# Mycelium

**Your AI agent should think before it codes.**

AI has made building cheap. It hasn't made *deciding* cheap. Agents will jump from an idea to a pull request without asking why, who for, or whether anyone needs it.

The gap shows up the same way across every AI-native team: the agent is fast, confident, and glad to build something nobody asked for. Mycelium puts the deciding back. It doesn't replace your judgment — it gives the agent enough feedback that the judgment that ships is still yours.

**Other tools accelerate delivery. Mycelium makes the agent earn the right to start.**

Built using itself, and released as open source.

**Prerequisite:** Claude Code, signed in — or another supported agent (see [install paths](docs/install-paths.md)). Then, inside Claude Code:

``` /plugin marketplace add haabe/mycelium /plugin install mycelium@haabe-mycelium /mycelium:start # one command: setup + 10-minute discovery ```

Plugin install is brownfield-safe; no project-root files are modified. Skills are namespaced `/mycelium:<name>`, and `/myc<Tab>` expands the prefix. Legacy install + migration: [`docs/install-paths.md`](docs/install-paths.md).

This README orients you and gets you installed. Full docs live at [`docs/`](docs/README.md): mental model, how-to guides, theory grounding, receipts.

## What it does

You have an idea. You run `/mycelium:start`. The agent doesn't open an editor; it asks four questions. What's the problem, who has it, what's the riskiest thing you're assuming, and what's the smallest move that would test it. Ten minutes in, you have a written brief and the agent points to the riskiest thing you assumed and asks if you want to test it before building anything.

You can say no. A weekend hack gets lighter prompts than a team product, and you can decline depth at any step. What the agent won't do is silently skip past missing evidence and call the work done. It stops where you'd want to be stopped.

## What it feels like

Three moment