Code ScoutA project of mine.
A free AI coding agent that can run on your own machine.
AI coding got expensive fast. Subscriptions and API bills add up. Most stacks route your repository through someone else's cloud. Code Scout is my answer. It is a native desktop app that can run local language models, so your code stays on your machine.
Download Code ScoutSource on GitHub
- Price
- Free to download, no sign-up
- Runs on
- macOS, Windows, Linux
- Status
- Alpha
§ 1 What it is
Code Scout is a desktop AI coding agent. It is my own project. So read this page as a description by the person who makes it. It is not a review.
You point it at a repository and at a model. It plans the work, runs it with real tools, reads what comes back and goes again. The model can be one that runs on your own computer, so your code stays on your machine. Cloud providers only see what you send them. That only happens if you add your own keys.
§ 2 Why it exists
AI coding got expensive fast. Subscriptions and API bills add up. Most stacks send your repository to someone else's cloud to do the work.
Local models avoid both problems. They bring one of their own, a narrow context window. A local model can see far less of the project at once. Code Scout is built around that limit. I describe it as a serious loop for small models.
§ 3 How the loop works
Code Scout separates planning from execution, so each role works in a smaller, focused context. An orchestrator handles strategy, tool selection and the order of steps. It can be a local or a hosted model, depending on how you set it up. A coder carries those steps out against your repository. It is usually the local model you point it at.
- PlanThe orchestrator decides the steps and picks the tools.
- RunThe coder runs them on your machine: shell, files and git.
- InspectEvery step returns real output: logs, diffs, test results.
- FixThat output becomes the input for the next move.
- RepeatThe loop goes round again, carrying forward only what the next step needs.
Local runs and tool round-trips can be slow or quiet. A visible heartbeat shows whether the agent is running a tool, streaming output or waiting on the model. You are not left wondering whether it is thinking or has hung.
§ 4 What is in it
Real tool access
Shell, files and git in one workbench, so the agent can act instead of only suggesting.
Repo-aware context
Project structure, conventions and notes are kept in plain
.codescoutfiles that the model can reuse.Benchmark mode
Run the same prompts and tool loop against local weights and hosted endpoints. Compare latency, output quality and cost. Then choose the stack for the task.
A visible heartbeat
You can see when the agent is running a tool, streaming output or waiting on the model.
Small context by design
Planning and execution are split. Each step carries forward only what the next one needs.
A native app
Built with Tauri 2 and Rust. It is not an Electron app. It has direct access to the filesystem and the shell.
§ 5 Backends and platforms
- Local backends
- Ollama, LM Studio and llama.cpp, plus any OpenAI-compatible server you point it at.
- Cloud, on your terms
- Add your own keys for hosted providers when you want a hosted orchestrator or a heavier model. The core loop does not need them.
- Platforms
- Native builds for macOS (Apple Silicon and Intel), Windows and Linux.
- Built with
- Tauri 2 and Rust.
- Price
- Free to download. No sign-up required.
- Source
- Available on GitHub.
§ 6 Where it stands
Alpha
Code Scout is in alpha. Expect rough edges and changes from one release to the next. This page gives no benchmarks, version numbers or user numbers. For the current build, see the release notes on GitHub.
Disclosure. I make Code Scout. This page describes my own product.
Free to download, no sign-up
Download Code Scout
Pick the build for your platform from the latest release on GitHub. The source code is public too.
Latest release on GitHubSource on GitHub
These links leave this site. I checked this page against the GitHub repository on 3 October 2026.