In-App Agent Chat

Author from a chat inside the app

The in-app chat is the fastest way to work with an agent in Looky: nothing to install, nothing to configure beyond one thing — your own AI provider key. Open a workspace, open the chat, describe the change you want, and the agent drafts the Malloy and YAML for you. You review, validate and approve before anything goes live.

• Zero install. No Python, no CLI, no editor. If you can open your Looky instance in a browser, you have everything.
• Bring your key. The chat works with Claude (Anthropic), Gemini (Google) and Groq keys today, with more providers on the way. The key is pasted once, in the chat's settings.
• You approve what ships. The agent drafts and validates; publishing always requires your explicit yes — same rule as everywhere else in Looky.

How it works

1. Open the chat. With a workspace selected, use the chat button in the corner of the app.
2. Add your provider key. Pick your provider, paste your key, choose a model. The key is saved in your browser only.
3. Describe the change. "Add a monthly revenue trend to the sales dashboard." The agent reads your workspace and the official docs, then drafts the files in a private working copy inside your browser.
4. Review and validate. Browse the draft files right in the chat, ask for adjustments, and run validation — the same checks the CLI runs.
5. Approve the publish. When validation is clean and you say yes, the change goes live and the view refreshes in place.

If you are looking at a dashboard or visualization while you chat, the agent knows — "change this title" just works. A small chip above the message box shows what context rides along, and you can remove it with one click.

Where your work and your data live

• The draft lives in your browser. Until you publish, every edit exists only in a private sandbox in your browser — invisible to other users and to your instance.
• Your key stays yours. It is stored in your browser, travels only with each request you make, and is never kept by Looky.
• Your data never enters the conversation. The agent works with your content files and schemas — models, visualizations, dashboards — not with the rows in your database. Query results render for you in the browser, outside the chat.
• The agent reads the official docs. Its authoring contract is the same public documentation you are reading now, so it does not invent options that do not exist.

See what every turn costs

Because the chat runs on your key, it shows you what it spends — no console-diving on your provider's site to find out what a conversation cost.

• Under every reply, a muted footer states that turn's consumption, split the way your provider bills it: fresh input tokens, output tokens, cached reads, and cache writes — plus how many model calls the turn took.
• In the header, a running total keeps the whole conversation's tokens (and estimated cost) in view.
• Tokens are the measurement; dollars are an estimate. The ≈ $ figure is computed from your provider's listed prices at the time of the turn — your provider's invoice is always the authority.
• Failed turns show their spend too. If a turn errors out after some work, the tokens it consumed are stated rather than hidden.

The record lives in your browser alongside the conversation — reload the page and the history brings its costs back with it.

When a turn does not finish

The chat keeps going when it can, and asks you when it cannot:

• A brief provider hiccup is retried in front of you. You see the wait; you do not have to resend.
• If the connection drops, what already finished stays in the conversation. Continue picks up from there — completed tool work is not run again. The conversation is sent again, so a provider that can reuse a cached prefix will; one that cannot bills that prefix as fresh input.
• If the model stops mid-task — it ran out of steps, hit a length limit, or described a next step instead of doing it — the chat says so and offers Continue. That next reply is another full answer on your key, so it waits for you.

When to use which

The chat is one rung on a ladder — use as much or as little of it as you like:

• In-app chat — quick edits and new content without leaving the browser. Zero setup.
• Claude Code / Cursor workbenches — your own agent working on local files, with git and your editor around it. More control, more setup.
• By hand — the CLI and your editor, no agent at all. Everything the agents do, you can do yourself.