Proposal: third_party notebook — building a tax agent with verified, source-cited skills (OpenAccountants)

I'd like to contribute a `third_party/OpenAccountants/` notebook demonstrating the **remote MCP connector** pattern for a domain where hallucinated answers are costly: tax.

The notebook shows Claude:
1. connecting to a hosted MCP server via the Messages API MCP connector;
2. loading a **source-cited, accountant-verifiable skill at compute time** instead of relying on training-data tax figures;
3. computing US Schedule C / self-employment tax against current-year figures (e.g. the 2025 $176,100 Social Security wage base, the 92.35% net-earnings adjustment); and
4. handing off to a licensed human for sign-off via the server's `request_accountant_review` tool.

It uses [OpenAccountants](https://www.openaccountants.com) (open source, ~1,000 skills across 192 jurisdictions; a subset section-by-section verified by named CPAs/CAs/EAs, the rest research-verified with primary-source citations) as the knowledge layer. The public MCP endpoint requires no API key to read skills.

This fits alongside the existing `third_party/` integrations (Pinecone, VoyageAI, Wikipedia, WolframAlpha, etc.) and demonstrates a real MCP + verified-knowledge pattern rather than a product pitch. The notebook is drafted and lint-clean (no output cells, no secrets, `ANTHROPIC_API_KEY` from env).

Happy to open the PR — wanted to propose via issue first per CONTRIBUTING. Disclosure: I'm a maintainer of OpenAccountants. OpenAccountants is not affiliated with Anthropic.