Braintrust API + MCP vs Navige

Braintrust API + MCP

OpenAPI 3.0.3 spec with 75 paths, 429 and Retry-After declared on 154 operations. Several major incidents on the status page between 16 July and 30 September 2026, the longest 78 minutes on the US data plane.

Navige

Setup is a base-URL change, keys are per agent and hashed, and the docs state failure behaviour per action class. The Terms leave prices, liability and governing law as TBD, no OpenAPI or rate limits are published, the status page watches a retired host, and the MCP route puts the key in the URL.

Braintrust API + MCP

1. Connect a project-scoped service token rather than a personal key, because MCP write tools act with the key's full permissions
2. Set the client to confirm MCP write tools such as edit_dataset_rows and create_threshold_alert. The server has no read-only mode
3. Cache sql_query results. Starter and Pro allow about 20 queries a minute and return 429
4. Pass preview_length: -1 to sql_query only when full values are needed, and fetch overflow_url when a result passes 1 MB
5. Upgrade to Python braintrust 0.28.0 or TypeScript 3.23.1 or later and rotate any provider keys traced before July 2026

Navige

1. Send the Navige key in x-navige-key on the proxy. Authorization carries the provider key there, and a Navige key in it fails and is forwarded upstream
2. Act only on allow or approved from /enforce. Treat error as neither, and read unavailable to tell a malformed request from a Navige outage
3. Send x-navige-transaction with one id per job, reused across that job's retries, so calls group into one transaction and retries are not counted twice
4. On a 429 from the proxy, check for a blocking budget before retrying. Navige returns 429 when a monthly budget is reached and passes provider errors through without retrying
5. Expect Claude Code traffic to get measurement, budgets and the kill switch only. Context trimming, routing, the output ceiling and prompt caching are switched off for it