Lexis: a local model writes the command, plain code decides if it runs

Lexis turns plain English typed in a terminal into one shell command, written by a small model on your own computer. It's for developers who know what they want but not the flags for find, tar or lsof. The model never decides what runs: a command executes only if code proves it read-only or you approve it.

The model proposes, code decides

Three times, the 1.5b model answered a harmless question with a drastic command. For "disk usage of /" it proposed remounting / as noexec. So Lexis treats the model as a suggestion box, and ordinary code decides what happens next.

A shell hook catches any line that isn't a real command. The local model (llama.cpp, llama-server) returns one command, and the policy engine in lexis/src/policy.mjs assigns it a risk level. A command proven read-only runs at once. Anything else shows an approval card: Enter runs it, e edits it, Esc cancels. A critical command makes you type yes.

The model can't lower a risk floor, and no flag skips the critical confirmation. sudo, stopping a service and removing a package raise the floor. A remote script piped to a shell, or an encoded command, always asks. rm -rf on your home directory, the root or a wildcard is critical, and the card shows the blast radius: how many items and how many bytes.

The hook leaves real commands alone. A line with a real command, flags or paths goes straight to the shell, so a missing gh still prints "command not found". Piped and agent-spawned shells never reach Lexis.

757 of 919, and how I counted

These numbers are mine, from the README and lexis/model-cards/METHODOLOGY.md. I froze the test cases before training and picked the models on a separate selection half. The final half, 919 cases, ran once per model. A case passes only if the first command is right. The grader accepts known-good forms and sometimes runs the command in a container, and the methodology file says it understates accuracy slightly.

Model	Final half, 919 cases	Warm p50 latency	Size
lexis-1.5b	757 (82.4%)	about 0.25 s	986 MB
lexis-0.5b	714 (77.7%)	about 0.1 s	398 MB
Qwen2.5-Coder-0.5B-Instruct, untrained	188 (20.5%)	not measured	not measured

Latency is from one machine, an Apple M4 Pro, with the Q4_K_M files. Seed noise on the selection score is about 15 cases.

Both models are a full fine-tune of Qwen2.5-Coder, one epoch, on about 140,000 rows. Every row passed checks on its target OS: valid syntax, a command that exists, and every flag present in the man page or --help. A sample also ran in a Docker container as an unprivileged user, and rows overlapping the test sets were dropped. In one targeted round, 3,591 rows went in and 2,711 survived the checks. A coding agent wrote 62,502 of the rows, which passed the same checks.

What didn't help

Model size gave the biggest gain: 5 points for 1.5b over 0.5b on the same data. Most of the rest did nothing.

• More epochs and a higher learning rate overfit.
• Two rounds of DPO raised pair accuracy from 66% to 77%. The selection score went from 836 to 835.
• Targeted data gained 5 and 9 cases on the final half, inside the noise. The 1.5b model ships without it because it lost ground on Windows.

Where Lexis still gets it wrong

I read 217 failures of the 1.5b model. About 40 were unknown tools. About 30 were invented commands or flags, like rm -sf and dnf ininstall. About 25 missed a detail, about 12 wrote when asked to look, and about 8 had broken syntax.

Web search is on by default in auto mode and sends queries to DuckDuckGo. Turn it off with lx config set webSearch.enabled false. The apple-fm provider is experimental and weaker. The training data isn't published, so you can't rebuild the models from the repository.

Lexis against cloud shell assistants

GitHub Copilot CLI and Warp AI are cloud assistants, and a larger hosted model may handle more requests than a 1.5b one. I ran no head-to-head test, so I claim no accuracy advantage. What Lexis offers is documented: the model runs on your device, it needs no account or API key, it has no telemetry, and the policy engine is separate from the model.

Try Lexis

• macOS or Linux: curl -fsSL https://lexis.hridya.tech/install.sh | bash
• Windows: irm https://lexis.hridya.tech/win.ps1 | iex
• npm, with Node 20 or later: npm install -g @hridyacodes/lexis, then lexis setup

Then run lx plan "biggest 10 files in my Downloads" or lx explain 'rm -rf ~'. The demo on the site matches fixed example requests to fixed commands, and nothing runs on your computer.

Source: github.com/hridaya423/lexis (Apache-2.0). Models: lexis-1.5b and lexis-0.5b. Full method: METHODOLOGY.md.