm run build	Compile TypeScript to build/
npm run typecheck	Type-check without emitting
npm start	Build and run the stdio server
npm run start:http	Run the Streamable HTTP server on $PORT (default 3000)
npm run skill	Regenerate skills/kitbash/SKILL.md from the rubric
npm run check:examples	Confirm every repo named in the rubric still exists
npm run snapshot -- owner/name …	Pre-warm the offline snapshot
kitbash-mcp verify < slate.json	Verify a slate from the shell

🤝 Contributing

Contributions are welcome, from code to docs to a better worked example in the rubric. Start with the Open Dev Society contribution guidelines, then look for issues labelled good first issue.

git clone https://github.com/Open-Dev-Society/kitbash.git && cd kitbash
npm install
claude mcp add kitbash-dev -- node "$(pwd)/build/index.js"

Things that will bite you in this codebase:

• stdout is the JSON-RPC channel. One console.log anywhere corrupts the protocol and the server dies silently. Log with console.error. src/index.ts reroutes console.log/info/debug to stderr as a backstop.
• Local imports need explicit .js extensions. Under Node16 ESM resolution, leaving them out typechecks clean and crashes at runtime.
• src/types.ts is the frozen contract. Every other module depends on it and nothing else.
• src/rubric.ts is the product. After editing it, run npm run skill and npm run check:examples, and commit the regenerated skill.

🛡️ Security

Please don't open public issues for security vulnerabilities. Report them privately via GitHub security advisories or email opendevsociety@gmail.com.

What it touches. Kitbash's only network requests are read-only GET /repos/{owner}/{name} calls to api.github.com. It writes nothing outside its own snapshot file.

Your GitHub token. A token is optional — without one you get GitHub's 60 requests an hour. When GITHUB_TOKEN, GH_TOKEN, or gh auth token is available, it is sent as an Authorization header to api.github.com and to nothing else, needs no scopes, and is never logged or written to disk.

Pinned execution. The skill's verify command pins an exact commit instead of tracking main, so the code cannot change under whoever runs it. Once the npm package is published, npx -y kitbash-mcp@<version> pins an immutable published version instead.

Third-party audits. The skill is scanned by Gen Agent Trust Hub, Socket, and Snyk; current results are on its skills.sh listing.

📜 License

Kitbash is licensed under the GNU General Public License v2.0.

🙏 Acknowledgements

• Model Context Protocol and its TypeScript SDK
• The GitHub REST API, which makes the fact-check possible
• Zod for the schemas at every trust boundary
• Every maintainer of the 144-star library that does exactly one thing well. This tool exists to send people your way.

---

If Kitbash saved you from building something, give it a ⭐. It helps other developers find it.

Built openly by Open Dev Society.