Introduction

picoo is a tiny local tool runtime for AI agents. Instead of loading a big MCP-style tool schema into context (or burning a multi-turn browser-exploration loop just to click through a form), an agent runs one small declared op:

picoo run github/create-issue --input repo=picoo-dev/picoo --input title="Bug: foo"

What’s an op?

An op (short for “operation”) is a single, narrow, declared action: the smallest unit picoo knows how to run. Concretely, it’s two files: a manifest (picoo.toml) that declares its name, description, and input schema, plus a sibling script that does the work. The contract between them is deliberately boring: JSON in on stdin, {"result": ...} or {"error": {...}} out on stdout, exit code mirrors success. No SDK, no schema negotiation, no persistent server.

Why not just use MCP?

MCP servers are great for discovering capability at runtime. But once an agent already knows exactly what it wants to do (“publish this draft to Medium,” “open a GitHub issue”), re-deriving that action through a live tool-listing handshake and a multi-turn exploration loop is pure token tax. picoo ops are meant to be authored once (by a human, or by an agent driving a browser/API manually) and replayed deterministically forever after.

Install

curl -fsSL https://picoo.dev/install.sh | sh

Then check your machine has what your ops will need:

picoo doctor

Next

• Write your first op
• CLI reference