dev-tools-mcp

An MCP (Model Context Protocol) stdio server that gives Claude Code - or any MCP client - five structured development tools:

Prerequisites

• Node.js ≥ 18

• Claude Code installed

• For the RLS tool: psql on your PATH and a running Supabase/Postgres instance

Related MCP server: saas-audit-mcp

Quick start

# 1. Clone the repo
git clone https://github.com/briancox730/dev-tools-mcp.git
cd dev-tools-mcp

# 2. Install dependencies
npm install

# 3. Build
npm run build

# 4. Register with Claude Code (local scope - this project only)
claude mcp add --transport stdio dev-tools -- node /absolute/path/to/dev-tools-mcp/build/index.js

# OR register globally (available in all projects)
claude mcp add --transport stdio --scope user dev-tools -- node /absolute/path/to/dev-tools-mcp/build/index.js

Alternative: edit .claude.json directly

Add this to the mcpServers key in ~/.claude.json (global) or .claude.json in your project root:

{
  "mcpServers": {
    "dev-tools": {
      "type": "stdio",
      "command": "node",
      "args": ["/absolute/path/to/dev-tools-mcp/build/index.js"]
    }
  }
}

Then restart Claude Code.

Verify it's working

Inside Claude Code, run /mcp - you should see dev-tools: connected with 5 tools listed.

Tool details

1. run_e2e_tests - Playwright E2E runner

Runs Playwright with the JSON reporter and parses results into structured output.

Inputs:

• project_dir (required) - absolute path to your project

• test_pattern - glob or file path to filter, e.g. "tests/auth.spec.ts"

• headed - run in headed mode (default: false)

• start_server - use Playwright's webServer config (default: true)

• timeout_seconds - kill after this many seconds (default: 120)

What you get back:

{
  "success": true,
  "summary": { "passed": 8, "failed": 1, "skipped": 0, "total": 9 },
  "tests": [
    { "name": "Auth > should redirect unauthenticated users", "status": "passed", "duration_ms": 1200 },
    { "name": "Auth > should show dashboard after login", "status": "failed", "duration_ms": 3400, "error": "Expected element to be visible..." }
  ]
}

Tip: Make sure your playwright.config.ts has a webServer section so Playwright auto-starts your dev server:

export default defineConfig({
  webServer: {
    command: 'npm run dev',
    port: 3000,
    reuseExistingServer: !process.env.CI,
  },
});

2. execute_sql_as_role - Supabase RLS tester

Executes SQL as a specific Postgres role with JWT claims, then rolls back. Never mutates data.

Inputs:

• connection_string (required) - e.g. "postgresql://postgres:postgres@localhost:54322/postgres"

• sql (required) - the query to run

• role - Postgres role (default: "authenticated")

• user_id - UUID set as auth.uid()

• claims - additional JWT claims, e.g. { "app_role": "parent" }

Example prompt for Claude Code:

"Run SELECT * FROM children as an authenticated parent user with id abc-123 and confirm they can only see their own children."

What you get back:

{
  "success": true,
  "role": "authenticated",
  "user_id": "abc-123",
  "query": "SELECT * FROM children",
  "rows": [["id-1", "abc-123", "Alice"], ["id-2", "abc-123", "Bob"]],
  "row_count": 2
}

3. typecheck - TypeScript checker

Runs tsc --noEmit and parses the output into structured errors.

Inputs:

• project_dir (required)

• tsconfig - relative path to tsconfig (default: "tsconfig.json")

• files - check only specific files

What you get back:

{
  "success": false,
  "error_count": 2,
  "errors": [
    { "file": "src/utils.ts", "line": 42, "column": 5, "code": "TS2345", "message": "Argument of type 'string' is not assignable..." },
    { "file": "src/api.ts", "line": 18, "column": 12, "code": "TS2339", "message": "Property 'foo' does not exist on type..." }
  ]
}

4. npm_run - Structured npm script runner

Runs any npm script with CI=true and FORCE_COLOR=0, then parses Vitest/Jest output.

Inputs:

• project_dir (required)

• script (required) - e.g. "test", "test:unit", "lint"

• args - extra args passed after --

• env - extra environment variables

What you get back:

{
  "success": false,
  "exit_code": 1,
  "script": "test",
  "summary": { "total_tests": 14, "passed_tests": 12, "failed_tests": 2 },
  "file_results": [
    { "file": "src/auth.test.ts", "status": "failed", "tests_failed": 2 },
    { "file": "src/utils.test.ts", "status": "passed", "tests_passed": 5 }
  ],
  "raw_stdout": "..."
}

5. nextjs_build - Next.js build validator

Runs next build in production mode and parses errors and page output.

Inputs:

• project_dir (required)

• env - extra env vars for the build

• timeout_seconds - (default: 180)

What you get back:

{
  "success": true,
  "error_count": 0,
  "errors": [],
  "pages": [
    { "path": "/", "size_kb": 5.42, "type": "static" },
    { "path": "/dashboard", "size_kb": 12.1, "type": "dynamic" },
    { "path": "/api/auth", "size_kb": 0, "type": "ssr" }
  ],
  "build_duration_ms": 24500
}

Development

# Watch mode
npm run watch

# Run the unit tests (Vitest)
npm test

# Test interactively with the MCP Inspector
npm run inspect

Tests

The output parsers are the trickiest, most regression-prone part of the codebase. They turn free-form Vitest/Jest/tsc/next build console output into structured JSON. Those pure functions are unit-tested with Vitest under test/; run them with npm test. CI (see .github/workflows/ci.yml) runs npm ci, the TypeScript build, and the test suite on Node 20 and 22 for every push and pull request.

Customization ideas

• Add a lint tool that parses ESLint JSON output

• Add a prisma_migrate tool for structured migration status

• Add a docker_compose tool to manage test containers

• Wire the RLS tool to Supabase CLI (supabase db test) instead of raw psql

License

MIT © 2026 Brian Cox