The AI assistant

The Assistant is the chat front door to Upliftr's AI debugging agents. Describe a bug or a flow in plain English and the agents read your suites, runs, issues, and plans, reproduce the problem, root-cause the failure, and write or edit the YAML for the checks they save. They are read-only by design: they investigate and diagnose, but never mutate your code, data, or infrastructure. It's the fastest way to go from "something's broken" or "I want to check this" to a root cause and a re-runnable check, with no YAML to hand-author.

Open it from Assistant in the sidebar of the dashboard.

What it can do

The agents work entirely over your own data, with read-only access. They never make things up. They can:

• Generate suites from a plain-English description, or by crawling a live URL.
• Edit suites, "add a case for an expired coupon", "make the login test also check the avatar loads", reading the suite first and changing only what's needed.
• Debug failures, "why did my last run fail?", it reads the run and root-causes the failure to the exact failing case, step, or error.
• Read across your project, suites, recent runs, filed issues, and plans, to answer questions with real context.
• Run suites, when you ask, and after confirming (runs spend your model budget).

The agents are scoped to debugging and checking your app. Ask them something unrelated and they'll point you back to debugging or building checks.

A few things to try

Generate a suite for the login flow at https://app.example.com. I will give you a test account.

Add a case to checkout.upliftr.yaml that rejects an empty cart.

Why did run_… fail?

Summarize my open issues by severity.

How it works

• You review before it saves. When the assistant writes a new suite (or substantially rewrites one), it shows you the proposed YAML first so you can tweak or discuss it. It saves only once you're happy, or if you tell it to just save.
• It gathers what it needs. For a flow behind a login, it asks for a test account (username, password, and the login path if it's not /login) and logs in before crawling, so it can reach pages behind auth instead of producing a thin suite. Credentials go straight to the browser, never to the model.
• It confirms before anything destructive or costly. Deleting a suite or starting a run always asks first.
• It's grounded. Every statement about your project comes from a tool result in the conversation, if it hasn't read something, it says so and offers to fetch it.

Bring your own key

The agents run on your Anthropic API key (set it in Settings → AI), self-hosted in your own environment. Each turn spends tokens billed to your own account, there's no per-seat AI markup. If the key is missing or out of credit, the assistant will tell you instead of failing silently.

Prefer a form?

The Assistant is the front door, but the structured paths are still there: Generate (from a URL, your docs, or a spec) and the YAML editor on the Suites page. Use whichever fits, they all produce the same suites.