mcpknife

The Swiss Army knife for MCP servers. One command to generate, transform, and add UIs to MCP servers — with Unix pipes.

mcpknife boot --prompt "Hacker News API https://github.com/HackerNews/API" \
  | mcpknife mod --prompt "combine new + best into get_trending" \
  | mcpknife ui \
  | mcpknife export --output-dir ./my-server

That's it. Four stages, one pipeline, zero glue code. You go from API docs to a standalone MCP server with custom tool composition and auto-generated UI in a single command.

What it does

mcpknife unifies three tools under one CLI, plus an export command:

| Command | What it does | Underlying tool | |---------|-------------|-----------------| | mcpknife boot | Generate an MCP server from a prompt + API docs | mcpboot | | mcpknife mod | Transform, rename, hide, or combine tools on an MCP server | mcpblox | | mcpknife ui | Auto-generate interactive UIs for MCP server tools | mcp-gen-ui | | mcpknife export | Dump the pipeline as a standalone MCP server project | built-in | | mcpknife deploy | Deploy an exported project to the cloud | built-in |

Each stage reads an upstream server URL from stdin and writes its own URL to stdout. Standard Unix pipes connect them. export is a terminal stage that crawls the pipeline via _mcp_metadata and generates a self-contained Node.js project.

Install

npm install -g mcpknife

One install. Three tools. No extra setup.

Usage

Generate an MCP server from API docs

mcpknife boot --prompt "Yahoo Finance API https://finance.yahoo.com/api"

Point it at API documentation and describe what you want. mcpknife boots an MCP server with tools derived from the API.

Transform tools on an existing server

mcpknife mod --upstream "npx some-mcp-server" --prompt "hide write tools, expose read-only"

Reshape any MCP server's tools using natural language. Rename, combine, filter, or create synthetic aggregation tools.

Add UI to any MCP server

mcpknife ui --upstream-url http://localhost:3000/mcp

Auto-generates interactive UIs for every tool on the server.

Export a standalone server

mcpknife boot --prompt "Free Dictionary API https://dictionaryapi.dev/" \
  | mcpknife mod --prompt "Create one tool: synonyms and antonyms" \
  | mcpknife ui \
  | mcpknife export --output-dir ./dict-server

Append mcpknife export to any pipeline to dump a self-contained Node.js project. The exported server runs independently — no mcpknife, mcpboot, mcpblox, or mcp-gen-ui required. Just cd dict-server && npm install && node server.js.

Deploy to the cloud

# Export and deploy to Fly.io in one pipeline
mcpknife boot --prompt "Free Dictionary API https://dictionaryapi.dev/" \
  | mcpknife export \
  | mcpknife deploy --name dict-api

# Deploy an existing exported project
mcpknife deploy ./dict-server --name dict-api --region lax

# Set extra environment variables
mcpknife deploy ./dict-server --name dict-api --env DB_URL=postgres://...

# Tear down
mcpknife deploy --destroy --name dict-api

Deploy takes an exported project directory (positional arg or piped from mcpknife export) and deploys it. The server is protected with bearer auth — an API key is auto-generated (or pass --api-key). The /health endpoint remains open for monitoring.

Currently supports Fly.io (--target fly, the default). Requires the fly CLI installed and authenticated (fly auth login).

Compose with pipes

The real power is composition. Each stage is a separate process connected by pipes:

# Generate → Transform → UI in one pipeline
mcpknife boot --prompt "GitHub API https://docs.github.com/en/rest" \
  | mcpknife mod --prompt "combine issues + pulls into get_activity" \
  | mcpknife ui

# Stack multiple transforms
mcpknife mod --upstream "npx some-server" --prompt "rename tools to snake_case" \
  | mcpknife mod --prompt "add synthetic summary tool" \
  | mcpknife ui --standard openai

# Full pipeline: generate → transform → export → deploy
mcpknife boot --prompt "Yahoo Finance" \
  | mcpknife mod --prompt "combine tools" \
  | mcpknife export \
  | mcpknife deploy --name yahoo-finance

Configuration

Create ~/.mcpkniferc to set defaults across all pipe stages:

{
  "provider": "anthropic",
  "model": "claude-sonnet-4-20250514",
  "apiKey": "sk-ant-..."
}

No more repeating --provider and --api-key on every command. CLI flags still override config values when you need them to.

A project-level ./.mcpkniferc overrides user-level settings, so teams can share provider/model defaults without sharing keys.

| Field | CLI flag | Description | |-------|----------|-------------| | provider | --provider | anthropic or openai | | model | --model | LLM model ID | | apiKey | --api-key | API key (or use env vars) | | verbose | --verbose | Verbose logging |

Precedence: CLI flags > env vars > project .mcpkniferc > user ~/.mcpkniferc

How it works

mcpknife is a thin multiplexer. It doesn't implement MCP logic — it delegates entirely to the three underlying tools by spawning them as child processes with inherited stdio.

cli.ts → config.ts → resolve.ts → args.ts → spawn.ts → child process

This means:

• Zero overhead. stdio: 'inherit' — the OS kernel handles the plumbing, not mcpknife.
• Zero coupling. mcpknife only knows the CLI contract. Internal changes to the underlying tools don't affect it.
• Pipe protocol works natively. Each tool detects whether stdout is a pipe and adjusts its behavior (port selection, URL output) automatically.
• Signals propagate cleanly. Ctrl+C forwards to the child. Exit codes pass through.

Subcommand reference

Run mcpknife <command> --help for full options:

mcpknife boot --help
mcpknife mod --help
mcpknife ui --help
mcpknife export --help
mcpknife deploy --help

Development

npm install
npm run build          # esbuild → dist/cli.js
npm test               # vitest
npm run dev            # tsx (no build needed)

License

Apache-2.0