The gateway
OpenRouter is the door to most of the catalogue — GPT, Gemini, DeepSeek, Qwen, Kimi and the rest — behind a single key. Kolkrabbi treats it as the default route, and starts on a model that costs nothing before you have decided anything.
Commands checked against kolk help on .
01 / FREE FIRST
The zero-config path is a real one, not a trial.
For a new session with no model chosen, Kolkrabbi asks OpenRouter's live catalogue for zero-cost, tool-capable models and prefers the strongest coding-oriented one. If the catalogue is unreachable, it falls back to OpenRouter's guaranteed zero-cost openrouter/free router.
A saved model, -m, or a custom effort-tier map beats discovery every time, because those are explicit decisions and discovery is a guess.
/key stores it. It only ever goes to openrouter.ai — never to a local endpoint, never to a --base-url you pass.
02 / THE DIAL
This is the part that is unlike other agents. Elsewhere, effort means how many thinking tokens one model spends. Here it decides which model runs the turn, and how wide the agent work goes.
| Effort | What it selects |
|---|---|
low | A cheap, fast model. Routine edits, renames, questions about the codebase. |
medium | A balanced model. The default working tier. |
high | A strong reasoning model, for work that has bitten you before. |
max | A frontier model, and wider agent orchestration. |
03 / THE LEDGER
Every call is appended to a JSONL file you own. Nothing is uploaded, and there is no account to log into.
/stats aggregates per model: how many calls, how many tokens, what it cost, how long it took, and the 1–5 ratings you gave with /rate. /dash serves the same numbers as a loopback-only page.
The status line carries mode, model, effort, context use and what the session has cost so far, so the number is in front of you while you decide, not afterwards.
Gateway turns, local turns and subscription turns land in the same file. Subscription turns are labelled API-equiv. and never pooled with metered spend — a Max subscriber's marginal cost is zero, and a number that pretends otherwise is worse than no number.
KEEP READING
Local AI coding agent Ollama coding agent Compared with other agents