cursor-opus-4-7-high
claude-opus-4-7-high
No HTML artifact for this solution. Browse the source.
READMEcoding-agent-ui/cursor-opus-4-7-high/ez-claw/README.md
ez-claw
A tiny coding agent with a browser chat UI. Built for the coding-agent-ui eval.
- Harness: Cursor
- Model:
claude-opus-4-5(Anthropic) — swap viaEZ_CLAW_MODEL - Stack: Python 3.10+ · FastAPI · WebSockets · vanilla JS frontend (no build step)
The agent has these tools:
bash— run a shell command in the working directory.read_file— read a file (with optional line range).write_file— overwrite / create a file.edit_file— exact string replacement in a file.list_dir— list a directory.
Tool use streams live to the browser so you can watch the agent think and act.
Run it
Requires Python 3.10+ and an Anthropic API key.
cd backend
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
export ANTHROPIC_API_KEY=sk-ant-...
# Optional: pin which directory the agent's tools operate in.
# Defaults to the current working directory at launch.
export EZ_CLAW_WORKDIR="$PWD/../scratch"
mkdir -p "$EZ_CLAW_WORKDIR"
python -m ez_claw
Then open http://localhost:8765 in your browser and start chatting.
Configuration
| Env var | Default | Notes |
|---|---|---|
ANTHROPIC_API_KEY |
required | Your Anthropic key. |
EZ_CLAW_MODEL |
claude-opus-4-5 |
Any Anthropic model with tool use. |
EZ_CLAW_WORKDIR |
$PWD |
Sandbox-ish working dir for the agent's tools. The agent can still escape via shell — see warnings. |
EZ_CLAW_PORT |
8765 |
HTTP/WS port. |
EZ_CLAW_MAX_TURNS |
40 |
Max tool-use iterations per user message. |
Warnings
The agent has real shell and filesystem access on your machine. There is no sandbox. Run it under an account / in a directory you're comfortable with. You were warned.
Layout
ez-claw/
├── README.md
├── backend/
│ ├── requirements.txt
│ └── ez_claw/
│ ├── __init__.py
│ ├── __main__.py
│ ├── server.py # FastAPI + WebSocket
│ ├── agent.py # Anthropic tool-use loop
│ └── tools.py # bash / read / write / edit / list
└── frontend/
├── index.html
├── styles.css
└── app.js
Outcome
status
submitted
score
—
verdict
Builds and serves cleanly. Not yet evaluated end-to-end against the acceptance criteria.
Stack
pythonfastapiwebsocketsanthropicvanilla-js
notes
Run with `python -m ez_claw` after exporting ANTHROPIC_API_KEY.