Kolkrabbi — AI skill for Claude Code

Kolkrabbi

**Chat, code, and ordered agents in one fast CLI — any model, any provider, with a 100% local rating dashboard.**

Kolkrabbi (binary: **`kolk`**) is an open-source, model-agnostic AI coding agent for the terminal — an alternative to Claude Code and Codex CLI. It runs any model on [OpenRouter](https://openrouter.ai), *your own* Claude or ChatGPT subscription through that vendor's own CLI, or a local model on Ollama, vLLM, or any OpenAI-compatible endpoint.

Think Claude Code, but: separate chat, code, and agent modes, an effort dial that selects *which model* and agent task width instead of just thinking tokens, and every call tracked locally so you learn which models actually earn their cost.

Go, two dependencies, a single static binary under 9MB, milliseconds to start.

curl -fsSL https://kolkrabbi.francomichetti.com/install.sh | sh   # macOS and Linux
kolk                                                              # opens a session

*Kolkrabbi* is Icelandic for octopus — *kol* ("coal") + *krabbi* ("crab"). Fitting: roughly two-thirds of an octopus's neurons live in its arms. Many arms, one small terminal, and many model providers within reach.

The three modes

/mode chat    plain conversation, no tools — cheap and instant
/mode code    the coding loop: read/write/edit files, run commands,
              iterate until done (Claude-Code style)
/mode agent   orchestration: plan the work as tasks with real
              dependencies, route each one to a model slot, run the
              independent ones concurrently, then synthesize one answer

Code is the default, so plain `kolk` is ready for file and command work. Switch to chat when you want a tool-free conversation, or agent when a longer task benefits from decomposition and isolated working contexts.

The effort dial

/effort low | medium | high | max

`ultrathink` scales thinking on one vendor's model. Kolkrabbi's effort scales across providers: each level maps t