about

The model should work out what you want. Not do the work.

That one sentence is the whole design. A small local model reads your request and emits a JSON plan. From there, ordinary code validates the plan, expands it into concrete steps, asks before it writes, and runs it. The model never runs a command, never writes a shell string, and never owns a safety decision.

the name

knaif is a portmanteau of knife and AI — which is why the wordmark brackets the AI. Think of a Swiss Army knife: a compact tool where every blade does one thing well and folds cleanly away. Each skill is a blade. You carry the ones you need; the handle stays the same.

The brackets do double duty. In the plan pipeline, the model is the only non-deterministic step — genuinely fenced in, with tested code on either side:

why not just use an agent

Most agents use the model as a general-purpose brain. Every request goes through it, trivial or not — "trim this video to 10 seconds" gets the same multi-thousand-token treatment as "refactor my auth layer". That buys three problems:

knaif inverts it. The model is a thin intent layer; the capability lives in hand-written, tested code. The model picks compress_video and extractstarget_size_mb: 25. Code turns that into a probe → recipe → preview → confirm → run workflow and renders the exact command. Same input, same output, every time.

what that costs to run

Every cloud agent pays for inference per use, forever — real electricity, at data-centre scale, for work that is mostly deterministic. knaif pays the AI cost once, during development: you spend it building and tuning a skill, and ship something whose runtime cost is one short call to a 4B model on the user's own machine.

Our measured comparison puts a number on it: roughly $0.058–0.14 every single time someone asks, against $0 and a tenth of the latency, for work all four systems completed equally well. One request is pocket change. A habit is not, and neither is a million people with the same habit.

Write once. Run free, run local, run the same way every time.

what it is not

• Not a general coding agent. It will not refactor your repo.
• Not a shell the model drives. It cannot execute or even emit commands — only a plan that code checks first.
• Not a cloud service. No account, no key, no upload.
• Not magic for tiny models. Below about 4B, accuracy drops, and our own evaluation says so plainly.
• Only as capable as its skills. That is the trade, and it is deliberate.

The narrow scope is the point. knaif wins by being a sharp tool for a few deterministic jobs, not a dull one for everything.