No MCP servers
Claude Code speaks MCP, so it reaches Jira, Slack, Drive and anything else with a server. Kolkrabbi does not yet. If your workflow depends on MCP, this is the difference that matters most.
Compare
Kolkrabbi
vs Claude Code
Claude Code is Anthropic's agentic coding tool: it reads a codebase, edits files, runs commands, and runs in the terminal, in VS Code and JetBrains, in a desktop app, on the web and on a phone. It is proprietary and runs Claude models, across six deployment paths from the Anthropic API to Bedrock, Vertex and Microsoft Foundry. Kolkrabbi is an open-source terminal coding agent that is not tied to one vendor's models: it runs anything on OpenRouter, a local model on Ollama or vLLM, or the Claude and ChatGPT subscriptions you already pay for, by driving each vendor's own CLI.
Every claim below was checked against primary documentation on . Both projects move quickly. If something here is out of date, it is a bug — say so.
01 / WHERE KOLKRABBI IS BEHIND
What Claude Code does better.
This section comes first on purpose. A comparison that opens with its own strengths is an advertisement, and you should not trust the rest of it.
The terminal, and only the terminal
Claude Code runs in your IDE, a desktop app, a browser and a phone, and a session moves between them. Kolkrabbi is a terminal program; a remote device can watch a session and answer its permission prompts, but it cannot start one.
Both sandbox the bash tool; Claude Code's network rules are finer
Claude Code's sandboxed Bash tool uses Seatbelt on macOS and bubblewrap on Linux and WSL2, with filesystem and network isolation down to allowed domains. Kolkrabbi's sandbox is opt-in: /sandbox on uses Seatbelt on macOS and Landlock on Linux, confines writes to the project and temp, denies credential paths, and its network policy is allow or deny with nothing in between.
macOS and Linux only
Claude Code supports native Windows and WSL. Kolkrabbi cross-builds for Windows in CI and does not support it.
One person, one codebase
Claude Code is a funded product with a team, a release channel that auto-updates, and skills, hooks and subagents that have been through many more hands than Kolkrabbi's have.
02 / WHERE KOLKRABBI IS AHEAD
What Kolkrabbi does differently.
Not better at everything — different in specific ways, each of which you can check in the source or in the first ten minutes of using it.
Not one vendor's models
Claude Code runs Claude. Kolkrabbi runs whatever OpenRouter lists, a local model on Ollama or vLLM, or Claude and ChatGPT through their own CLIs — and lets you switch mid-session.
Both subscriptions, one terminal
If you pay for Claude and for ChatGPT, Kolkrabbi can run turns on either, and switching between a gateway model and a plan mid-session keeps the transcript and the project context. It never sees, stores or refreshes either credential: each vendor's own CLI does the login, and Kolkrabbi drives that CLI as a subprocess. One caveat, because it is the first thing you would hit: starting a session still expects a gateway key even when a plan answers the turns. That is on the list to remove, not a decision.
The dial picks the model
Elsewhere, effort means thinking tokens on one model. In Kolkrabbi, low, medium, high and max select which model runs the turn and how wide the agent work goes — so cheap work can go to a cheap model without you retyping anything.
Your own numbers, on your own disk
Every call is appended to a JSONL file you own. /stats shows cost, latency, tokens and your own 1–5 ratings per model. Claude Code shows usage on your account; Kolkrabbi shows it in a file you can grep.
No telemetry at all
Claude Code sends operational telemetry unless an organization turns it off. Kolkrabbi contacts a server when you ask it to, and not otherwise — no analytics, no background version check, no cloud sync.
03 / SIDE BY SIDE
The whole thing, in one table.
Where a cell says a fact was not established, it means exactly that: we did not find it in the documentation we read, and we would rather leave a gap than guess.
| Capability | Kolkrabbi | Claude Code |
|---|---|---|
| License | Apache-2.0 | Proprietary. Needs a Claude subscription or an Anthropic Console account. |
| Distribution | One static Go binary under 11 MiB, no runtime to install | Native installer, Homebrew, WinGet, apt/dnf/apk, or npm. Native installs update themselves in the background. |
| Model providers | OpenRouter, plus Anthropic, OpenAI, Ollama, vLLM and LiteLLM directly | Claude models only — but across six deployment paths: Anthropic API, Console, Amazon Bedrock, Claude Platform on AWS, Google Cloud’s Agent Platform, Microsoft Foundry. |
| Local models | Yes — Ollama, vLLM, or any OpenAI-compatible URL | No local model path. |
| A subscription you already pay for | Claude, through your own claude CLI, and ChatGPT, through your own Codex CLI |
Claude, its own. |
| Which model runs the turn | The effort dial picks it: low, medium, high, max, ultra map to models you choose | You pick the model. |
| Checkpoints | Before every write. /diff, /changes, /rewind, /undo |
Yes. /rewind or double-Esc, the 100 most recent per session, kept 30 days. |
| Changes made by shell commands | Not checkpointed | Not checkpointed either — their docs say so plainly. |
| Cost and quality data | A local JSONL you own: calls, tokens, cost, latency and your own 1–5 ratings | A usage dashboard on your account. |
| Telemetry | None. It contacts a server when you ask it to | Operational telemetry, which an organization can turn off through managed settings. |
| MCP | Not yet | Yes |
| Hooks and slash commands | Markdown slash commands; post-edit, post-write and session-end hooks | Hooks, skills and slash commands |
| Subagents | Dependency-ordered tasks sharing one working tree | Parallel subagents, coordinated by a lead agent |
| Execution sandbox | Opt-in. /sandbox on uses Seatbelt on macOS and Landlock on Linux; writes confined to the project and temp, credential paths denied, network allow or deny. Off by default |
Yes. Sandboxed Bash tool with Seatbelt on macOS and bubblewrap on Linux and WSL2; filesystem isolation and network isolation by allowed domain |
| Operating systems | macOS and Linux | macOS, Linux, WSL and native Windows |
| Where it runs | The terminal. A remote device can watch a session and answer its prompts | Terminal, VS Code, JetBrains, desktop app, web and mobile |
04 / THE TWIST
You can run Kolkrabbi on Claude Code.
The obvious question about this page is why you would use one when you already have the other. You do not have to choose.
Your own claude, driven as a subprocess
Kolkrabbi's Anthropic backend spawns the claude binary you installed and
logged into yourself, unmodified, in a cleared environment, and translates its output
into Kolkrabbi's own event stream. Your Claude subscription answers the turn.
In that mode, Claude Code is the agent
It runs its own tools. Kolkrabbi's permission rules and path jail do not gate them, and the interface says so on every provider-executed line rather than once at the start. Kolkrabbi is the front end and the recorder, not the agent.
The same ledger across both
Turns answered by your Claude plan land in the same local stats file as turns answered
by an OpenRouter model or a local one, so you can compare them. 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 none.
05 / TRY BOTH
Ten minutes settles more than a table.
Both are one command. Run them on the same repository, on the same task. When we did that with a local 8B model, all three scored zero — the pre-registered benchmark is where side-by-side results will land as they exist.
| Tool | Install |
|---|---|
| Kolkrabbi | curl -fsSL https://kolkrabbi.francomichetti.com/install.sh | sh |
| Claude Code | curl -fsSL https://claude.ai/install.sh | bash |
ALSO COMPARE
Kolkrabbi vs Codex CLI Kolkrabbi vs OpenCode All comparisons