Minimal MCP server exposing Auggie CLI as tools for Q&A and code implementation.
- ask_question: Repository Q&A via Auggie’s context engine.
- implement: Implement a change in the repo; dry-run by default.
- Node.js 18+
- Python 3.10+ available on the system (used internally; no manual setup needed)
- Auggie CLI installed (check by running
auggie --version
) - see installation guide
Retrieve your token via the Auggie CLI:
# Ensure Auggie CLI is installed and on PATH
auggie --version
# Sign in (opens browser flow)
auggie login
# Print your token
auggie --print-augment-token
Provide the token in either of these ways:
- Cursor/Claude config (recommended): set it under
env
for the server
{
"mcpServers": {
"auggie-mcp": {
"command": "npx",
"args": ["-y", "auggie-mcp@latest"],
"env": { "AUGMENT_API_TOKEN": "YOUR_TOKEN" }
}
}
}
- Shell environment (macOS/Linux)
One-off for a single command:
AUGMENT_API_TOKEN=YOUR_TOKEN npx -y auggie-mcp --setup-only
Persist for future shells (zsh):
echo 'export AUGMENT_API_TOKEN=YOUR_TOKEN' >> ~/.zshrc
source ~/.zshrc
Security tip: never commit tokens to source control. Prefer per-machine environment variables or your client's secure config store.
Use this MCP config in Cursor (global or per-project):
{
"mcpServers": {
"auggie-mcp": {
"command": "npx",
"args": ["-y", "auggie-mcp@latest"],
"env": { "AUGMENT_API_TOKEN": "YOUR_TOKEN" }
}
}
}
This will:
- download the wrapper package,
- create a local Python venv inside the package,
- install
requirements.txt
, and - launch the Python server in
stdio
mode.
# Install deps into the package's local venv (no global installs)
npx -y auggie-mcp --setup-only
# Run the server (stdio). Useful for quick smoke-tests.
npx -y auggie-mcp
# Optional: start HTTP mode for manual debugging
npx -y auggie-mcp -- --http
Edit ~/Library/Application Support/Claude/claude_desktop_config.json
and add:
{
"mcpServers": {
"auggie-mcp": {
"command": "npx",
"args": ["-y", "auggie-mcp@latest"],
"env": { "AUGMENT_API_TOKEN": "YOUR_TOKEN" }
}
}
}
- Default:
implement
runs in dry‑run mode. No files are written, no shell runs; you get a proposed diff. - Enable writes: set
dry_run: false
. - Recommendation: use a feature branch and review the diff before merging.