An AI agent
for your terminal.
He works in your shell: reads files, writes code, runs it and fixes what breaks. Before he asks a model, he checks his library of things he has already solved. You pay per task.
npx virgil-cliNo download page, no account, no credit card. One command and Virgil is in your terminal.
npx virgil-cli— one command, needs Node 18+- 15 free credits — no card, no account, no signup
- Tell it what to do. It acts, you watch — and you see what every task costs.
1 credit is 1 cent.
No subscription, ever. You start with 15 free credits — no card, no account, no signup. After every task Virgil shows exactly what it cost and what's left. You never pay for a model sitting idle, and you're never throttled for spending too little.
A real day with Virgil. Reading is free, repeats come straight from your library, and you see every cent — no plan, no surprise bill.
The library
Most CLIs send every request to a model, so the tenth fix costs as much as the first. Virgil keeps four local libraries: speech, debug, code and UI/HTML. Every request is checked against them first. A known answer costs 0 calls, a known pattern 1 call, and only something new goes through the full chain. The result is stored, so the next similar task is cheaper.
virgil brain. The key stays in your system keychain. Two days free, then 5 € a month for the library updates.What if your model is 90% trained?
The more you run Virgil, the more its library fills with your patterns — your fixes, your scripts, the shapes of your work. They come back answered locally: no model call, in seconds, for next to nothing. Push it far enough and most of your day runs on your own machine. That's the whole point — your Virgil becomes yours. Keeping the library fed with fresh patterns, once you're a regular, is a flat 5 €/month — run as much as you want.
Building a weather tool
A real run. Virgil had to find the weather API himself, hit a KeyError, read his own code, fixed it and checked two cities.
Same idea, different billing
npx virgil-cli — instant, no accountCommon questions
A note on the v0.1.4 security flag
Virgil is built in the open and ships fast. An early version (0.1.4) let Virgil share small code snippets with its shared library, so every user's copy learns from patterns someone already solved — and an automated scan (OSV MAL-2026-17630) flagged that data leaving the machine before we had documented it. That was a fair call: anything that leaves your computer should be spelled out first. So we spelled it out. The excerpts that feed the shared library are now capped and disclosed line by line, and you can stop contributing them any time with virgil learn off (or VIRGIL_NO_LEARN=1). To actually solve a task, Virgil still sends what that task needs to its reasoning model — like any AI CLI — and if you want our servers out of the loop entirely, bring your own key with virgil brain, so requests go straight to your provider. Your files, keys and shell never leave on their own — no background uploads, no hidden telemetry, no watermarks. The full list is in the README under “What leaves your machine”.
Is it really pay-per-use?
Yes. No subscription. You buy credits (1 credit = 1 cent) and spend them only when Virgil does work. A short question costs a fraction; a real multi-step task, a few. Reading and listing files are free.
What does a typical task cost?
A quick question is around 0.03 credits. A hard, self-correcting task — write a script, run it, fix errors, verify — is roughly 3 credits, about three cents. The exact number is shown when it finishes.
Which operating systems?
macOS, Linux and Windows. Virgil detects your OS and uses the right commands. You need Node 18 or newer; then npx virgil-cli or npm i -g virgil-cli.
Is my code sent anywhere?
Virgil runs on your machine and only sends what a task needs to its reasoning model, like any AI CLI. Nothing runs in the background, and he asks before changing or running anything.
How is this different from Claude Code?
Same idea, an agent in your terminal. The differences: pay per use instead of a subscription, no account to start, the cost of every task shown, and the library that makes repeated work cheaper.
Can Virgil run commands without asking me?
No — by default every shell command, file write, and file edit shows you a preview and waits for y (once) or a (always for this kind of step). There is a --auto flag that skips confirmations entirely; it exists for scripted pipelines where you want zero interruptions. Don't use it in an unfamiliar directory or on a production machine — you'd be running an AI agent with no gate. Reading and listing files never require confirmation.
Light a lantern in your terminal.
npx virgil-cli