⏱️ speedrun-mcp

A Model Context Protocol server for speedrun.com — let an AI assistant query games, categories, leaderboards, world records, players and their personal bests, and (with an API key) submit and moderate runs.

"What's the current Super Mario 64 16-star world record, and who holds it?"

Built on speedrun.com's official REST API. The read tools need no account or API key — add a key (see Authenticated features) to unlock identity reads and, optionally, run submission and moderation. Results come back as compact, model-friendly JSON (player ids resolved to names, durations formatted, subcategory variables labeled).

Example

Ask "the SM64 16-star world record?" and the model calls get_world_record, which returns resolved JSON. An excerpt:

{
  "game_name": "Super Mario 64",
  "category_name": "16 Star",
  "world_record": {
    "players": ["Suigi"],
    "time": "14m 35.5s",
    "date": "2023-03-22",
    "video": "https://youtu.be/1_vkwkniHuI"
  }
}

Tools

A typical flow: search_games → list_categories (and list_variables for subcategories) → get_leaderboard / get_world_record. Use list_platforms / list_regions when you need an id for the platform / region filters.

With write tools enabled (see below), submit_run, verify_run, reject_run, set_run_players and delete_run are also available.

Result scope and provenance

Search tools, list_runs, list_unverified_runs, and list_notifications return an object with results, returned, offset, limit, has_more, and next_offset. This replaces their earlier bare-list output. Pass next_offset as offset to continue with the same filters. has_more: null means the API omitted pagination metadata, so completeness is unknown. get_series.games uses the same envelope; continue with game_offset. pagination_note explains the continuation evidence: an API next-page link can lead to an empty page and does not establish a total result count.

get_game_records follows all pages. Notifications scan up to scan_limit source records and report scanned; an empty unread result is not evidence of no unread notifications when has_more is true or unknown.

Run rows preserve account/guest identity in player_details, all video links in videos, and separate source commentary in video_text. Personal-best rows include game/category IDs, level, and raw variable choices. Variable details and subcategory maps are keyed by variable ID to avoid collisions between names.

Leaderboard applied_filters includes the API's system filters and resolved variables; requested_filters separately records the call's filters, including historical dates. A missing requested timing is marked unavailable rather than replaced by the primary time. Each displayed time identifies its source field. returned_runs and omitted_from_response count rows from the fetched response, not the full leaderboard.

Write errors that leave completion uncertain explicitly warn against automatic retries. A success response that cannot be parsed preserves its HTTP status and resource location when supplied.

Install & run

Requires Python 3.10+.

# from PyPI
pipx install speedrun-mcp        # or: uv tool install speedrun-mcp

# from source
git clone https://github.com/williamcodes/speedrun-mcp
cd speedrun-mcp
pip install -e .

The server speaks MCP over stdio:

speedrun-mcp          # console script
python -m speedrun_mcp # equivalent

Use with Claude Desktop / Claude Code

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

{
  "mcpServers": {
    "speedrun": {
      "command": "speedrun-mcp"
    }
  }
}

If you installed from source into a virtualenv, point command at that interpreter, e.g. "command": "/path/to/.venv/bin/speedrun-mcp".

For Claude Code:

claude mcp add speedrun -- speedrun-mcp

# with authenticated features (optional):
claude mcp add speedrun \
  -e SPEEDRUN_API_KEY=your-key-here \
  -e SPEEDRUN_ENABLE_WRITES=1 \
  -- speedrun-mcp

One-click install in Claude Desktop

Each GitHub release ships a speedrun-mcp-<version>.mcpb bundle. Download it and open it with Claude Desktop (double-click, or Settings → Extensions → Install from file). Claude Desktop installs Python and the dependencies itself; nothing else is needed. The extension settings expose the optional API key and the writes toggle described below.

Authenticated features

An API key is entirely optional. With no key, the server exposes only the public read tools (leaderboards, games, players, the moderation queue) and works exactly as described above — no account required. Adding your key unlocks more:

Read-only is the default. Just adding a key never changes anything on speedrun.com — you get identity reads, and everything keeps working perfectly. If a write tool is invoked while writes are off, it doesn't silently fail; it returns:

This server is in read-only mode, so this write action is disabled. To allow run submission and moderation, set the environment variable SPEEDRUN_ENABLE_WRITES=1 (alongside SPEEDRUN_API_KEY) and restart the server.

So the way to switch to read-write mode is always discoverable from the error itself.

Getting your API key

1. Log in to speedrun.com.
2. Go to your account settings.
3. In the left-hand nav, find the Developers section and click API Key.
4. Copy the key shown there.

