Groundcrew

A local stdio MCP server that gives coding agents safe, deterministic access to common local developer operations.

For coding agents: See AGENT_QUICKSTART.md for the fastest path to using Groundcrew effectively.

What problem it solves

Coding agents often do expensive, fragile things for simple local operations:

• Reading a file, writing the same contents elsewhere, then deleting the original instead of moving it
• Reading whole files to find one line instead of searching
• Rewriting whole files for tiny text changes
• Editing without checking git state
• Dumping huge outputs into context

Groundcrew gives them safe local primitives for those actions, saving time and tokens.

With vs. without Groundcrew

Example 1: Moving a file

Without Groundcrew:

1. Read src/utils/old.ts (500 lines, ~8KB context)
2. Write contents to src/utils/new.ts (another ~8KB output)
3. Delete src/utils/old.ts
4. Verify the move worked (read both paths)

With Groundcrew:

{ "tool": "move_path", "arguments": { "from": "src/utils/old.ts", "to": "src/utils/new.ts" } }

Result: ~50 bytes of JSON in, ~100 bytes of JSON out. No file contents touch the context.

Example 2: Finding a function definition

Without Groundcrew:

1. Read src/server.ts (800 lines, ~12KB context) — "Hmm, not here"
2. Read src/routes.ts (600 lines, ~9KB context) — "Not here either"
3. Read src/handlers.ts (400 lines, ~6KB context) — "Found it!"

With Groundcrew:

{ "tool": "search_context", "arguments": { "pattern": "function handleRequest", "path": "src" } }

Result: Returns the exact file, line number, and surrounding context (~200 bytes per match). If results are capped, a suggestion field tells you how to narrow the search.

Verifiable claims

These examples describe real patterns we see in agent traces. You can verify the difference yourself:

1. Check your agent's context window usage during file operations
2. Count how many round-trips it takes to accomplish simple tasks
3. Compare token estimates: reading a 500-line file (~4KB-8KB) vs. a 50-byte tool call

What it is

Groundcrew is a local stdio MCP server that exposes workspace-scoped, output-capped, dry-run-friendly wrappers around common developer operations.

• TypeScript / Node.js LTS
• Distributed through npm
• stdio MCP transport only
• No HTTP server, no daemon, no background work
• Safe by default

Install and run instantly:

npm install @b7r_dev/groundcrew
npx groundcrew

Or run directly without installing:

npx @b7r_dev/groundcrew

Or run from source:

npm run dev

What it intentionally does not do

• No arbitrary shell execution tool
• No network operations
• No deploy / migration / database tools
• No hidden background work
• No remote services, auth, telemetry, or web UI
• No HTTP transport or daemon mode

Safety model

• All paths resolve against a workspace root (default: current working directory)
• Resolved paths must stay inside the workspace
• Path traversal outside the workspace is rejected
• Absolute paths outside the workspace are rejected
• Home paths (~) are rejected
• Binary files are refused for preview and edit tools
• Mutating tools default to dry-run
• Overwrites are refused by default
• Output is capped for every tool that can produce substantial output
• Search results are bounded
• All spawned CLI operations have timeouts

MCP configuration

Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "groundcrew": {
      "command": "npx",
      "args": ["-y", "@b7r_dev/groundcrew"],
      "cwd": "/path/to/your/project"
    }
  }
}

Generic MCP stdio config

{
  "name": "groundcrew",
  "transport": "stdio",
  "command": "npx",
  "args": ["-y", "@b7r_dev/groundcrew"],
  "cwd": "/path/to/your/project"
}

Tools

Workspace

• workspace_info — Return basic workspace facts: cwd, workspace root, platform, available tools, limits, and safety settings. Use this before touching files to know your environment.
• detect_project — Identify project type and important files. Detects Node, Go, Python, Rust, Make, and Git.
• list_project_commands — List known project-native commands without running them. Parses package.json scripts and Makefile targets. Labels each with risk level (safe, medium, risky).

Filesystem

• path_info — Return metadata for a path without reading contents: exists, type, size, modified time, binary guess. Use this to check a path before operating on it.
• ensure_dir — Create a directory (mkdir -p) safely scoped to workspace.
• move_path — Move or rename a file/directory using filesystem primitives. Use this instead of reading a file, writing it elsewhere, and deleting the original.
• copy_path — Copy a file or directory safely. Refuses directory copy unless recursive: true. Refuses overwrite by default.
• safe_move_path — Copy, verify byte equality, then delete original. Cautious move for files only.

Search & Preview

• find_files — Find files by substring query or simple glob. Respects ignored directories (.git, node_modules, dist, etc.). Use this instead of dumping an entire tree.
• search_context — Search text with line context using rg (ripgrep). Use this to find content inside files instead of reading whole files to look for one line.
• preview_file — Read a bounded slice of a text file. Refuses binary files. Use this instead of reading entire large files into context.

Git

• git_status — Cheap git preflight: branch, dirty state, short status entries. Use this before making edits to know the repo state.
• git_diff_stat — Compact summary of git changes.
• git_diff — Bounded git diff for review.

Edits

• replace_text — Exact deterministic text replacement with dry-run diff preview. Defaults to dryRun: true. Use this instead of rewriting an entire file for tiny changes.
• replace_regex — Controlled regex replacement with dry-run diff preview. Defaults to dryRun: true and maxReplacements: 1. Use this for patterned edits.

Default limits

| Limit | Value | |-------|-------| | maxOutputBytes | 20,000 | | maxResults | 100 | | maxPreviewBytes | 20,000 | | maxPreviewLines | 400 | | Spawned command timeout | 10,000 ms |

v0.1 limitations

• Only stdio transport (no HTTP)
• No arbitrary shell execution
• safe_move_path supports files only, not directories
• search_context requires rg (ripgrep) to be installed
• Glob matching in find_files is basic (* and ** only)
• No configuration file support yet
• Single workspace root per process

Development

# Install dependencies
npm install

# Run in development mode
npm run dev

# Run tests
npm test

# Run linter
npm run lint

# Run type check
npm run typecheck

# Build for distribution
npm run build

Project structure

src/
├── index.ts          # Entry point, stdio transport
├── server.ts         # MCP server registration, tool routing
├── config.ts         # Limits, ignored dirs, workspace root
├── errors.ts         # GroundcrewError class
├── tools/
│   ├── workspace.ts  # workspace_info, detect_project, list_project_commands
│   ├── filesystem.ts # path_info, ensure_dir, move/copy/safe_move
│   ├── search.ts     # find_files, search_context
│   ├── preview.ts    # preview_file
│   ├── git.ts        # git_status, git_diff_stat, git_diff
│   ├── edits.ts      # replace_text, replace_regex
│   └── project.ts    # detect_project, list_project_commands
└── utils/
    ├── path.ts       # resolveWorkspacePath, isInsideWorkspace
    ├── limits.ts     # capOutput, capResults
    ├── spawn.ts      # runCommand with timeout/byte cap
    ├── binary.ts     # isBinaryFile, refuseBinary
    └── diff.ts       # unifiedDiff generation
test/
├── safety.test.ts
├── filesystem.test.ts
├── preview.test.ts
├── edits.test.ts
└── project.test.ts

License

MIT