Your agent's data, without the dump.
Caveman Connect gives an agent token-efficient access to 850+ connectors — OAuth apps, MCP servers, REST APIs and your own functions — as one tool on the wire instead of a catalog in the prompt.
git clone https://github.com/JuliusBrussee/agent-sdk.gitRetrieval answers by the pound. You pay for every token it dumps in.
The usual way to connect an agent to data is to retrieve chunks and paste them into the prompt. Recall improves as the pile grows, so the pile grows — and the model rereads all of it on every turn that follows.
Most of what that costs is context the answer never used: neighbouring chunks, repeated headers, whole documents fetched for one field.
Anything in. One schema on the wire, one tool out.
- 850+ connectors
- Caveman Connect reaches OAuth apps, MCP servers, REST APIs and your own functions from inside the Agent SDK. Anything in.
- one schema on the wire
- The catalog does not ride in the prompt. The model sees one bounded tool and reaches what it needs from inside it.
- typed and capped
- Every call is validated against its schema, budgeted, metered and abortable — the smaller wire is not a looser one.
- on the receipt
- Every run hands back one receipt: model calls, tool calls, subagents, catalog cost basis, stop reason and resume state.
The shape of it. Your tools, your keys, one tool out.
OAuth apps · MCP servers · REST APIs · your keys
your loop, exact-pinned adapters
850+ connectors are reachable through Caveman Connect. The marks above stand for the tools you already run and the keys you already hold — they are an illustration, not an endorsement or an integration claim.
What powers it.
Connect the data.
Skip the dump.
git clone https://github.com/JuliusBrussee/agent-sdk.git@caveman-ai/agent v0.2 documents current main and is not on npm yet, so clone the repo to test it. Token reductions from the local proxy are estimates labeled inferred, never verified savings.