Treat the key like a password — anyone who has it can act as you on speedrun.com. If it ever leaks, regenerate it from that same page.

Using your key

Add the key to your MCP client config under env. It is read only from the environment — never passed as a tool argument — so it can't leak into the model's context or transcripts. Add SPEEDRUN_ENABLE_WRITES=1 only when you want writes to actually run; with the key alone you stay safely read-only.

{
  "mcpServers": {
    "speedrun": {
      "command": "speedrun-mcp",
      "env": {
        "SPEEDRUN_API_KEY": "your-key-here",
        "SPEEDRUN_ENABLE_WRITES": "1"
      }
    }
  }
}

Or with Claude Code:

claude mcp add speedrun -e SPEEDRUN_API_KEY=your-key-here -- speedrun-mcp
# add -e SPEEDRUN_ENABLE_WRITES=1 as well if you want the write tools

Keep the key out of version control — put it in your client config or a local, git-ignored .env, never in a committed file. All tools carry MCP read-only / destructive hints so clients can flag the write and moderation actions.

Local environment file

Copy .env.example to .env and fill in the settings you need. The template leaves the API key empty and disables writes. Git ignores .env.

The server reads exported environment variables and does not load .env automatically. To load your local file and start the server from a shell:

set -a
. ./.env
set +a
speedrun-mcp

Notes & limits

• Reads need no key; writes are opt-in. Leaderboards, games, players and the moderation queue are open reads. Run submission and moderation need SPEEDRUN_API_KEY and SPEEDRUN_ENABLE_WRITES (see above).
• Rate limit: speedrun.com allows 100 requests/minute per IP and responds with HTTP 420 when exceeded; the client surfaces a clear error if you hit it.
• Game and category arguments accept either an id (o1y9wo6q) or an abbreviation (sm64). For precise subcategory leaderboards (e.g. 16 Star), discover the variable/value ids with list_variables and pass variables={variable_id: value_id}.
• Errors are explanatory. Invalid ids/filters raise an error that includes speedrun.com's own message — e.g. passing a level to a full-game category returns "The selected category is for full-game runs, but a level was selected."

Output shape

• Times reflect the leaderboard's sort timing. When you pass timing (realtime / realtime_noloads / ingame), the reported time / time_seconds match that ranking, not the game's default timing.
• get_leaderboard returns returned_runs (the number of runs returned, bounded by top and ties — not the full board size) and a runs list with resolved player names, formatted times, and labeled subcategories.
• get_world_record returns world_record (the place-1 run, or null if the board is empty) plus tied (a list of any other runs sharing first place).
• get_user_personal_bests returns returned (how many came back, capped by limit) and total_available (the player's true PB count), plus the personal_bests list with game/category names and resolved players.

Development

The package uses a flat src/speedrun_mcp/ layout:

• client.py handles HTTP requests, API errors, and pagination.
• format.py converts API payloads into tool results without network access.
• server.py owns MCP tools, configuration, and the shared client's lifecycle.
• __main__.py provides the python -m speedrun_mcp entry point.
• __init__.py exposes mcp on demand, so importing the client or format helpers does not initialize the server.

Tests live in tests/. Unit tests cover each layer; package and MCP protocol tests cover imports and startup. test_live.py contains the live API checks, selected with the network marker.

pip install -e ".[dev]"
pre-commit install

# The checks run by CI:
ruff check .
ruff format --check .
mypy
pytest -m "not network"

# Apply safe lint fixes and formatting locally:
ruff check --fix .
ruff format .

# Optional: include live API tests.
pytest

# Build the Claude Desktop bundle (needs Node for npx):
npx -y @anthropic-ai/mcpb validate manifest.json
npx -y @anthropic-ai/mcpb pack . speedrun-mcp.mcpb

Releasing

Bump the version in pyproject.toml, server.json and manifest.json (a test checks they agree), tag v<version>, and publish a GitHub release. The release workflow then runs CI, publishes the wheel to PyPI, attaches the .mcpb bundle to the release, and publishes the new version to the MCP registry with both the PyPI package and the bundle listed. Registry auth uses GitHub OIDC, so no token is stored; PyPI verifies ownership through the mcp-name comment at the top of this README.

Ruff checks source and tests for common bugs, security issues, async mistakes, overly complex functions, and pytest mistakes. It also sorts imports and formats Python code. Tests may use assert; the other lint rules apply to both source and tests. Print and debugger statements are rejected because the server uses stdout for the MCP protocol.

Ruff is pinned to the same version in the dev dependencies, its required-version setting, and pre-commit. Update all three together. The hooks apply safe lint fixes and formatting before running mypy; CI checks without changing files.

License

MIT