If your AI agents run Gemini on Vertex AI, you can now point those calls at condense and let it manage their context. They still run on your own Google account, with your own credential and the same request format. This went live on September 30.
Swap your Google host for this base URL, keep the rest of the path, and add your condense key in the x-condense-auth-token header.
https://api.condense.chat/vertex
Calls to :generateContent and :streamGenerateContent?alt=sse get compressed. You can send a Google access token as a bearer token, or an express mode key in x-goog-api-key or ?key=.
To manage context, condense compresses the earlier turns of long conversations and adds one short instruction at the end of your systemInstruction. Your own system text stays first and unchanged, so Gemini's implicit cache keeps working. A request that names an explicit cache with cachedContent doesn't get the instruction. Short conversations aren't compressed, and responses come back in Gemini's format.
A few things are forwarded unchanged and aren't compressed: :streamGenerateContent without alt=sse, countTokens, cachedContents and Vertex's OpenAI-compatible endpoint. To get compression, stay on Gemini's native format. This route is for Vertex AI only, so calls to the Gemini Developer API at generativelanguage.googleapis.com aren't covered.
The API reference has the full details.