mcp-tooling
DarojaAI/mcp-tooling
MCP tool servers for OpenClaw agents: capability layer wrapping external APIs and host operations
Instalação
npx convez-ai DarojaAI/mcp-tooling1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
mcp-tooling
MCP tool servers for OpenClaw agents: capability layer wrapping external APIs and host operations
What is this?
This repo provides MCP (Model Context Protocol) servers that expose external capabilities to OpenClaw agents:
- Travel booking (Duffel API) — search flights, book, manage bookings
- VM operations (planned) — host diagnostics, log tailing, process control
- Calendar/scheduling (planned) — Cal.com integration
- Payments (planned) — Stripe integration
Each server is a capability adapter: it wraps a third-party API with tool-calling interfaces that agents can use.
Architecture
See dev-nexus architectural boundaries for the separation of concerns:
- mcp-tooling (this repo): Capability Layer — adapts external APIs
- dev-nexus: Analysis Layer — repository intelligence, code/data analysis
- openclaw-gateway: Orchestration Layer — agent runtime, multi-step workflows
Directory structure
mcp-tooling/
├── runtime/ # Shared framework (BaseTool, registry, stdio/HTTP servers)
├── servers/ # Per-capability servers (duffel/, vm-ops/, cal/, etc.)
├── terraform/ # Hetzner VM provisioning
├── scripts/ci/ # CI validation scripts
├── scripts/deploy/ # In-VM install script (called by deploy workflow)
├── config/ # Data contract (GitHub Actions vars/secrets)
├── docs/ # Architecture decisions, authoring guides
└── tests/ # Framework + server tests
Quick start
Running a server locally (stdio)
# Install dependencies
pip install -e .
# Set up secrets
cp servers/duffel/config.example.env /etc/mcp-tooling/secrets.env
# Edit secrets.env with your Duffel API key
# Run the Duffel server
python -m servers.duffel
The server speaks MCP over stdio. Connect with any MCP client (e.g., openclaw-gateway).
Running a server locally (HTTP)
python -m servers.duffel --http --port 8765
curl http://localhost:8765/healthz
Deploying to Hetzner
The deploy workflow is in .github/workflows/deploy-duffel-hetzner.yml and provisions a Hetzner VM via terraform/ (using the shared terraform-hcloud-linux-vm module), then runs scripts/deploy/install-vm.sh on the VM to install the Duffel MCP server as a systemd service.
Prerequisites (one-time, per environment):
- Create a GitHub environment at
Settings → Environments(e.g.dev,prod). - Configure the required secrets and environment variables on that environment.
- Create the HCX S3 bucket (e.g.
terraform-state-mcp-tooling). - Register an SSH key in your Hetzner project.
To deploy:
- GitHub Actions → "Deploy Duffel to Hetzner" → Run workflow
- Choose environment (dropdown is auto-populated from your GitHub envs via
sync-environment-options.yml) - Choose action:
plan(preview) orapply(execute) - Review plan output, then re-run with
apply
To tear down: Run the workflow with action destroy. The next apply will re-create the VM.
Adding a new server
See docs/authoring-server.md for the step-by-step guide.
TL;DR:
- Create
servers/<name>/ - Write tools that inherit from
runtime/base.py:BaseTool - Register them in
servers/<name>/server.py - Add tests in
servers/<name>/tests/
The runtime handles stdio/HTTP, health checks, allowlists, and secrets loading — you just write the tool logic.
Data contract
All GitHub Actions environment variables and secrets are documented in config/dat-contract.yaml.
See docs/github-actions-secrets.md for the rendered reference (auto-generated from the contract).
Contributing
- Boundary check: If adding code to
runtime/orservers/<name>/tools/, confirm it doesn't belong in dev-nexus (analysis) or openclaw-gateway (orchestration) per architectural boundaries. - Contract changes: If editing
config/dat-contract.yaml, update.env.exampleanddocs/github-actions-secrets.mdto match. - Tests: CI runs
pytest tests/— keep it green.
License
MIT