Coding Guides MCP Server

A Model Context Protocol (MCP) server that provides access to coding guides and best practices for AI assistants like Claude and GitHub Copilot.

What is this?

This MCP server exposes coding guidelines and style guides as resources that can be accessed by MCP clients. It's designed to extend or replace AGENTS.md files by providing a structured way to serve coding practices and guidelines to AI assistants during development.

Features

• Resource-based API: Exposes coding guides through MCP resources
• GitHub integration: Loads guides from GitHub repositories over the web
• Automatic caching: Caches downloaded guides locally for offline access
• Fallback support: Uses local cache or directory when network is unavailable
• Simple file-based storage: Guides can be stored as Markdown files locally
• Official MCP SDK: Built on the Python mcp SDK (MCPServer, formerly FastMCP)
• Easy integration: Works with any MCP-compatible client (Claude Desktop, Cline, etc.)

Available Resources

• guides://list - Lists all available coding guides
• guides://{guide_name} - Retrieves the content of a specific guide (e.g., guides://python.md)

Installation

From Source

# Clone the repository
git clone https://github.com/delian/codeguide-mcp.git
cd codeguide-mcp

# Install with uv (recommended)
uv sync

# Or with pip
pip install -e .

With Docker

docker build -t codeguide-mcp .
docker run -i codeguide-mcp

In VS Code

Or search for codeguide-mcp in the Extensions view MCP servers list (type @mcp in the Extensions search bar), or add it manually to .vscode/mcp.json:

{
  "servers": {
    "codeguide-mcp": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "delian/codeguide-mcp"]
    }
  }
}

Configuration

Configure the server by creating a config.toml file or setting environment variables:

GitHub Configuration (Recommended)

To load guides from a GitHub repository:

github_repo = "owner/repository"  # e.g., "delian/codeguide-mcp"
github_path = "guides"            # Path to guides directory in repo
github_branch = "main"            # Branch to fetch from
  cache_dir = ".guides-cache"       # Local cache directory
log_level = "INFO"

Local Directory Configuration

To use local guides only:

guides_dir = "guides"
log_level = "INFO"

Environment Variables

• GUIDES_GITHUB_REPO - GitHub repository (format: owner/repo)
• GUIDES_GITHUB_PATH - Path to guides directory in repository (default: guides)
• GUIDES_GITHUB_BRANCH - Branch to fetch from (default: main)
• GUIDES_CACHE_DIR - Local cache directory (default: .guides-cache)
• GUIDES_DIR - Local directory containing guide files (default: guides)
• GUIDES_LOG_LEVEL - Logging level (default: INFO)

Transport (see Remote deployment):

• GUIDES_TRANSPORT - stdio, streamable-http, or auto (default: auto — HTTP when a PORT env var is present, stdio otherwise)
• PORT - Port to listen on in HTTP mode; takes precedence over GUIDES_PORT (Cloud Run injects this)
• GUIDES_HOST - Bind address in HTTP mode (default: 0.0.0.0)
• GUIDES_HTTP_PATH - MCP endpoint path (default: /mcp)
• GUIDES_STATELESS_HTTP - Handle each request independently (default: true; required when replicas autoscale)
• GUIDES_ALLOWED_HOSTS - Host header allowlist enabling DNS-rebinding protection (default: empty = no Host validation)

Behavior

1. Network available + GitHub configured: Fetches guides from GitHub and caches them locally
2. Network unavailable: Uses local cache if available
3. No cache available: Falls back to local guides_dir if configured

Remote deployment (Google Cloud Run)

The same image serves both transports: it speaks stdio over a pipe by default, and switches to Streamable HTTP when a PORT env var is present — which Cloud Run always injects. No separate image or entrypoint is needed.

1. Publish the image

docker build -t delian/codeguide-mcp:0.1.0 -t delian/codeguide-mcp:latest .
docker push delian/codeguide-mcp:0.1.0
docker push delian/codeguide-mcp:latest

2. Deploy

gcloud run deploy codeguide-mcp \
  --image=docker.io/delian/codeguide-mcp:0.1.0 \
  --region=europe-west1 \
  --allow-unauthenticated \
  --port=8080 \
  --set-env-vars=GUIDES_TRANSPORT=streamable-http,GUIDES_GITHUB_REPO= \
  --memory=512Mi --cpu=1 \
  --min-instances=0 --max-instances=4 --concurrency=40

GUIDES_GITHUB_REPO= (empty) makes the service serve the guides baked into the image. Leaving GitHub enabled adds a network round-trip per guide and runs into the unauthenticated GitHub API limit of 60 requests/hour per egress IP, after which the server silently falls back to those same baked-in files anyway.

