thousand-eyes-super-mcp 0.1.0

FastMCP server for the Cisco ThousandEyes API, generated dynamically from the official OpenAPI spec.

thousand-eyes-super-mcp

FastMCP server for the Cisco ThousandEyes API, generated dynamically from the official OpenAPI spec.

ThousandEyes is Cisco's SaaS for end-to-end network and application visibility — synthetic agent tests, BGP monitoring, internet weather, endpoint experience. This server exposes the entire v7 REST API as MCP tools, ready to drop into Claude Desktop, Claude Code, Cursor, or any MCP-aware client.

Status

Read-only by default. See CHANGELOG.md for the current release and what changed.

Install

uv tool install thousand-eyes-super-mcp

Configure

1. Copy .env.example to .env and fill in THOUSANDEYES_BEARER_TOKEN. Generate a token at https://app.thousandeyes.com under Account Settings → Users and Roles → Profile → User API Tokens.
2. (Optional) Edit thousand-eyes-mcp.yaml for runtime knobs — transport, splitting cap, retries, pagination, default account group.

Run

thousand-eyes-mcp                                          # stdio (Claude Desktop)
thousand-eyes-mcp --transport sse --host 0.0.0.0 --port 8000
thousand-eyes-mcp --read-write                             # enable mutating endpoints

CLI subcommands

thousand-eyes-mcp fetch 7.0.88              # download a spec without starting the server
thousand-eyes-mcp fetch --all-known         # download every known version
thousand-eyes-mcp list-versions             # offline: enumerate known + cached versions
thousand-eyes-mcp discover-versions         # [experimental] diff DevNet vs KNOWN_SPEC_URLS
thousand-eyes-mcp --diff 7.0.88 7.0.89      # diff two on-disk spec versions

How it works

• Bearer token auth. ThousandEyes issues long-lived OAuth2 bearer tokens out-of-band; the dispatcher sets Authorization: Bearer <token> on every request. No login flow, no refresh.
• Dynamic tool generation. At startup, the loader merges every spec file under specs/<version>/, groups operations by tag, and splits oversized groups via an adaptive algorithm (section → sub-tag → URL path depth) so no single tool overwhelms the LLM with too many actions.
• Stable action names. Derived from (method, path, tag) rather than upstream operationId, so naming stays consistent across spec releases.
• Pagination auto-follow. Cursor pagination via _links.next.href and offset/limit pagination are detected from spec parameters and stitched automatically. Override per-call with reserved params _max_pages, _page_size, _auto_follow.
• Configurable retry. 429/502/503/504 are retried with exponential backoff + jitter. Mutating methods are not retried by default; opt in with retry_mutating: true.
• Transports. stdio (default), SSE, streamable-http. HTTP transports refuse to bind a non-loopback host with auth.type=none unless --insecure-allow-public is passed.

Docker

docker compose up --build

License

Apache-2.0 — see LICENSE.