31,762 downloads, 740 tool calls: shipping the same tools for agents
Shipping the same tools twice, once for people and once for agents
Who this was for. Us. Mamba Labs builds data tools and sells them on a public marketplace. This is the second way we ship them.
The drawing
The problem
Our tools were only reachable through the marketplace. An AI agent couldn't call one without a custom integration for that tool. 47 tools meant 47 integrations, and nobody writes 47 integrations.
What we built
A thin agent server for every actor, plus one suite server that exposes the whole set through a single connection. An agent connects once and can discover and call any tool.
Every server passes a container build check that runs a protocol handshake before release. The rule exists because a server that installs but can't complete a handshake looks shipped and does nothing.
All of it is published to a public package registry and listed in three third party directories. We track downloads, tool calls, and directory grades weekly.
What it produced
Ten weeks of weekly data:
- 47 agent servers live.
- 31,762 package downloads across 44 servers.
- 740 tool calls across 47 servers, on one directory that reports calls.
- The suite server exposes 24 tools through one connection.
Look at the gap. 31,762 downloads against 740 calls. Downloads are not use. A tool call is someone's agent actually doing work.
We publish both numbers because the second one is the one that matters. It's small, and it's real.
Two publishing failures needed repair passes along the way. Both are documented.
Try it
Every tool prints its price per result. Connect the suite server and your agent sees the whole set. If your agent needs a tool we don't ship yet, reach out and tell us what the row should carry.
Related tools: Apify, Pinecone, GTM Suite agent server
Some of the links on this page are affiliate links.