mc-mcp

A Model Context Protocol (MCP) server for the metacontract (mc) framework, providing smart contract development support via Foundry integration and semantic documentation search.

---

Overview

mc-mcp is an extensible MCP server designed for the metacontract (mc) framework. It enables AI-powered smart contract development workflows by exposing tools such as:

• mc_search_docs_semantic: Performs semantic search over mc documentation and user-configured sources, returning structured JSON results.
• mc_setup: Initializes a new Foundry project using the metacontract/template. (Set setup.force=true in the config if the directory is not empty.)
• mc_test: Runs Foundry tests (forge test) in your workspace.
• mc_deploy: Deploys contracts using the script specified in mcp_config.toml. Supports dry-run (default) and broadcast mode (broadcast: true argument).
• mc_upgrade: Upgrades contracts using the script specified in mcp_config.toml. Supports dry-run and broadcast mode.
• mc_lint: (Coming Soon) Lints project files for best practices and errors.
• mc_erc7201_slot: Calculates the namespaced storage slot using ERC-7201.

---

Configuration and Project Root

Project Root is always specified by the MC_PROJECT_ROOT environment variable.

• The MC_PROJECT_ROOT environment variable is required because path resolution is delegated to the MCP client implementation. This ensures consistent and reliable behavior across different environments.
• Set MC_PROJECT_ROOT to your project root directory before running any mc-mcp command or tool.
• All configuration, including mcp_config.toml, is expected to be in this directory.
• If MC_PROJECT_ROOT is not set or does not exist, mc-mcp will return an error and not start.
• If mcp_config.toml is missing, mc-mcp will use built-in defaults and log a warning.

Example: Project Structure

/path/to/your/project/
├── mcp_config.toml   # (optional) Project configuration
├── contracts/
├── scripts/
└── ...

Configuration File (mcp_config.toml)

• Place this file directly under your MC_PROJECT_ROOT directory.
• If omitted, mc-mcp will use default settings for all configuration values.
• See below for a sample config and field descriptions.

---

Getting Started

1. Install mc-mcp (as a CLI tool)

# recommended for most users
cargo install mc-mcp

# Or, to use the latest development version from GitHub:
cargo install --git https://github.com/metacontract/mc-mcp.git

# Alternatively, you can clone and build manually:
git clone https://github.com/metacontract/mc-mcp.git
cd mc-mcp
cargo build --release

2. Add to your MCP client

mc-mcp works with any MCP-compatible client, such as:

• Cursor
• Cline
• RooCode

Example: Cursor

Add your MCP server in the settings (./cursor/mcp.json or global settings):

{
  "mcpServers": {
    "mc-mcp": {
      "command": "mc-mcp", // or "/path/to/mc-mcp/target/release/mc-mcp",
      "env": {
        "MC_PROJECT_ROOT": "/absolute-path/to/your/project"
      }
    }
  }
}

Example: Cline

Cline auto-detects MCP servers in your workspace. For advanced configuration, see Cline's documentation.

Example: RooCode

RooCode supports MCP integration out of the box. See RooCode's documentation for details.

{
  "mcpServers": {
    "mc-mcp": {
      "disabled": false,
      "timeout": 60,
      "command": "mc-mcp",
      "transportType": "stdio",
      "alwaysAllow": ["mc_setup", "mc_search_docs_semantic", "mc_test", "mc_deploy", "mc_upgrade"], // if needed
      "env": {
        "MC_PROJECT_ROOT": "/absolute-path/to/your/project"
      }
    }
  }
}

Then, your MCP client starts the mc-mcp server automatically.

• On first startup, the prebuilt documentation index (prebuilt_index.jsonl.gz) will be downloaded from the latest GitHub Release if not present.

3. Initialize a new mc project

You can simply instruct your agent to "setup mc project".

or, manual setup

forge init <your-project-name> -t metacontract/template
cd <your-project-name>

4. Develop your smart contract

• Use your MCP-compatible Agent/IDE to interact with mc-mcp
• Design, search docs, and run TDD cycles (mc_test, mc_search_docs_semantic, etc.)