The MCP endpoint is then https://<service-url>/mcp:

gcloud run services describe codeguide-mcp --region=europe-west1 \
  --format='value(status.url)'

Cloud Run answers on two hostnames for the same service — the SERVICE-PROJECTNUMBER.REGION.run.app form printed by gcloud run deploy, and the older SERVICE-HASH-REGIONCODE.a.run.app form that status.url reports. Both are equivalent; either works in a client config.

3. Point clients at it

See Connecting to a remote server below for the per-client configuration.

Pulling from Docker Hub

Cloud Run deploys public Docker Hub images directly, but caches them for only an hour and re-pulls anonymously afterwards, so a scale-up can hit Docker Hub's anonymous pull limits and fail to start instances. For anything beyond casual use, mirror through an Artifact Registry remote repository:

gcloud artifacts repositories create dockerhub \
  --repository-format=docker --location=europe-west1 \
  --mode=remote-repository --remote-docker-repo=DOCKER-HUB

gcloud run deploy codeguide-mcp \
  --image=europe-west1-docker.pkg.dev/PROJECT_ID/dockerhub/delian/codeguide-mcp:0.1.0 \
  ...

Notes on running it publicly

• --allow-unauthenticated makes the endpoint world-callable. The server is read-only, but the clear_cache prompt is reachable by any caller and drops the in-memory caches, and traffic drives autoscaling cost — keep --max-instances capped. To restrict access, omit the flag and have clients send an identity token, or front the service with Cloud Armor / API Gateway.
• GUIDES_STATELESS_HTTP must stay true unless you also enable session affinity, since Cloud Run may route a session's requests to different instances.
• GET / returns 404 by design; only /mcp is served. Cloud Run's default startup probe is a TCP check on $PORT, so this is fine — don't configure an HTTP health check on /.
• Set GUIDES_ALLOWED_HOSTS to your service hostname to enable Host-header validation if you expose the service under a custom domain.

Publishing to the MCP Registry

The MCP servers list in the VS Code Extensions view (type @mcp in the search bar) is fed by the GitHub MCP Registry, which ingests from the official MCP Registry. Publishing there is therefore how this server becomes discoverable in VS Code — no VS Code extension of its own is required.

server.json holds the registry metadata: the Docker image for clients that want to run it locally, and the hosted URL for clients that would rather not. Ownership of the image is proven by the io.modelcontextprotocol.server.name label in the Dockerfile, whose value must equal .name in server.json.

Authenticate once (an interactive device-code flow), then run the publish script:

mcp-publisher login github     # namespace io.github.<your-username>/*
tools/publish.sh

tools/publish.sh does the whole release: it checks the required tooling and Docker login, verifies that server.json and pyproject.toml agree on the version and that the Dockerfile label matches the server name, builds and pushes :VERSION and :latest, validates server.json against the live registry, publishes, then reads the entry back to confirm.

tools/publish.sh --dry-run          # everything except push and publish
tools/publish.sh --version 0.2.0    # bump server.json + pyproject + image tag, then release
tools/publish.sh --skip-build       # reuse images already on Docker Hub

Install mcp-publisher from the registry quickstart if you do not have it. After publishing, inclusion in GitHub's curated list may need a request to partnerships@github.com.

Adding Guides

Using GitHub (Recommended)

If you've configured github_repo, simply add Markdown files to the specified directory in your GitHub repository. The server will automatically fetch and cache them.

Using Local Directory

Add Markdown files to the guides/ directory. Each file will be automatically available as a resource.

Example:

echo "# Python Style Guide\n\nUse PEP 8..." > guides/python.md

Usage with MCP Clients

The server can be consumed two ways:

Local mode needs no network and no hosting; remote mode lets a team share one deployment and keeps the guides identical for everyone.