Skip to content
kodakoda

Quickstart

koda needs two things: the koda binary, and a model for it to talk to. This page gets you both and then puts them to work.

Pointing koda at your server100×30 · 0:13
/setup walks the endpoint, the model and the key, and writes the config for you.

From nothing to a fixed bug

Section titled “From nothing to a fixed bug”

  1. Install koda

    Section titled “Install koda”

    The installer builds from source and offers to install Rust first if you do not have it. It puts the binary in ~/.local/bin — no sudo, no system directories.

    Terminal window
    curl -fsSL https://raw.githubusercontent.com/simpletoolsindia/koda/master/install.sh | bash

    If the bin directory is not on your PATH, the installer tells you the one line to add. Full options — system-wide installs, the interactive menu, building by hand — are on the Installation page.

  2. Get a model running

    Section titled “Get a model running”

    Ollama is the shortest path. Any OpenAI-compatible server works — see LLM providers for LM Studio, llama.cpp, vLLM, MLX and the cloud.

    Terminal window
    brew install ollama          # or: curl -fsSL https://ollama.com/install.sh | sh
    ollama serve                 # leave this running
    ollama pull qwen2.5-coder:14b

    qwen2.5-coder:14b wants roughly 10 GB of RAM. On a 16 GB machine use qwen2.5-coder:7b instead and expect weaker tool use.

  3. Point koda at it

    Section titled “Point koda at it”
    Terminal window
    cd ~/your-project
    koda -u http://localhost:11434/v1 -m qwen2.5-coder:14b

    Or run koda, type /setup, and fill in the endpoint, model and API key in the form — it fetches the model list from whatever URL you type and writes your config, so you never pass those flags again.

  4. Ask for something

    Section titled “Ask for something”

    Type a request the way you would say it to a colleague. koda decides which files to read and which tools to use.

    the discount test fails — fix it

    It will read the failing test, look up the function in the symbol graph, and propose an edit.

  5. Approve the change

    Section titled “Approve the change”

    Before any file is written you get a diff and a prompt:

    edit cart.py
      @@ -14,3 +14,3 @@
    14  def apply_discount(amount, percent):
    15 -    return amount - percent
    15 +    return amount * (1 - percent / 100)
    

    apply this edit?   y once · a always · n deny

    Press y to apply it once, a to stop being asked about this tool for the rest of the session, or n to deny — which tells the model to ask you what to do instead.

  6. Let it check its own work

    Section titled “Let it check its own work”

    Ask it to run the suite. Commands ask for approval the same way.

    run the tests

What to try next

Section titled “What to try next”
  • ctrl+p cycles the mode. PLAN removes the write and command tools entirely, which is the right place to start an unfamiliar task — see Modes & autonomy.
  • Type / to see every command. /settings opens one page holding everything koda can be configured to do.
  • Prefix a line with ! to run a shell command yourself with no model call and no tokens spent: !git status.
  • Mention a file with @ to pull it into the message — @report.pdf summarise this works on PDFs, spreadsheets and screenshots too.
  • /undo puts back every file the agent changed in the last turn.