docs(mcp): add anonymous Parallel Search config example

[georgeatparallel]

This adds a commented config example so Chainlit developers can try Parallel web search and page extraction without creating an account or adding an API key.

Chainlit has no automatic search default or bundled web-search MCP provider. Its existing examples are GitHub and generic SSE/HTTP servers, with MCP disabled. This gives developers a working search endpoint to try through that same setup.

I checked all 19 products on Artificial Analysis's Search API leaderboard, using its August 31, 2026 data. Parallel Fast has the lowest time per task at 19.2 seconds, compared with 28.6 for Perplexity Medium, 45.9 for Tavily Basic and 63.0 for Firecrawl. That's about 70% less time than Firecrawl in this benchmark.

The research scores are a good reason to try it too. On DeepSearchQA, Parallel Fast scores 80 F1, ahead of Exa Auto and Brave LLM Context at 78, You.com Highlights at 77, Firecrawl and Tavily Basic at 74, both Keenable modes at 70, and TinyFish Web at 67. Perplexity Medium still leads the overall Search Index, 80 to Parallel Fast's 73. Source: Artificial Analysis.

These are Search API results in AA's fixed harness, where time per task combines derived model time and measured search time. The anonymous MCP uses Fast mode, but AA did not test this Chainlit integration.

The comments cover setup, disconnecting, the inputs sent to Parallel, and anonymous rate limits. MCP stays disabled by default, no server is preconfigured, and user-server permissions stay unchanged.

Validation:

• Repository lint and format wrappers pass for backend/chainlit/config.py with the locked Ruff 0.14.7; git diff --check passes.
• Verified default config generation and load_settings() with the exact uncommented example.
• Exercised the source connect_mcp and disconnect_mcp handlers with a real websocket session and MCP client: disabled gate, named-server connection, anonymous tool discovery, web_search, web_fetch, and cleanup passed.

The focused handler check isolated UI emitter initialization and loaded the handler definitions without the full server module. HTTP routing, browser behavior, full app startup, and the full test suite were not tested. No repository build ran.

I work at Parallel, which operates this service. This contribution was prepared with coding-agent assistance.

---

Summary by cubic

Adds a commented Parallel Search MCP server example to the generated config, showing developers how to enable web search and page extraction without an account or API key. MCP stays disabled by default, no server is preconfigured, and user-server permission settings are unchanged.

^{Written for commit 7d4c4a0. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 1 file

_{Re-trigger cubic}