if it has an api,it has anmcp server.
Today, engineers use heku to keep tool-heavy agents cheap and to wire up services that don't have an MCP server yet — describe the API as JSON, call it in the same session. The same runtime lets agents install and author their own tools, no human in the loop.
Two things.
Every day._
Not a roadmap. This is what engineers use heku for right now.
Cut token usage
Lazy discovery keeps the cold-start manifest flat. Install 5 configs or 200 — the agent only pays for the tools it actually searches for and calls.
tools/list · 529–757 tokens · flat regardless of config countFill the MCP gaps
No MCP server for that service yet? Describe it as JSON — endpoint, auth, a few tool definitions — and it's callable in the same session. No waiting on someone else to ship a server.
heku.create_config → live tool, same sessionOne runtime.
Three ways in._
heku, heku hub, and the console aren't three products — they're one loop. Discover a config on the hub, run it on the runtime, operate it from the console.
Discover on the hub.
Browse or publish connector configs — Stripe, Linear, Slack — and install any of them with one command.
Run on the runtime.
heku loads the config and serves it as a live MCP tool. Lazy discovery keeps the manifest flat as you add more.
Operate from the console.
Chat with your tools, edit configs live, and check auth status — pointed at your local runtime or a hosted one.
Three rules. Zero exceptions.
Agents author their own tools.
heku.create_config lets an agent describe a new integration and start calling it in the same session — no human round-trip.
Context is a budget.
Lazy discovery keeps the cold-start manifest flat and bounded whether you install 5 configs or 200.
Local-first runtime.
Credentials live in .env files on your machine. Nothing routes through us.