What koda is
koda is a coding assistant that lives in your terminal. You describe what you want in plain language — “the discount test fails, fix it” — and it works out which files to read, proposes the change, and, once you approve, makes it. It can run your tests, search your project, drive a debugger, and explain what it did.
The difference from a cloud chatbot is where the model runs. koda talks to an OpenAI-compatible chat endpoint, and the endpoint it expects by default is one on your own machine: Ollama, LM Studio, llama.cpp, vLLM, MLX. Your code is read by a process you started, on hardware you own. Cloud endpoints work identically — OpenAI, OpenRouter, Groq, Together, Mistral, anything speaking the same API — but that is a decision you make, not the default you inherit.
What you get
Section titled “What you get”One binary. About 6 MB, starting in roughly 3 milliseconds, with nothing to install alongside it. No sidecar process per tool, no runtime, no package manager at run time. The TUI, the agent loop, the tools, the symbol graph and the HTTP client are all compiled into the same Rust executable.
Approval before action. Every file write shows you a unified diff before it is applied
and again afterwards. Every shell command asks. In PLAN mode the write and command tools
are not offered to the model at all, so nothing on disk can change while you are still
deciding what to do.
Tools that do real work. Reading and editing files, searching, running commands — but also a symbol graph of your project, a real debug adapter for stepping through a running program, subagents with their own context windows, durable memory, web search and fetch, a headless browser for pages that need JavaScript, and PDF/Word/Excel reading.
Deliberate handling of small context windows. A local model with an 8k window spends most of it on tool results it can no longer use. koda curates each request instead of truncating the conversation — see Prompt & context.
What it is not
Section titled “What it is not”koda does not have a hosted service, an account, or a subscription. There is nothing to sign up for and no telemetry. It also does not bundle a model: you bring the endpoint, and which model you pick matters more to the result than anything else on this site.
It is built and tested primarily on macOS with Apple Silicon. Linux and Windows are supported by the installers and the code, and the binary has no platform-specific runtime requirements.
Two editions
Section titled “Two editions”koda ships on two branches. Both are maintained; neither is merged into the other.
| Edition | Branch | What is different |
|---|---|---|
| Official | master |
The standard build. browse drives agent-browser, which koda ships and installs itself. |
| Uncensored | uncensored |
Everything in Official, plus a stealth browsing stack so the browse tool is not flagged by bot-detection. Drives your real Google Chrome, non-headless. |
To switch an existing checkout, git checkout master or git checkout uncensored, then
re-run ./install.sh. The install one-liners honour KODA_BRANCH if you want a specific
branch.
Where to go next
Section titled “Where to go next”If you want it running, start with the Quickstart — install, a model, and a first fix, in about five minutes. If you want to understand the machine before you run it, start with Architecture.