MCP limitations

The Cavuno MCP server is a JSON-oriented proxy to the Operator API. It is intentionally not a general-purpose browser, scraper, filesystem, scheduler, or integration platform.

Sandbox boundaries

• Outbound fetch is blocked inside search and execute.
• execute can call only cavuno.request.
• Every call starts in a fresh isolate with no persistent state.
• Environment variables and the local filesystem are unavailable.
• execute can send multipart file uploads with a FormData body, but it cannot read the local filesystem. The bytes must already be available inside the invocation, and callers must let cavuno.request generate the multipart Content-Type boundary.
• CPU and wall-clock time are bounded, so long-running workflows need an external orchestrator.

Product boundaries

An MCP client decides how prompts, approvals, retries, schedules, and other tool connections work. Cavuno does not become a native Ahrefs, Semrush, Firecrawl, ATS, or newsletter integration merely because the same agent can access both systems. Google Search Console is an explicit native exception when it is connected for the Board: MCP can run finalized Search Analytics and indexed URL Inspection reads, but not live tests or indexing requests.

Use the right surface

Use MCP for interactive agent work, the CLI for deterministic local and CI tasks, and the REST API for application code. The MCP cookbook shows the handoff explicitly when another system is involved.