Every AI coding tool makes a quiet trade on your behalf. You paste your code somewhere, and in exchange you get help. Usually that somewhere is a vendor's server, wrapped in an editor you didn't choose, tied to a model you can't swap, billed on a subscription you can't audit.
Vexi was built to refuse that trade. It is a CLI that runs on your machine, talks directly to whichever AI provider you already pay for, and keeps everything else local.
There is no Vexi server
This is the part people assume is marketing language, so it's worth being precise. There is no backend. No account system, no telemetry endpoint, no analytics beacon, no "anonymous usage statistics."
When you run Vexi, exactly one network destination is involved: the API of the provider whose key you supplied. Your code goes to Groq, or Anthropic, or your own Ollama instance, under that provider's privacy policy, which you already accepted. It never passes through us, because there is no us to pass through.
Everything else stays on disk:
- Your API key lives in
~/.vexi/config.jsonwith owner-only permissions. - Your project memory lives in
.vexi/inside the project itself. - Your session recordings are local files you can delete, diff, or gitignore.
You can verify all of this. The source is MIT-licensed on GitHub, and the network layer is small enough to read in an afternoon.
Bring the key you already have
The second trade most tools make is locking you to one model. That made sense when there was one good model. It makes much less sense now, when the frontier moves every few weeks and the price difference between providers can be two orders of magnitude.
Vexi supports 13 providers and detects which one you're using from the shape of the key itself. Paste a key starting with sk-ant- and you're on Claude. Paste gsk_ and you're on Groq. There is no provider dropdown to configure, because the key already told us.
npm install -g vexi-cli
vexiIf your endpoint isn't in the list (a local Ollama server, LM Studio, a corporate proxy), vexi setup takes a raw OpenAI-compatible URL, fetches the available models, and verifies the connection with a real inference request before it saves anything. You find out your config is wrong immediately, not on your first prompt.
The terminal is not a limitation
Choosing the terminal wasn't nostalgia. It's where the rest of your toolchain already lives. Your compiler, your test runner, your package manager, your git remote. Vexi can drive all of them because it's standing in the same place you are.
That's what makes something like /push possible: stage the changes, draft a commit message from the actual diff, push. No context switch, no second window, no copying a suggestion out of a chat panel and hoping it still applies.
What this costs you
Honesty matters more than pitch here, so: a local-first tool gives up some things. There is no shared team workspace. There is no cloud history you can open from another laptop. There is no hosted index of your codebase that makes retrieval instant.
What you get instead is a tool with no account to lose access to, no pricing page that can change, no server that can go down, and no third party accumulating your source code. For a lot of developers, especially anyone working under a contract that says code doesn't leave approved infrastructure, that's not a preference. It's the requirement.
Vexi is MIT licensed and free. If it stops being useful to you, the uninstall is one command and it takes nothing with it.