[ docs · guides ]
Quickstart
Install Grist, sign in, attach your own provider key, and run your first task — five minutes, most of it waiting on npm.
Install
npm install -g grist-ai
That puts the grist CLI on your path. Prefer a GUI? The same harness ships as a desktop app — download the Mac desktop build and sign in there instead. Everything below works the same either way.
Sign in
grist auth login
Pick Grist and sign in — your account is created on the spot, no invite needed. Prefer a key? Paste a Grist API key (it starts with grist_sk_) at the same prompt, or create and paste one from the dashboard later.
Bring a provider key
Grist never sells inference. Routing and the harness are Grist; the model calls ride your provider key, billed to your provider account at zero markup. Attach a key from OpenRouter, Vercel AI Gateway, or any OpenAI-compatible endpoint — either during grist auth login or later in the dashboard under provider settings.
No provider key yet? The dashboard can issue a grist_sk_ key against your Grist account instead. Your own provider key is the better default: your spend, your limits, your data relationship.
Run your first task
Move into any git repository and start the interactive session:
cd your-repo
grist
Type a task in plain language — a bug fix, a small feature, a refactor — and press enter. For one-shot work, skip the interface and run headlessly:
grist run "add pagination to the orders endpoint, matching the users one"
Check the setup
grist doctor
grist doctor is the no-spend health check: it verifies your sign-in, gateway reachability, and which mode the routing gate is in, without buying any inference. If something is wrong, it says what — run it before anything else.
What happens on a run
- The gate routes. Your task is scored once, up front, and sent to the cheapest rung of the ladder that can handle it.
- Specialists work. The run executes in your working tree — exploring, editing, and verifying with tools, delegating pieces to specialist subagents when that pays.
- You get a diff and a number. Every run ends in a reviewable diff plus a cost summary, so you always know what changed and what it cost.
Your source never leaves the machine. Tools run locally; only the prompt goes out, and your provider key pays for the answer.