Claude Code
done9.52 s
median time a task
replay of run coding.repair-held-devfix-claude-code
Hanzo Dev is a coding agent. It reads your repository, makes the change, runs the tests and shows you the diff — in your terminal, on the web, in the cloud and in Slack.
Hanzo Dev and Claude Code were handed the same 44 coding tasks (single-fault repairs), one try each. Each clock replays that side's median time a task, at the speed it ran.
3.02× faster over the 42 tasks both solved, 95% interval 2.06–4.36×.
9.52 s
median time a task
replay of run coding.repair-held-devfix-claude-code
2.29 s
median time a task
replay of run coding.repair-held-devfix-hanzo-dev-dc
By how Hanzo Dev settled a task:
One agent
The same agent, the same models and the same account in every place it runs.
In your terminal, on the web, in the cloud and in Slack
In any repository. It reads the code, makes the change and runs commands in a sandbox. Bare hanzo opens the same session.
Describe an app. Dev writes it, runs it and deploys it, with the database, the sign-in and the storage already there.
Hand a task to Hanzo’s cloud and close the laptop. dev cloud diff shows what it did; dev cloud apply puts it in your checkout.
An org admin connects the workspace in Dev’s settings, and the team talks to it from their channels.
Install
Linux, macOS and Windows. Dev updates itself the way it was installed.
curl -fsSL https://hanzo.sh | sh
Linux and macOS. One native binary, its checksum verified.
npm i -g hanzo
Every platform, Windows included. Run it as hanzo dev; the agent downloads on first run.
brew install hanzoai/tap/hanzo-dev
macOS and Linux.
Then sign in. Hanzo IAM links the machine to your account.
dev login # pick an account: Hanzo is the default
dev login --hanzo # straight to your Hanzo account
dev login --with-api-key # or paste a Hanzo API key, hidden
Dev starts on enso-auto through api.hanzo.ai, so every call is metered on your Hanzo account. Every model Hanzo serves is a flag or a /model away.
01Models
Kai is Dev’s own controller, not an adviser. At each step it returns an operation Dev can run, or says plainly that it abstains. The cheapest step that holds is the one taken.
02Controller
One candidate, its preconditions hold: it runs, and no model is asked.
Where to fix, what to do and how hard to try, asked of Kai in one structured call.
A model writes the edit, asked for as a strict schema.
The same, given room to reason.
A failure seen before with nothing changed picks a different action, or escalates. It never replays the loop.
Nothing counts as done because a model said so.
03Proof
A step is settled by what the verifier records — the exit status, the tests, the diff — never by what anything expected.
Preconditions are checked just before an action runs, and an edit voids every decision made against the files as they were.
Every decision, fallback, run and verification is a line with a stable id and its reason, credentials redacted, so its time and cost can be read back.
Commands run in a sandbox and are approved inside it. Ask before every action, or lift the fence for one run. A repository’s own tool servers stay off until you say you trust it.
04Sandbox
Hanzo Gen is the harness: Claude Code, Codex, Cursor and the rest, on Hanzo’s tools and models and linked to your account. Dev is the agent it opens when none is named.
05Gen
Open source
Dev is Apache-2.0 on GitHub. Build it, change it, send a pull request — hanzoai/dev.
Questions
Read
Paste and ship
Read https://hanzo.ai/skill.md and use Hanzo Dev in my project. Start with: dev