It used your server and left it alone
A server already listening on 127.0.0.1:11434 is adopted read-only. Kolkrabbi never stops it — it may be your own shell, or a service holding your Ollama Cloud sign-in.
Walkthrough
A coding agent
on Ollama.
Six steps, no account, no key, nothing sent anywhere. At the end you will have made a real edit to a real file with a model running on your own machine, undone it, and seen the turn recorded at $0.00 beside whatever your cloud turns cost.
Commands checked against kolk help on . Kolkrabbi runs on macOS and Linux.
01 / WALKTHROUGH
Six steps.
Steps 1 and 2 are one-time. After that, opening a local session is one command.
-
01 — INSTALL KOLKRABBI
One command. It picks a writable directory on your
PATHand verifies the download's SHA-256 against the release checksums before it installs anything.curl -fsSL https://kolkrabbi.francomichetti.com/install.sh | sh -
02 — HAVE A MODEL
Kolkrabbi does not install Ollama and does not pull models behind your back. Use the Ollama you already have. A coding-tuned model in the 7B–14B range is the usual starting point on a laptop.
ollama pull qwen2.5-coder:14b -
03 — OPEN A SESSION ON IT
Point Kolkrabbi at the local endpoint. No key is involved — and your OpenRouter key, if you have one, is never sent here. It only ever goes to openrouter.ai.
cd your-project kolk --base-url http://localhost:11434/v1 -m qwen2.5-coder:14b -
04 — SEE WHAT THIS MACHINE CAN DO
Inside the session,
/localiareads your actual hardware and the models on disk, and will tell you whether a given model fits before you spend twenty minutes finding out that it does not./localia /localia models /localia plan qwen2.5-coder:14b -
05 — MAKE A REAL EDIT, THEN TAKE IT BACK
Ask for something small and specific. Kolkrabbi checkpoints files before every write, so you can look at exactly what changed and undo it.
add a docstring to the top of the busiest function in this file /changes # which files this session touched /diff # what it did, as a diff /rewind # put the files back, keep the conversation -
06 — RATE IT, THEN READ THE LEDGER
This is the part that pays off later. Rate the turn, and the rating is stored next to the cost and latency in a JSONL file on your disk. After a week you can see which models are actually worth their price.
/rate 4 /stats # calls, tokens, cost, latency, your ratings /dash # the same numbers as a loopback-only page
02 / WHAT JUST HAPPENED
What Kolkrabbi did to your system.
Worth knowing, because it is less than you might expect.
It started one, on its own port
When nothing is listening and ollama is on your PATH, Kolkrabbi starts a server on a loopback port it chooses, with a curated environment, says so once in the transcript, and stops only that process. The server's lifetime is the session's.
On macOS, a killed session can leave it running
Linux gets a death signal, so the server dies with Kolkrabbi. macOS has no equivalent and Windows has no job object yet. A server that outlives a killed session is found and adopted by the next one, which is the safe direction — but it is a real difference and the code says so.
03 / IF IT GOES WRONG
The three things that actually go wrong.
In roughly this order.
The model does not appear
Kolkrabbi probes the literal address 127.0.0.1:11434 and nothing else — deliberately, since OLLAMA_HOST can point at another machine or at ollama.com. If your server is elsewhere, pass it yourself with --base-url. Run /doctor to see what Kolkrabbi can actually reach.
It is unbearably slow, or it swaps
The model does not fit. /localia plan <model> compares a model against your hardware and says so before you commit to it. Drop a size class, or a quantisation.
It edits nothing and just talks
Code mode needs a model that can call tools, and at 8B on 16 GB, none we tried could do it under a real agent prompt. We measured this: three harnesses, three tasks, two runs each, 0 of 18, including Kolkrabbi. If a model will not use the file tools, that is the model, not the harness. Expect to need something considerably larger than 8B for code mode, and treat the walkthrough above as the plumbing check it is.
KEEP READING
Local AI coding agent OpenRouter coding CLI Compared with other agents