@oscarmarin/mcp-devtools

mcp-devtools

AI-native developer tools via Model Context Protocol. A production-grade MCP server that gives AI agents (Claude, Cursor, Copilot, Continue, ...) safe, scoped access to your local development environment.

Why

The MCP ecosystem is full of single-purpose tutorials and vendor-locked adapters. There is no well-maintained, multi-tool, framework-agnostic, production-quality MCP package for everyday developer tooling.

mcp-devtools fills that gap with 13 tools across 4 categories (filesystem, database, process, OpenAPI) and two transport modes (stdio + HTTP), built on patterns refined in production at DailyBot: retry with jitter, structured logging, typed error taxonomy, and security-first design.

Quick start

stdio (default)

npx @oscarmarin/mcp-devtools

Add it to Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "devtools": {
      "command": "npx",
      "args": ["-y", "@oscarmarin/mcp-devtools"]
    }
  }
}

Or Cursor (~/.cursor/mcp.json): same block.

HTTP transport

Create a mcp-devtools.json in your project root:

{
  "transport": "http",
  "port": 3333
}

Then start the server:

npx @oscarmarin/mcp-devtools

The MCP endpoint will be available at http://localhost:3333.

Tools

Group	Tools
Filesystem	read_file, write_file, list_directory, search_files, get_file_info
Database	query_db, list_tables, describe_table
Process	run_command, read_logs, get_env
OpenAPI	parse_openapi, call_api

Per-tool reference: docs/tools/.

Configuration

Configuration is loaded by cosmiconfig from mcp-devtools.json, .mcp-devtoolsrc, or the mcpDevtools key in package.json. See mcp-devtools.example.json and docs/configuration.md for the full schema.

Zero-config is supported: running npx mcp-devtools with no config uses schema defaults (RNF-05).

Security

Three non-bypassable controls:

• Filesystem scope boundary. Every path is resolved to an absolute and compared against config.scope. Symlinks that escape scope throw SCOPE_VIOLATION.
• Command allowlist. run_command only executes binaries whose basename is in allowedCommands. Invocation uses spawn(file, args) (no shell), so shell-injection via the command argument is structurally impossible.
• Database read-only mode. When readOnly: true, all SQL is parsed and INSERT/UPDATE/DELETE/DROP/CREATE/GRANT are rejected. Result sets are capped (default 200 rows). Queries run in BEGIN READ ONLY ... ROLLBACK on PostgreSQL.

Additional safety measures:

• Secret masking. get_env automatically masks values matching common secret patterns (SECRET, TOKEN, PASSWORD, KEY, etc.).
• OpenAPI host restriction. call_api only sends requests to hosts listed in the spec's servers array. Requests to unlisted hosts are rejected.
• Output capping. All tools cap their output to prevent context flooding (100KB for commands, 1MB for files, 200 rows for queries).

Contributing

git clone https://github.com/marin1321/mcp-devtools.git
cd mcp-devtools
npm install
npm run dev       # tsup --watch
npm run test      # vitest
npm run typecheck # tsc --noEmit
npm run lint      # eslint .

See CONTRIBUTING.md for the full workflow and CODE_OF_CONDUCT.md for community guidelines.

License

MIT © Oscar Humberto Marin Molina — oscarmarindev.com