Run your agents, not your infrastructure.

mcpd is a tool to declaratively manage Model Context Protocol (MCP) servers, providing a consistent interface to define and run tools across environments, from local development to containerized cloud deployments.

Built by Mozilla AI

📚 mcpd official docs

---

Today, mcpd launches MCP servers as subprocesses using STDIO (Standard Input/Output) and acts as an HTTP proxy between agents and the tools they expose. This enables agent-compatible workflows with support for secrets, runtime arguments, and reproducible configurations, no matter where mcpd is running.

We're developing a Kubernetes operator, guided by our internal roadmap, to extend mcpd for deploying and managing MCP servers as long-lived services in production. It will use the same .mcpd.toml configuration and proxy model, making it easier to scale and manage lifecycles without changing the developer experience.

The Problem

ML teams build agents that work perfectly locally. Operations teams often get handed Python scripts with little context and told "make this production-ready across dev/UAT/prod." The gap between local development and enterprise deployment kills AI initiatives.

mcpd solves this with declarative configuration, secure secrets management, and seamless environment promotion - all while keeping the developer experience simple.

Why mcpd?

Zero-Config Tool Setup
No need to clone multiple repos or install language-specific dependencies. mcpd add and mcpd daemon handle everything.

Language-Agnostic Tooling
Use MCP servers in Docker containers or written in Python, JavaScript, or TypeScript via a unified HTTP API.

Declarative Configuration
Version-controlled .mcpd.toml files define your agent infrastructure. Reproducible, auditable, CI-friendly.

Enterprise-Ready Secrets
Separate project configuration from runtime variables, and export sanitized secrets templates. No need to commit secrets to Git ever again.

Seamless Local-to-Prod
Same configuration works in development, CI, and cloud environments without modification or risky hand-offs.

Built for Dev & Production

Development Workflow	Production Benefit
mcpd daemon runs everything locally	Same daemon runs in containers
.mcpd.toml version-controlled configs	Declarative infrastructure as code
Local secrets in ~/.config/mcpd/	Secure secrets injection via control plane
mcpd config export exports version-control safe snapshot of local configuration	Sanitized secrets config and templates for CI/CD pipelines

Features

• Focus on Developer Experience via mcpd CLI
• Declarative configuration (.mcpd.toml) to define required servers/tools
• Run and manage language-agnostic MCP servers
• Secure execution context for secrets and runtime args
• Smooth dev-to-prod transition via the mcpd daemon
• Rich CLI and SDK tooling, see supported languages below:

🚀 Quick Start

Prerequisites

You need the following installed before running mcpd:

• Docker
• npx (Node.js package runner)
• uvx (Universal virtual environment manager)

If you are developing mcpd, you will also need:

• Go (v1.25.1+ recommended)

Installation

via Homebrew

Add the Mozilla.ai tap:

brew tap mozilla-ai/tap

Then install mcpd:

brew install mcpd

Or install directly from the cask:

brew install --cask mozilla-ai/tap/mcpd

via GitHub releases

Official releases can be found on mcpd's GitHub releases page.

The following is an example of manually downloading and installing mcpd using curl and jq by running install_mcpd:

function install_mcpd() {
    command -v curl >/dev/null || { echo "curl not found"; return 1; }
    command -v jq >/dev/null || { echo "jq not found"; return 1; }

    latest_version=$(curl -s https://api.github.com/repos/mozilla-ai/mcpd/releases/latest | jq -r .tag_name)
    os=$(uname)
    arch=$(uname -m)

    zip_name="mcpd_${os}_${arch}.tar.gz"
    url="https://github.com/mozilla-ai/mcpd/releases/download/${latest_version}/${zip_name}"

    echo "Downloading: $url"
    curl -sSL "$url" -o "$zip_name" || { echo "Download failed"; return 1; }

    echo "Extracting: $zip_name"
    tar -xzf "$zip_name" mcpd || { echo "Extraction failed"; return 1; }

    echo "Installing to /usr/local/bin"
    sudo mv mcpd /usr/local/bin/mcpd && sudo chmod +x /usr/local/bin/mcpd || { echo "Install failed"; return 1; }

    rm -f "$zip_name"
    echo "mcpd installed successfully"
}

via local Go binary build

# Clone and build
git clone [email protected]:mozilla-ai/mcpd.git
cd mcpd
make build
sudo make install # Install mcpd 'globally' to /usr/local/bin

Using mcpd

# Initialize a new project
mcpd init

# Add an MCP server
mcpd add time

# Set the local timezone for the MCP server
mcpd config args set time -- --local-timezone=Europe/London

# Start the daemon in dev mode with debug logging
mcpd daemon --dev --log-level=DEBUG --log-path=$(pwd)/mcpd.log

# You can tail the log file
tail -f mcpd.log

API docs will be available at http://localhost:8090/docs.

Deploy Anywhere

mcpd is runtime-flexible and infrastructure-agnostic:

• ⚙️ Works in any container or host with uv and npx
• ☁️ Multi-cloud ready (AWS, GCP, Azure, on-prem)
• ♻️ Low resource overhead via efficient server management

SDKs

mcpd SDKs

Language	Repository	Status
Python	mcpd-sdk-python	✅
JavaScript	mcpd-sdk-javascript	✅

mcpd plugin SDKs

Plugin SDKs are built using the mcpd plugin Protocol Buffers specification.

Language	Repository	Status
Go	mcpd-plugin-sdk-go	✅
.NET	mcpd-plugin-sdk-dotnet	✅

More on plugins soon!

💻 Development

Build local code:

make build

Run tests:

make test

Validate Mozilla AI registry (when modifying registry files):

make validate-registry

Run the local documentation site (requires uv), dynamically generates command line documentation:

make docs

---

🤝 Contributing

Please see our Contributing to mcpd guide for more information.

📄 License

Licensed under the Apache License 2.0.