Tiny tools for AI agents. Less context, less compute.
Give an AI agent a bunch of tools and it has to read every single one of their descriptions before it can do anything — even the ones it never ends up using. That can be tens of thousands of tokens spent before a single useful thing happens. And for a lot of real tasks (posting to a website, filling out a form), agents end up clicking around a browser step by step instead, burning even more tokens and time just figuring out where things are.
Say you want an agent to publish a post to Medium. Without Picoo, it has to click through the page step by step every time — open the editor, type the title, type the body, find the publish button, confirm — that's many turns, and it can click the wrong thing.
Write that whole flow once, as a tiny script. From then on, the agent does it in a single call:
picoo run alice/medium/publish --input title="My Post" --input body="Hello world"
No page to click through. No tool list to read first. One script, one call, done — the agent only pays for exactly the thing it's using, when it's using it.
cargo build --release
Create an op:
picoo create alice/hello/world --runtime python --description "say hello" --input name:string
Fill in the generated script (~/.picoo/ops/alice/hello/world/main.py), then run it:
picoo run alice/hello/world --input name=world
Full command reference and conventions: see SKILL.md.
Early days — v0.1. The core loop (create, run, list, search) works. Sandboxing and a shared registry are still on the way.
MIT