Help that does the work, not just points at it

Most help buttons hand you a manual. ToRun's In-App Assistant is a floating bubble on every screen that actually does the thing - fills the form in front of you, looks up the answer, and carries out multi-step actions on your behalf. It sees exactly what you see, asks before it changes anything, and lets you undo with one click.

Try it on any page · See which plans include it

It understands the page you're on

The assistant isn't a generic chatbot bolted to the corner. Before it answers, it takes a snapshot of your screen - the visible fields, the one you're focused on, the form's rules, and any errors showing - so its suggestions fit this task, not a guess about it.

Land on the pricing page and it can pre-fill your organisation details from context. Stuck on a validation error and it explains the fix in plain language. Drowning in a long form and it drafts every field for you to review. Because it reads the same context you do, the help is specific instead of generic - and you stay in control the whole time.

It can act, through real tools

When help means doing something, the assistant uses a catalog of tools - typed, well-defined actions it's allowed to call, each with a clear contract. A tool invocation isn't a hidden side effect; it's a named operation with inputs you can see and a result that's recorded. The assistant can chain several together to complete a multi-step task, and each step that changes your data is surfaced for approval before it runs.

Tools run inside a sandbox policy that decides what's reachable from a given surface, so the assistant's reach is bounded by design - it can do exactly what the page allows and nothing beyond it.

Nothing changes without your yes

Reading, searching, translating, and drafting suggestions happen instantly - there's nothing to approve when nothing is being changed. But the moment the assistant wants to do something that alters your data, it stops and shows you the exact proposal first as an approval request. You can edit it, accept it, or ignore it.

And approval never blocks you. The request waits quietly while you keep working; the action applies when you're ready. If something still lands wrong, every assistant action is reversible for 24 hours - one click rolls it back, cleanly.

Safe by design, on every screen

Each page decides what the assistant is allowed to touch. Form-fill is available where there are forms; it's switched off on sensitive surfaces like payment screens - no global kill-switch required. Every action is written to a tamper-evident audit trail, and anything sensitive you type (like a password) is redacted before it's ever logged. You get a helper that's genuinely useful and genuinely accountable.

Quietly handling support, too

The same assistant powers AI support deflection: ask a how-do-I question and it answers from the platform's own help corpus, grounded and cited - resolving most questions on the spot. When it genuinely can't help, it escalates into a real support ticket with the full conversation attached, so nothing falls through the cracks. Most questions get answered instantly; the rare hard one reaches a human with context already in hand.

What the assistant can do

Smart form fill bi-input-cursor-text - drafts every field from the context already on your screen, so long forms take a fraction of the time. You review and submit - it never hits "save" without you.
Context-aware answers bi-bullseye - it reads the page, the focused field, and any errors before responding, so the help fits the task in front of you.
Knowledge base lookup bi-search - ask a question and it searches your connected knowledge base, returning grounded answers with sources rather than guesses.
Acts through real tools bi-tools - a catalog of typed, sandbox-bounded actions it can call and chain to finish multi-step tasks, each one named and recorded.
Built-in translation bi-translate - rewrite or translate any text on the spot across the languages ToRun supports, no copy-paste into a separate tool.
Approval gates for changes bi-shield-check - anything that modifies your data shows up as a clear proposal you can edit or decline first. Async by design, so reviewing never stops your flow.
24-hour undo + audit trail bi-arrow-counterclockwise - every action is logged to a tamper-evident trail, sensitive input is redacted, and you can reverse anything the assistant did for a full day after the fact.