I took inspiration from Cloudflares codemode where they took 2500 endpoints and turned them into consumable ~1000 token MCPs for agents to consume.
This basically takes your APIs, databases, and docs and compresses them into 2 MCP tools (~1,000 tokens) instead of N tools (100K+ tokens). Claude was burning tokens (and a lot of them for me) on just tool definitions. This compresses that considerably.
It's a typed SDK layer and not just a RAG for APIs. The search index holds function signatures instead of data.
We dogfood the Express, OpenAPI, and Database adapters daily. But the other adapters are implemented with tests but less mileage -- LAMBDAs is just an idea.
Would love to get feedback -- or even see if it's helpful for other people or if the recent expansion of context windows makes this irrelevant.
I took inspiration from Cloudflares codemode where they took 2500 endpoints and turned them into consumable ~1000 token MCPs for agents to consume.
This basically takes your APIs, databases, and docs and compresses them into 2 MCP tools (~1,000 tokens) instead of N tools (100K+ tokens). Claude was burning tokens (and a lot of them for me) on just tool definitions. This compresses that considerably.
It's a typed SDK layer and not just a RAG for APIs. The search index holds function signatures instead of data.
We dogfood the Express, OpenAPI, and Database adapters daily. But the other adapters are implemented with tests but less mileage -- LAMBDAs is just an idea.
Would love to get feedback -- or even see if it's helpful for other people or if the recent expansion of context windows makes this irrelevant.