# AgentCanvas

> A friendlier way to steer the app an AI is building for you. AgentCanvas turns any app/codebase into a plain-English map of what it does, lets non-technical people change how it works in their own words, and turns each change into a structured request that an AI coding agent implements.

## What it is

AgentCanvas is a local, agent-agnostic tool. It reads a workspace and presents the app's behaviour as simple **journeys** made of **When / Do / If** steps (with if/else branches), readable plain-English flows, not code. A person edits a step ("add a step", "change what happens", "add a rule", "remove a step") in their own words. Nothing is mutated directly: each edit is written as a pending change request.

## Who it's for

Non-technical people (operators, founders, power users) who want to understand and steer an app an AI agent is building, without reading code.

## How an AI coding agent uses it

1. Start it: `agentcanvas start --workspace .` (install: `pip install agentcanvas`). It serves a local web canvas and prints a URL.
2. The user edits flows in the browser. Each edit is written to `.agentcanvas/pending/` as a plain-English request: a `.md` (human/agent brief) and a `.json` (structured fields).
3. For each pending request: read it, implement the change in code, run the relevant tests, then re-index with `agentcanvas index --workspace .` and report what changed.
4. Keep polling `.agentcanvas/pending/` for new requests.

Works with Claude Code, Codex, Cursor, Antigravity, or any agent that can read files and run commands. Copy-paste fallback is always available.

## Key facts

- Local-first, runs from the command line.
- Agent-agnostic (no vendor lock-in).
- The source of truth is the `.agentcanvas/` folder in the workspace.
- Edits are requests, not direct code changes, safe and reviewable.
