Compare
Codex CLI is OpenAI's terminal coding agent: Apache-2.0, installed
by script, npm, Homebrew or a release binary, and signed into with the ChatGPT plan you
already pay for. It reaches local and third-party models through
model_providers in config.toml, speaks MCP, and sandboxes the
commands it runs. Kolkrabbi is also Apache-2.0 and also reaches local
models — the differences are narrower than the marketing on either side suggests, and
worth stating precisely.
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 If something here is out of date, it is a bug — say so.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
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.
Codex has sandbox_mode — read-only, workspace-write, or danger-full-access — with a matching approval policy, always on. Kolkrabbi's sandbox is one mode and opt-in: /sandbox on confines the bash tool with Seatbelt on macOS or Landlock on Linux, writes limited to the project and temp, credential paths denied, network deny optional. The gap is now the default and the number of modes, not the capability.
Codex's MCP support is deep: OAuth and bearer auth, per-tool approval modes, timeouts. Kolkrabbi has none yet.
Codex runs on Windows through PowerShell. Kolkrabbi cross-builds for Windows in CI and does not support it.
Codex is OpenAI's own client for OpenAI's models, and it will always know things about them first.
02 / WHERE KOLKRABBI IS AHEAD
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.
Codex reaches other providers through configuration, but it is OpenAI's client and defaults to OpenAI. Kolkrabbi treats every provider as a first-class route, including two subscriptions and a gateway with the whole catalogue behind it.
Codex uses your ChatGPT plan. Kolkrabbi can run turns on your ChatGPT plan and on your Claude plan, by driving each vendor's own CLI as a subprocess, and a gateway model and a plan can trade places mid-session. It never sees or stores either credential. 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.
Low, medium, high and max select which model runs the turn and how wide the agent work goes, rather than only how much thinking one model does.
Every call lands in a JSONL file on your disk. /stats turns it into cost per model, latency, token counts and your own 1–5 ratings — the data you need to decide which model is worth its price.
No runtime, no package manager, two dependencies in the module graph, and a release whose checksums are verified by the installer and signed with keyless Cosign.
03 / SIDE BY SIDE
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 | Codex CLI |
|---|---|---|
| License | Apache-2.0 | Apache-2.0 |
| Distribution | One static Go binary under 11 MiB | Install script, npm, Homebrew cask, or release binaries |
| Model providers | OpenRouter, plus Anthropic, OpenAI, Ollama, vLLM and LiteLLM directly | OpenAI, plus any OpenAI-compatible provider through model_providers.<id>.base_url |
| Local models | Yes — Ollama, vLLM, or any OpenAI-compatible URL | Yes — ollama and lmstudio are reserved built-in provider ids |
| A subscription you already pay for | Claude, through your own claude CLI, and ChatGPT, through your own Codex CLI |
ChatGPT Plus, Pro, Business, Edu or Enterprise, by signing in |
| Which model runs the turn | The effort dial picks it: low, medium, high, max map to models you choose | You pick the model. |
| Checkpoints | Before every write. /diff, /changes, /rewind, /undo |
Not documented in the configuration reference we checked |
| Cost and quality data | A local JSONL you own: calls, tokens, cost, latency and your own 1–5 ratings | Not documented in the configuration reference we checked |
| Telemetry | None | Not established here — check their current docs |
| MCP | Not yet | Yes, and thoroughly: mcp_servers.<id> with OAuth or bearer auth, per-tool approval and timeouts |
| 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 deny optional. Off by default |
Yes. sandbox_mode is read-only, workspace-write or danger-full-access, with approval_policy of untrusted, on-request or never |
| Operating systems | macOS and Linux | macOS, Linux and Windows through PowerShell |
04 / THE HONEST PART
A lot of comparison pages claim that only their tool can talk to Ollama. That is not true here, and pretending otherwise would make everything else on this page less believable.
ollama and lmstudio are built inThey are reserved provider ids in Codex's configuration, and any other
OpenAI-compatible endpoint can be added with base_url and
wire_api. vLLM works.
Kolkrabbi uses the Ollama you already have and never installs one. A pulled model
appears in /model even while Ollama is idle, and choosing it starts the
server for the session. vLLM and LiteLLM connect over an OpenAI-compatible URL.
The gap is not whether a local model is reachable. It is whether the tool is built to move between a local model, a gateway model and a subscription in one session, and to record what each one cost and how good it was. That is the part Kolkrabbi is for.
05 / TRY BOTH
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 |
| Codex CLI | curl -fsSL https://chatgpt.com/codex/install.sh | sh |
ALSO COMPARE
Kolkrabbi vs Claude Code Kolkrabbi vs OpenCode All comparisons