KuroCode

In development

Nothing leaves this machine.

KuroCode is a desktop coding agent that works against models running on your own hardware. It reads your repository, edits files, runs commands and drives git — and every one of those steps waits for your approval.

Session transcript
invoice-apifeat/json-exportqwen3-coder:30b · localhost:11434
  1. ›Add a --json flag to the export command, with tests.
  2. Reading the command and its tests…
  3. readsrc/cli/export.ts
  4. editsrc/cli/export.ts+18 −4
  5. edittest/export.test.ts+26

  6. Run this command?

    runpnpm test export

  7. runpnpm test export
  8. 14 passed in 2.3s
  9. Flag added. Tests cover both the object and array shapes.
  10. Stopped. The command never ran, and no process started.
  11. That is the whole idea: the agent proposes, you decide.

Try it: the agent is waiting for permission.

local

No account. No telemetry. No round trip.

Most coding agents are a client for someone else’s server. KuroCode is the whole thing — the agent loop runs inside the app on your desk.

  • Your code stays on your disk

    Files are read and written straight from the local filesystem. Nothing is uploaded to be indexed, embedded or reviewed.

  • There is nobody to sign up to

    No account, no license server, no session that expires because a backend went down.

  • The app does not phone home

    No analytics, no crash pings, no usage counters. The only network traffic is the model endpoint you configured.

models

Bring your own model.

Point KuroCode at whatever you already run. Local endpoints are the default path; cloud keys work too, if you decide you want them.

  • OllamaLocal
  • LM StudioLocal
  • OpenAI-compatibleAny endpoint
  • AnthropicAPI key
  • OpenAIAPI key

Model and reasoning effort are set per session, so a throwaway question and a refactor of the build system do not have to share a model.

permissions

Every tool call waits for you.

The agent reads files, searches the tree, edits, patches, runs shell commands and fetches URLs. What it may do without asking is your call, and you can change it mid-session.

Ask
Every edit and every command stops for approval. The default, and the one to use on a repository you care about.
Auto-edit
File edits go through; shell commands still stop and wait. Good for a long refactor you are watching.
Bypass
Nothing stops. For a scratch worktree where the worst case is `git checkout .`.

There is also a plan mode: the agent investigates and writes down what it intends to do, without touching a single file.

workspace

Built for repositories, not for snippets.

A session is attached to a project on disk, with the tools you would otherwise be alt-tabbing to.

  • Git and worktrees

    Branch, diff and commit locally. Give a session its own worktree so two tasks never fight over the same files.

  • Editor and diffs

    Read what changed in a real editor view before you accept it.

  • Terminal

    A live shell in the same session, so approved commands and your own commands share one context.

  • Streaming transcript

    Output as it happens — tokens, tool calls, results — not a spinner and then a wall of text.

remote

Pick it up from your phone.

A session started at your desk can be followed from another device on your own network. The desktop app stays the engine; the phone is only a window.

  • Off by default. You turn the gateway on in settings when you want it.
  • Pairing is a QR code, and each device gets its own token you can revoke on its own.
  • Traffic stays on your network — there is no relay in the middle.

faq

Questions

When can I download it?

It is in active development and not downloadable yet. That is exactly what the list is for: one email, on the day the first build ships.

Does my code ever leave my machine?

Only to the model endpoint you configure. Point it at Ollama or LM Studio on localhost and nothing crosses your network card. Add a cloud API key and prompts go to that provider, like any other client.

Which models actually work?

Anything served by Ollama, LM Studio or an OpenAI-compatible endpoint, plus Anthropic and OpenAI directly. Development is validated against a 30B coding model running on a local box.

Which platforms?

It is an Electron desktop app, so macOS, Windows and Linux. Which of those get a signed installer first will be in the launch email.

What will it cost?

Pricing is not settled yet, and the launch email will say plainly what it is. The app is proprietary and built on Apache-2.0 foundations, which are credited in the repository.

launch

Get the launch email

One message when the first build is downloadable. Nothing else.

Your address is used for the launch announcement and nothing else. It is never sold or shared, and one reply asking to be removed is enough.