DocsHub Server

An MCP server that provides AI clients with access to developer documentation via llms.txt files. Exposes tools, resources, and prompts.

Prerequisites

You must have uv installed on your machine. (macOS/Linux: curl -LsSf https://astral.sh/uv/install.sh | sh)

Configuration

DocsHub loads its project list from a docs_config.yaml file. Sources are tried in this order:

1. DOCSHUB_CONFIG environment variable — set to a local file path or a remote https:// URL pointing to a raw YAML file (e.g. a file hosted in a GitHub repository)
2. docs_config.yaml in the working directory
3. ~/.config/docshub/docs_config.yaml — user-level config
4. Bundled default — the docs_config.yaml in the DocsHub repo (used automatically as a fallback when no other config is found)

The YAML format:

projects:
  - name: "FastAPI"
    description: "FastAPI official documentation"
    llms_txt_url: "https://fastapi.tiangolo.com/llms.txt"

Using a remote config from a GitHub repository:

Set DOCSHUB_CONFIG to the raw file URL:

DOCSHUB_CONFIG=https://raw.githubusercontent.com/ORG/docshub/main/docs_config.yaml

You can pass this to any AI client as an environment variable in its MCP server configuration (see Client Setup below).

Environment variables

Variable	Default	Description
DOCSHUB_CONFIG	(none)	Local file path or https:// URL to the config file.
DOCSHUB_CONFIG_RELOAD_INTERVAL	900	Seconds between config reload checks. The server polls this source in the background and notifies clients when the project list changes. Set to 0 to disable polling entirely.
DOCSHUB_CACHE_TTL	1800	Seconds to cache fetched documentation in memory. Subsequent tool calls within the TTL window are served instantly without a network round-trip. Set to 0 to disable caching.

---

Client Setup

MCP server configuration

Claude Desktop, VS Code Cline, and GitHub Copilot CLI use the same JSON format. Add this block to the mcpServers object in each client's config file (see per-client instructions below):

{
  "mcpServers": {
    "docshub": {
      "command": "uvx",
      "args": [
        "--from", "git+https://github.com/rkratky/docshub",
        "docshub"
      ]
    }
  }
}

To load a remote docs_config.yaml, add an env key:

{
  "mcpServers": {
    "docshub": {
      "command": "uvx",
      "args": [
        "--from", "git+https://github.com/rkratky/docshub",
        "docshub"
      ],
      "env": {
        "DOCSHUB_CONFIG": "https://raw.githubusercontent.com/ORG/docshub/main/docs_config.yaml"
      }
    }
  }
}

---

Claude Desktop

1. Open your configuration file:
   • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
   • Windows: %APPDATA%\Claude\claude_desktop_config.json
2. Add the configuration block above to the mcpServers object.
3. Completely quit and restart Claude Desktop.

---

VS Code — Cline extension

1. Open VS Code and open the Cline extension sidebar.
2. Click the MCP Servers icon → Configure MCP Servers to open cline_mcp_settings.json.
3. Add the configuration block above to the mcpServers object.
4. Save. Cline automatically restarts the server.

---

GitHub Copilot (VS Code)

GitHub Copilot in VS Code uses VS Code's native MCP configuration, which has a slightly different format.

Option A — User settings (available across all workspaces):

1. Open VS Code user settings: Ctrl+, → Open Settings (JSON) (top-right icon).
2. Add the following:

{
  "mcp": {
    "servers": {
      "docshub": {
        "type": "stdio",
        "command": "uvx",
        "args": [
          "--from", "git+https://github.com/rkratky/docshub",
          "docshub"
        ]
      }
    }
  }
}

Option B — Workspace settings (scoped to a single project):

Create or edit .vscode/mcp.json in your project root:

{
  "servers": {
    "docshub": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "--from", "git+https://github.com/rkratky/docshub",
        "docshub"
      ]
    }
  }
}

After saving, open GitHub Copilot Chat — the docshub tools will be available automatically.

---

Claude Code

Run this command once to register the server at user scope:

claude mcp add --scope user docshub -- \
  uvx --from git+https://github.com/rkratky/docshub docshub

To use a remote docs_config.yaml, set the environment variable before running the command or add it to your shell profile:

export DOCSHUB_CONFIG=https://raw.githubusercontent.com/ORG/docshub/main/docs_config.yaml

---

GitHub Copilot CLI

Edit ~/.copilot/mcp-config.json (created automatically by the CLI the first time you run it; you can also create it manually) and add the configuration block above to the mcpServers object.

To use a remote docs_config.yaml, add an env key to the server entry:

{
  "mcpServers": {
    "docshub": {
      "command": "uvx",
      "args": [
        "--from", "git+https://github.com/rkratky/docshub",
        "docshub"
      ],
      "env": {
        "DOCSHUB_CONFIG": "https://raw.githubusercontent.com/ORG/docshub/main/docs_config.yaml"
      }
    }
  }
}

The config file location can be changed by setting the COPILOT_HOME environment variable.

---

Usage

DocsHub exposes three types of MCP primitives: tools (called automatically by the AI), resources (attached to context on demand), and prompts (predefined conversation starters).

Tools

The AI calls these automatically when you ask about documentation:

Tool	Description
list_available_docs	Lists configured projects. Always called first to discover what's available.
get_project_docs	Fetches documentation for a project. Tries llms-full.txt first (complete docs); falls back to llms.txt (an index of page URLs) if unavailable.
read_doc_page	Fetches a specific page by URL. Only needed when get_project_docs returned an llms.txt index.

Example: "Check the docs for FastAPI and show me how to write a route."

Performance note: fetched documentation is cached in memory for 30 minutes by default (DOCSHUB_CACHE_TTL); set to 0 to disable caching. The server also polls the config source in the background every 15 minutes (DOCSHUB_CONFIG_RELOAD_INTERVAL) and notifies connected clients when the project list changes, so doc sets can be added or removed without restarting the server; set to 0 to disable polling.

Resources

Resources provide structured access to the documentation registry and content. Attach them to your conversation context when you want to ground the AI in a specific project's docs.

URI	Description
docshub://projects	JSON list of all configured projects with names and descriptions.
docshub://project/{name}/docs	Raw documentation content for the named project.

Prompts

Prompts are predefined conversation starters for common documentation tasks. How to invoke them depends on your AI client (look for a prompt/slash-command picker or similar):

Prompt	Arguments	Description
query_docs	project_name, question	Answer a specific question using a project's documentation.
summarize_project	project_name	Produce a structured overview of a project's docs.