intervals-mcp-server

A Model Context Protocol (MCP) server for the Intervals.icu API. Connects Claude, Cursor, ChatGPT and any other MCP client to your training data — activities, events, wellness metrics, power curves, gear and custom items.

Zero-config first run: npx intervals-mcp-server@latest walks you through authentication in your terminal and saves everything for future runs.

Requirements

• Node.js ≥ 18.19

• An Intervals.icu account

Related MCP server: Intervals.icu MCP

Quick start

1. Authenticate (once)

npx intervals-mcp-server auth

The wizard will:

1. Open https://intervals.icu/settings in your browser

2. Ask you to copy two values from that page:

   • API Key — the "API Key" row, click Show

   • Athlete ID — shown on the same page (e.g. 123456 or i12345)

3. Verify them against the live API

4. Save them to ~/.config/intervals-mcp-server/config.json (permissions 0600)

5. Print ready-to-paste snippets for your MCP client

2. Add the server to your MCP client

Claude Code

claude mcp add intervals -s user -- npx intervals-mcp-server@latest

Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "intervals": { "command": "npx", "args": ["intervals-mcp-server@latest"] }
  }
}

Cursor — ~/.cursor/mcp.json:

{
  "mcpServers": {
    "intervals": { "command": "npx", "args": ["intervals-mcp-server@latest"] }
  }
}

Done — the client spawns the server via npx and all tools are available.

Why a config file instead of env vars? MCP clients spawn servers with a minimal environment, so shell exports like API_KEY are not inherited automatically. The wizard-written config file works everywhere. Env vars are still supported and take precedence when declared in the client config.

Alternative: environment variables

If you prefer declaring credentials in your MCP client config:

{
  "mcpServers": {
    "intervals": {
      "command": "npx",
      "args": ["intervals-mcp-server@latest"],
      "env": {
        "API_KEY": "your-api-key",
        "ATHLETE_ID": "your-athlete-id"
      }
    }
  }
}

Resolution order: API_KEY / ATHLETE_ID env vars → config file. INTERVALS_API_BASE_URL overrides the API base URL (default https://intervals.icu/api/v1).

Tools (20)

Tool names use stable snake_case identifiers, so prompts and integrations keep working across updates.

Data completeness & provenance

Activities can arrive partially — the activity record lands before all streams and server-side analysis finish. To keep the model from answering confidently from half-synced data, the activity tools surface native API signals and derive a verdict from them:

• get_activities — one-line Sync: verdict per activity (complete / pending / incomplete) plus last-sync time and upstream source

• get_activity_details — full Sync & Data Completeness section: analyzed, icu_sync_date, icu_sync_error, analysis_issues, stream_types, source / external_id

• get_activity_streams — warns when a requested stream type is missing (it may still be syncing) instead of silently omitting it

The verdict is marked (derived) — the API has no explicit completeness flag. Caveat per the Intervals.icu maintainer: analyzed may stay null on Strava-sourced activities while the data is fine; those are reported as pending rather than incomplete.

Stream quirk handled for you: the latlng stream stores latitude in data and longitude in a sibling data2 field; the streams tool zips them into [lat, lng] points so the longitude half isn't lost.

CLI reference

intervals-mcp-server                 # serve over stdio; with no credentials yet and an
                                     # interactive terminal, run the setup wizard instead
                                     # (exits after saving so clients spawn a clean process)
intervals-mcp-server auth            # (re)run the credential wizard
intervals-mcp-server serve           # start the server without the onboarding check
  --transport stdio|streamable-http  # transport (default stdio)
  --host <host>                      # HTTP host (default 127.0.0.1)
  --port <port>                      # HTTP port (default 8765)

Streamable HTTP transport

For remote setups (e.g. behind a tunnel for ChatGPT custom connectors):

npx intervals-mcp-server serve --transport streamable-http --port 8765
# MCP endpoint: http://127.0.0.1:8765/mcp

Development

pnpm install
pnpm build        # tsup → dist/
pnpm test         # vitest (114 tests)
pnpm typecheck    # tsc --noEmit
pnpm smoke        # end-to-end: spawn server, listTools, live API calls
                   # (needs API_KEY / ATHLETE_ID)

Publishing

Releases are published by GitHub Actions (.github/workflows/publish.yml):

# bump "version" in package.json, commit, then:
git tag vX.Y.Z
git push origin vX.Y.Z

The workflow installs dependencies, runs typecheck + tests + build, then publishes to npm with provenance using the NPM_TOKEN repository secret.

License

MIT