AI Chat

AI Chat is Revolte’s conversational interface for working with the AI agent directly. Describe a task in plain language, and the agent reads, writes, and runs code in a live sandbox—with as much or as little oversight as you want to give it.

​

Starting a thread

Before you send your first message, you configure where and how the agent should work:

• Application & repository — Scope the thread to one of your connected applications so the agent works against real repository context, complete with branch history and prior threads. If you don’t pick one, the agent starts in a blank local project you can link to a repository later.
• Base branch — Choose which branch the agent should start from; it creates its own working branch off of it, so your base branch is never touched directly.
• Sandbox size — Pick Small, Medium, or Large compute for the sandbox the agent runs in, based on how heavy the task’s build/test footprint is. See Sandboxes for details.

This means the same interface handles everything from “sketch out an idea with no repo attached yet” to “fix this bug on main, in a beefier sandbox because the test suite is slow.”

​

Thread details panel

A side panel keeps the thread’s context in view at all times, so you don’t have to leave the conversation to check what you’re working against:

• General — The project and application the thread is scoped to, with a link out to the App console.
• Repo details — The connected repo, base branch, and the working branch the agent is committing to.
• Sandbox Environment — Live status of the sandbox (e.g. Running), plus its instance type, vCPU, memory, and storage.

The same panel is also where you trigger Commit changes once you’re ready to push the agent’s work—see Committing your work.

​

Choosing a model and reasoning effort

The model picker allows you to switch between Claude, GPT, Grok, and DeepSeek models per thread, each with its own reasoning-effort tiers (e.g. low/medium/high) where the model supports it. That means you can match the model to the task: a fast, cheap model for routine changes, a higher-effort one for something that needs deeper reasoning.

​

Controlling how much oversight the agent gets

Every thread runs in one of two execution modes, which decide what happens when the agent wants to write or edit a file:

• Auto — File changes are approved automatically as the agent proposes them, so it runs uninterrupted end-to-end. Best for well-scoped tasks where you’re comfortable reviewing the result rather than every step.
• Manual — The agent pauses and batches up its proposed file writes/edits for your review before continuing. You see a diff (or full file preview for new files) for each pending change and can allow or deny them individually, or allow/reject the whole batch at once.

​

Prompt library

A built-in Prompt Library link sits right below the composer, giving you quick access to a set of reusable, pre-written prompts for common tasks—so you don’t have to write a good task description from scratch every time.

​

Committing your work

When the agent’s changes are ready to leave the sandbox, Revolte walks you through getting them into your repository. What that looks like depends on whether you started the thread with a repository connected: If you started the thread already scoped to a connected application (see Starting a thread), the first two steps are skipped entirely and committing goes straight to the commit dialog. This is on top of the automatic draft commit Revolte pushes if a sandbox is about to be torn down from inactivity—see Sandboxes: Inactivity timeout.

​

Usage and credits

AI Chat usage draws from your organization’s shared AI credit balance. If credits run out, active and new threads are blocked with a clear notice, and any company admin can top up credits without needing to involve engineering leadership for every request.

​

Next steps

• Sandboxes — Understand the execution environment behind every thread.
• Jira Workflow — See how the same agent runs autonomously from a Jira ticket instead of a chat prompt.