You want it working
Install koda, point it at a model, and fix your first bug. Plain English, copy-paste commands, about five minutes.
A coding agent that never leaves your machine.
koda drives the model you already run locally — Ollama, LM Studio, llama.cpp, vLLM, MLX — and any other server that speaks the OpenAI chat API. It reads and edits your code, runs your tests, and shows you a diff before it changes anything.
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/simpletoolsindia/koda/master/install.sh | bash
One binary, ~6 MB·Starts in ~3 ms·No runtime dependencies·Rust 1.82+·MIT
▌ the discount test fails — fix it plan 2/3 ████████░░░░ ✓ read the failing test ✓ fix apply_discount ◐ run the suite ✓ codegraph symbol apply_discount 4ms ✓ read cart.py (21 lines) 2ms ctrl+r ✓ edit cart.py (1 replacement) 3ms │ @@ -14,3 +14,3 @@ │ 14 def apply_discount(amount, percent): │ 15 - return amount - percent │ 15 + return amount * (1 - percent / 100) ✓ $ python3 -m pytest -q → exit 0 380ms Fixed. apply_discount subtracted the percent instead ofapplying it. EXEC ● ready · 3/3 steps @ file · ctrl+p mode · /help ❯
Two ways in
Section titled “Two ways in”You want to know how
One Rust binary, no sidecars. The agent loop, the tool protocols, how a request is curated to fit a small context window.
What it does
Section titled “What it does”Runs on your hardware
Your code is read by a model on your machine. Nothing is uploaded, no account, no metering. Cloud endpoints work too — that is your choice, not the default.
Asks before it acts
Every write shows a unified diff first. Every command asks. PLAN mode removes the
write tools entirely, so nothing on disk can change while you are still thinking.
Runs a real debugger
The debug tool drives your language’s actual debug adapter — breakpoints,
stepping, variables, logpoints. The model reads state instead of guessing at it.
Knows where things live
A symbol graph of the whole project is built on open, so the model asks where
apply_discount is defined rather than grepping for it.
Survives a small context
An 8k window fills up fast. koda supersedes stale reads, squeezes what is left on a ladder, and only then drops — never the opening request.
Remembers between sessions
Facts, procedures and style rules land as plain markdown in your project you can read, edit, commit or delete. Nothing is inferred behind your back.
Pick a starting point
Section titled “Pick a starting point”InstallationmacOS, Linux and Windows, from a one-liner or by hand.
LLM providersOllama, LM Studio, llama.cpp, vLLM, MLX, and any OpenAI-compatible endpoint.
ConfigurationEvery key in config.toml, what it does and what it defaults to.
Tools & safetyWhat the agent can do, what it must ask before doing, and how the sandbox works.
TroubleshootingBad tool calls, connection failures, models that lose the format.
FAQIs it free, is it private, which model should I run.