Build an AI agent, test it on your own computer, and publish it so others can find it.

The demo

A four-line login function looks fine at a glance. The Code Reviewer agent, running on a local model, is asked to review it before it is merged.

def login(conn, username, password):
    query = f"SELECT id FROM users WHERE username = '{username}' AND password = '{password}'"
    row = conn.execute(query).fetchone()
    return row is not None

Do not merge.

Critical: SQL injection on line 2. Typing ' OR '1'='1' -- as the username comments out the password check, so it logs in without a password.

query = "SELECT id FROM users WHERE username = ? AND password = ?"
row = conn.execute(query, (username, password)).fetchone()

This is what the review found, written out cleanly. The agent's own wording varies from run to run, and the attack string above is the working form of the one it suggests. Then the agent is published to a registry and found again by searching for it. The model is Qwen2.5-Coder 7B running through Ollama on the same laptop, so the review is free and the code never leaves the machine.

How it is built

What I built

• The desktop appStudio for writing and testing agents, Explore for finding them, and the screens around them. React and TypeScript in a Tauri shell.
• The background serviceA Rust service that runs agents, keeps the identity keys, checks what each agent declares it may do, and signs releases and run receipts.
• The agent language.agw: a small language for defining an agent, its instructions, inputs and outputs, with a parser and checker (agw check).
• The registryA Go service where signed agent releases are published, searched and resolved.
• Local and hosted modelsRuns on a local model through Ollama, or on hosted models with your own key.
• TestingAbout 2,800 unit tests in the background service and about 3,500 in the app, as of September 2026.

Get in touch