Generic tool names collide the moment a client connects two documentation servers — two search_docs tools and the model picks wrong. Generating the product name into the tool name solves that at the source, which matters more as these become common. The per-call analytics redaction is the other thoughtful piece: a query can contain something confidential, and being able to suppress it per call rather than per account is the right granularity.
Kapa's own public documentation server. Customers get their own subdomain; this is the one Kapa runs for its own docs. The tool names are generated per product, so they read as search_<product>_knowledge_sources.
- Search_<product>_knowledge_sources — agentic retrieval across every connected knowledge source, returning passages with their source URLs
- Get_<product>_knowledge_documents — the full document by its exact source URL
- give_feedback — the calling agent files feedback on the product, the documentation or the server itself
OAuth 2.1 bearer on the docs server. Customer servers can instead use an API key or an internal Kapa login. With an API-key server, a per-call flag redacts the query text from Kapa's analytics.
Paste a URL, then authorize — add the endpoint to your client, then approve the OAuth consent screen