Use this CLI-first toolkit to produce LLM-friendly documentation bundles (llms.txt, llms-full.txt, optional llms-ctx.txt, and fallback JSON) for any GitHub repository. The generator wraps DSPy analyzers, manages LM Studio model lifecycle with the official Python SDK, and guarantees output even when the primary language model cannot respond.

• Python 3.10 or later
• LM Studio server available locally (Developer tab → Start Server) or the CLI (lms server start --port 1234)
• GitHub API token in GITHUB_ACCESS_TOKEN or GH_TOKEN
• Optional: llms_txt when you want to produce llms-ctx.txt

python3 -m venv .venv
source .venv/bin/activate

Using uv:

uv venv
source .venv/bin/activate

pip install -e '.[dev]'

Using uv:

uv pip install -e '.[dev]'

Installing the editable package exposes the lmstxt CLI and the lmstxt-mcp server.

npx lmstudio install-cli
lms server start --port 1234

The server must expose an OpenAI-compatible endpoint, commonly http://localhost:1234/v1.

Open LM Studio, download the model (for example qwen/qwen3-4b-2507), and confirm it appears in the Server tab.

Run the CLI against any GitHub repository:

lmstxt https://github.com/owner/repo \
  --model qwen/qwen3-4b-2507 \
  --api-base http://localhost:1234/v1 \
  --stamp

The command writes artifacts to artifacts/owner/repo/. Use --output-dir to override the destination.

To run against a private repository you own, the GitHub API calls must be authenticated. The CLI reads a token from GITHUB_ACCESS_TOKEN or GH_TOKEN and sends it as a Bearer token.

Token options:

• Classic PAT: grant at least repo scope for private repo read access.
• Fine-grained PAT: grant read access to Contents and Metadata for the target repo.
• Org SSO: if your org enforces SSO, you must authorize the token in the GitHub SSO UI.

Environment setup examples:

export GITHUB_ACCESS_TOKEN="ghp_..."
# or
export GH_TOKEN="ghp_..."

You can also use a .env file; the CLI calls load_dotenv() on startup and reads the token into AppConfig.github_token.

GITHUB_ACCESS_TOKEN=ghp_...

Run the CLI normally once the token is available:

lmstxt https://github.com/<owner>/<repo>

If you run the MCP server, the same env vars must be present in the process environment that launches lmstxt-mcp (for example, in your MCP client config).

This package includes a FastMCP server that exposes the generator as an MCP tool and provides access to generated artifacts as resources.

• Asynchronous Processing: Tool calls return a run_id immediately while generation happens in the background.
• Tools:
  • lmstxt_generate_llms_txt: Trigger llms.txt generation.
  • lmstxt_generate_llms_full: Generate llms-full.txt from an existing run.
  • lmstxt_generate_llms_ctx: Generate llms-ctx.txt (requires llms_txt).
  • lmstxt_list_runs: View recent generation history and status.
  • lmstxt_read_artifact: Read generated files with pagination support.
  • lmstxt_list_all_artifacts: List all persistent .txt artifacts on disk.
• Resources:
  • Run-specific: lmstxt://runs/{run_id}/{artifact_name}
  • Persistent Directory: lmstxt://artifacts/{filename} (e.g., lmstxt://artifacts/owner/repo/repo-llms.txt)

# Default stdio transport
lmstxt-mcp

Add to your MCP client config (e.g., claude_desktop_config.json or config.toml):

{
  "mcpServers": {
    "lmstxt": {
      "command": "lmstxt-mcp",
      "env": {
        "GITHUB_ACCESS_TOKEN": "your_token",
        "LMSTUDIO_BASE_URL": "http://localhost:1234/v1"
      }
    }
  }
}

[mcp_servers.lmstxt]
command = "lmstxt-mcp"
startup_timeout_sec = 30
tool_timeout_sec = 30

[mcp_servers.lmstxt.env]
GITHUB_ACCESS_TOKEN = "your_token"
LMSTUDIO_BASE_URL = "http://localhost:1234/v1"

1. Collect repository material – the GitHub client gathers the file tree, README, package files, repository visibility, and default branch.
2. Prepare LM Studio – the manager confirms the requested model is loaded, auto-loading if necessary.
3. Generate documentation – DSPy produces curated content; on LM failures the fallback serializer builds markdown and JSON directly.
4. Assemble llms-full – curated links are re-fetched via raw GitHub URLs for public repos or authenticated API calls for private ones, with validation to remove dead links.
5. Unload models safely – the workflow first uses the official lmstudio SDK (model.unload() or list_loaded_models), then falls back to HTTP and CLI unload requests.

• src/lms_llmsTxt/ – core generation library, DSPy analyzers, and LM Studio helpers.
• src/lms_llmsTxt_mcp/ – MCP server implementation, asynchronous worker, and resource providers.
• tests/ – pytest coverage for the generator pipeline and MCP server.
• artifacts/ – sample outputs from previous runs.

source .venv/bin/activate
python -m pytest

Using uv:

source .venv/bin/activate
uv run pytest

All tests should pass, confirming URL validation, fallback handling, and MCP resource exposure.

Use this flow to rebuild locally and verify the installed CLI before tagging and publishing.

python3 -m pip install -U pip
python3 -m pip install build
python3 -m build
python3 -m pip install --force-reinstall dist/*.whl
lmstxt --help

Using uv:

uv pip install -U pip
uv pip install build
uv build
uv pip install --force-reinstall dist/*.whl
lmstxt --help

Note: this project uses setuptools_scm, so the version comes from git tags. Before tagging, you will see a post-release + date suffix.

Using uv:

uv pip install -e '.[dev]'

npx @modelcontextprotocol/inspector --config ./inspector.config.json --server lmstxt

Use the payloads in docs/mcp-inspector-payloads.md to verify specific tool behaviors.