Configuration Example

Create a file named mcp_config.toml in your project root:

# Reference sources for semantic search
[reference]
# The prebuilt index is downloaded to the system cache directory by default.
# You can override the path here if needed:
# prebuilt_index_path = "/custom/path/to/prebuilt_index.jsonl.gz"

[[reference.sources]]
name = "mc-docs"
source_type = "local"
path = "metacontract/mc/site/docs"

[[reference.sources]]
name = "solidity-docs"
source_type = "local"
path = "docs/solidity"

[[reference.sources]]
name = "user-docs"
source_type = "local"
path = "docs/user"

# Default scripts used by tools
[scripts]
deploy = "scripts/Deploy.s.sol"  # Used by mc_deploy
upgrade = "scripts/Upgrade.s.sol" # Used by mc_upgrade

# Optional: Settings for broadcasting transactions (used when broadcast=true)
rpc_url = "http://localhost:8545" # RPC endpoint URL
private_key_env_var = "PRIVATE_KEY" # Name of the env var holding the deployer's private key

• prebuilt_index_path ... (optional) Path to a prebuilt index (jsonl or gzipped jsonl). If set, it overrides the default cache location. The index will be loaded and upserted into Qdrant on startup.
• Each [[reference.sources]] must have name, source_type (usually local), and path (relative to the execution directory).
• [scripts].deploy and [scripts].upgrade specify the default Foundry script paths used by the mc_deploy and mc_upgrade tools respectively.
• [scripts].rpc_url and [scripts].private_key_env_var are required when using mc_deploy or mc_upgrade with the broadcast: true argument. forge will use these to send the transaction.
• All paths must exist and be directories, or indexing will fail.

See also: config.rs for the full config structure.

---

Features

• Layered (Onion) Architecture for maintainability and testability
• Rust implementation for performance and safety
• MCP SDK (modelcontextprotocol-rust-sdk) with #[tool] annotation-based tool definition
• Qdrant (embedded vector DB) for fast semantic search
• Local embedding generation using fastembed-rs
• Markdown parsing with comrak
• Flexible configuration via figment
• Structured logging with tracing
• CI/CD: Prebuilt documentation index is generated, compressed, and published as a GitHub Release asset

---

Architecture

• Library + Binary crate structure
• Domain / Application / Infrastructure layering
• Manual Dependency Injection
• Strict inward dependency rule (Entry Point → Application → Domain ← Infrastructure)
• Prebuilt index: Downloaded automatically from GitHub Releases on first run (not included in the crate)

---

Project Structure

• src/domain/ — Core business logic, traits, entities
• src/application/ — Use cases, DTOs
• src/infrastructure/ — Integrations (Qdrant, embeddings, file system, etc.)
• src/main.rs — Entry point, MCP tool registration, DI

---

Development

• TDD: Test-Driven Development is enforced (unit/integration tests)
• Integration tests: Use testcontainers for Qdrant, ensuring isolated and stable test environments.
• CI/CD: GitHub Actions for build, test, and prebuilt index artifact management

Testing & Troubleshooting

Some tests (especially integration tests and embedding-related tests) may fail due to OS file descriptor limits or cache lock issues.

Common Commands (via Makefile)

• All tests (include ignored tests):

  make test-all

• Unit tests (lib only, with cache lock cleanup, single-threaded):

  make test-lib

• Integration tests (single-threaded):

  make test-integration

• Clean embedding model cache:

  make clean-cache

Troubleshooting

• If you see Too many open files errors:
  • Increase the file descriptor limit in your shell before running tests: ulimit -n 4096
  • Run tests sequentially: cargo test -- --test-threads=1 (or use make test-lib / make test-integration)
• If you see .lock errors (related to embedding model cache):
  • Clean the cache: make clean-cache
• If tests involving forge commands fail unexpectedly:
  • Ensure the mock script setup within the specific test file (src/main.rs tests) is correct.
• Note:
  • The Makefile provides handy shortcuts for common tasks, but some OS or CI environments may require manual adjustment of file descriptor limits (ulimit).
  • See each Makefile target for more details.

---

Task Management & Progress

For up-to-date milestones, progress, and detailed task tracking, see: ➡️ mc-mcp Task List

---

License

MIT