iclaw 0.1.0

Autonomous iterative coding agent — builds until the vibe is right

InductiveClaw

Autonomous iterative coding agent — builds until the vibe is right.

InductiveClaw wraps the Claude Agent SDK in a persistent autonomous loop. Unlike Claude Code, which stops after completing one task and waits for the next prompt, InductiveClaw continuously builds, verifies, evaluates, and iterates on a software project until it reaches a quality threshold or exhausts its budget.

Installation

pip install iclaw

For screenshot-based visual evaluation (optional):

pip install iclaw[screenshot]
playwright install chromium

Requires the Claude Code CLI (npm install -g @anthropic-ai/claude-code) because the Claude Agent SDK shells out to claude for OAuth. Run claude login once before launching InductiveClaw if you plan to use a Max/Pro subscription.

Authentication

InductiveClaw prefers OAuth (your existing Claude Code / Max subscription login) by default, falling back to an API key. OAuth works via the Claude CLI, so you need claude installed and claude login completed on the machine running iclaw.

Option 1 — Max/Pro subscription (recommended):

claude login   # one-time setup
iclaw -g "your goal"

Option 2 — API key:

export ANTHROPIC_API_KEY=sk-ant-...
iclaw -g "your goal"

Option 3 — Explicit key:

iclaw --api-key sk-ant-... -g "your goal"

Usage

# Build a game from scratch
iclaw -g "Build a roguelike deckbuilder with pixel art style"

# Continue on existing project
iclaw -p ./my-game -g "Polish visuals, add sound effects, improve the tutorial"

# Quick prototype with low quality bar
iclaw -g "Simple snake game in the browser" --threshold 5 --max-iterations 10

# Disable screenshots for non-visual projects
iclaw -g "Build a CLI tool for CSV analysis" --no-screenshot

How It Works

1. You give InductiveClaw a goal
2. It initializes the project, creates a backlog, and builds the first feature
3. Each iteration: orient → plan → build → verify → evaluate → document → continue
4. The loop runs until quality threshold is met, max iterations reached, or you interrupt with Ctrl+C

Project state is persisted on disk (BACKLOG.md, EVALUATIONS.md, source files), so you can stop and resume at any time.

CLI Reference

usage: iclaw [-h] -g GOAL [-p PROJECT] [-m MODEL] [-t THRESHOLD]
             [--max-iterations N] [--eval-frequency N]
             [--use-api-key] [--api-key KEY]
             [--no-screenshot] [--port PORT]
             [--dev-cmd CMD] [-q] [-v]

Flag	Default	Description
-g, --goal	required	What to build
-p, --project	./project	Project directory
-m, --model	SDK default	Model override
-t, --threshold	8	Quality threshold (1-10)
--max-iterations	100	Max outer loop iterations
--eval-frequency	3	Self-evaluate every N iterations
--use-api-key	off	Prefer API key over OAuth
--api-key	env	Explicit API key
--no-screenshot	off	Disable visual evaluation
--port	3000	Dev server port for screenshots
--dev-cmd	auto	Dev server command (parsed but not consumed by the runtime)
-q, --quiet	off	Suppress verbose agent reasoning output (no effect unless -v is also set)
-v, --verbose	off	Print the agent's reasoning/tool output; --quiet disables this mode

Architecture

inductiveclaw/
├── __init__.py
├── __main__.py     # CLI entrypoint (argparse)
├── auth.py         # OAuth-first authentication
├── agent.py        # Outer autonomous loop + Agent SDK calls
├── tools.py        # Custom MCP tools (backlog, evaluate, screenshot, docs)
├── config.py       # Configuration + usage tracking
└── display.py      # Terminal output formatting (Rich)

License

MIT