Hey guys, I’d like to present my ongoing project I co-authored (with the help of codex-cli ). This extremely useful command line tool that provides a natural language way to describe what you want to happen at the terminal.
AI has brought a lot of budding new developers into the space and it is extremely useful for CLI beginners, making the terminal much more accessible for all.
It is also great for experienced users - very handy for formulating complex piped commands, or ones you rarely use like that specific but unusual thing you want to do with git but can’t quite remember …
It is bash based and works on WSL Ubuntu and macOS zsh amongst many other platforms. Naturally it leverages the OpenAI API, but will work with others too. It is pretty economical to run as token counts are generally low.
It incorporates a traffic light risk level indication by colouring the proposed command appropriately. “Dangerous” commands, e.g. things which will change things permanently like deletes, prompt for additional confirmation.
Adds a configurable “risk appetite” setting to allow CLAI to automatically execute lower-risk model-suggested commands without prompting, whilst preserving confirmation for higher-risk commands.
This makes usage much faster for more experienced users.
Changes:
Introduces risk_appetite config (default 0) with setup-wizard support and normalization.
Adds execution gating logic to skip confirmation prompts based on command risk + risk_appetite.
Updates README documentation and expands smoke tests to cover auto-run vs prompt behavior.
Adding reasoning support means you can run it much more reliably with Google’s Gemma 4 model locally (via Ollama) so you can now avoid any privacy issues and pay zero cost to run the helper (apart from some negligible extra electricity)
You can now enable history and command results in history without worrying about details about your local computer being shared with the cloud.
On Windows you will probably need a 16GB VRAM GPU to pull this off, or at least 24GB on a Mac, preferably 32.
Cost $37 to create the PR with Codex, so there was something in it for OpenAI even if it is likely I will probably run this fully locally going forward
CLAI has migrated from Bash to Go while retaining the familiar interface:
> clai check whether this server is running out of resources
The migration gives us a more maintainable, testable foundation with safer concurrency, typed provider integrations, reliable HTTP/JSON handling, native release binaries, and no runtime dependency on curl or jq.
New capabilities include:
Extensible MCP tool support, allowing independently maintained tools to be discovered without changing CLAI’s core code.
Lazy tool startup, so MCP servers run only when relevant and tools are enabled.
Optional command-result evaluation: after running a command, CLAI can make a follow-up LLM call to interpret its output in the context of the original request. This is
particularly useful for diagnostics, resource analysis, logs, and system-health checks.
Support for OpenAI, Anthropic, Gemini, Ollama, and compatible APIs.
Verified cross-platform release binaries and a streamlined bootstrap installer.
Continued use of existing configuration, history, command syntax, and safety controls.
CLAI remains simple at the command line, but now has a much stronger foundation for future development.
Well the internet is blowing up over AI again, but this time the excitement is not about OpenAI for once.
It’s about the new “System One” style models.
What’s a System One style model?
A System One-style model makes fast, focused judgments, such as classifying intent or assessing risk, rather than generating lengthy answers. It returns structured decisions and confidence scores that software can act on.
The key advantages of these models are:
Speed: Fast judgments without generating long responses.
Lower cost: Small, focused decisions can be cheaper than full LLM calls.
Predictable outputs: Structured choices and scores are easier to integrate into code.
Explicit uncertainty: Confidence scores help decide when to automate or ask a person.
Focused accuracy: Specialized models can outperform general-purpose models on particular decision tasks.
Actual gains depend on the model and workload; confidence isn’t a guarantee of correctness.
CLAI 2.1: Meet your terminal’s new second opinion
Ask for what you want in plain English. CLAI turns it into a shell command, explains what it does, and helps you decide whether to run it.
With CLAI 2.1, that decision gets a dedicated second opinion: optional System One integration, including TypeSafe’s Jev.
Your main LLM still writes the commands and explanations. System One handles focused judgments:
Understand your intent: Is this a command request, a question, or a request to clear conversation history?
Audit command risk: Independently assess the proposed command, escalating its risk rating when needed.
Keep auto-run in check: Apply your configured risk appetite, while requiring confirmation when the assessment is uncertain.
Why split the work? Generating a useful command and judging whether it is safe are different tasks. System One models are designed for fast, focused decisions with structured answers and confidence scores. That gives CLAI a practical way to combine flexible generation with explicit execution rules.
More automation when confidence supports it. More caution when it doesn’t.
The integration is optional: add your System One key, API endpoint, and model to enable it. Existing configurations continue to work.
Ready to try it? Get CLAI 2.1.0, with binaries for Linux, macOS, and Windows.
). This extremely useful command line tool that provides a natural language way to describe what you want to happen at the terminal.
by colouring the proposed command appropriately. “Dangerous” commands, e.g. things which will change things permanently like deletes, prompt for additional confirmation.
on github if you find it useful!
