AI-CLI

Shell tasks that should take thirty seconds stall because you cannot recall the precise syntax — and the usual fix, pasting your command history into a web chatbot, ships your system details off-premise. ai-cli takes a plain-English request, sends it to a locally running LLM, and surfaces the generated shell command in an editable buffer. You inspect it, change it if the model missed something, then execute or cancel. The full cycle stays on your machine.

The architectural decision that sets it apart is radical minimalism: the entire implementation is a single C file with zero external dependencies. You compile it, move the binary to your PATH, and it is done — no Python environment, no package manager, no runtime to keep alive. The vendor states it works with any LLM server that speaks the `/v1/chat/completions` API, which covers Ollama, LM Studio, and compatible local inference servers. Backend swaps are a config change, not a reinstall.

The `–memory` flag is where the tool earns repeat use: it maintains session context across successive requests within a working directory, so each follow-up instruction builds on the last without you restating the file name or permission target. This works well for two- or three-step jobs on a single target. It does not handle conditional branching — if step two depends on what step one returned, you are bridging that manually. Teams with workflows that require automated multi-step decision trees will hit that ceiling and move to a scripted or agent-based approach.

The interactive review step is not optional and cannot be suppressed — the docs describe no flag to run in batch or headless mode. That is a feature for workstations and a blocker for unattended scripts or CI pipelines. The tool is MIT-licensed and self-hosted by definition; there is no cloud component and no paid tier.