Connector tool results return unusable payloads: Browserbase navigate dumps the internal page object, Claude Code Remote list_triggers returns 180 KB on one line

Two connectors on the same account return results that are technically successful and practically unusable, both costing large amounts of context for information the caller cannot act on. Measured 2026-09-14 in Claude Cowork on the web.

1. Browserbase connector: navigate and start return the whole internal page object.

A single mcp__Browserbase__navigate call returns tens of kilobytes: the Playwright page object, the CDP connection, its WebSocket internals (including the signed connectUrl repeated four times), the full response headers, the CSP, and the set-cookie values of the site visited. The useful part is the final URL and the HTTP status. Six navigations in one session cost more context than the work they were for. browserbase/mcp-server-browserbase is archived, so there is no upstream tracker for the server itself.

Ask: return {url, status} (plus a short error on failure), or gate the rest behind an option.

2. Claude Code Remote connector: list_triggers returns one 180 KB line.

Error: result (181,726 characters across 1 line) exceeds maximum allowed tokens.

limit: 15 does not help, because each routine's record carries its full mcp_connections array (every connector on the account, with uuids and per-tool policy) and the entire stored session_request, including the routine's whole prompt. The harness writes it to a file, and the file is a single line, so Read's offset/limit cannot chunk it and every caller has to regex it. The same shape comes back from create_trigger.

Ask: a summary view, or field selection, so listing scheduled tasks returns id, name, schedule, enabled, next run and last run without the connector inventory and the prompt bodies.

Both are ordinary daily calls for an account that runs scheduled tasks and browser automation, so the cost lands on every session.