How the AI Coding Agent in OneCompiler Works

Updated

The AI coding agent in OneCompiler runs inside the browser-based editor, so there is nothing to install. You describe a change in plain language, and the agent reads your workspace, edits files in place, creates new ones, and cleans up after itself step by step. Every change lands behind a checkpoint you can roll back, and each modified line is marked directly in the editor.

What the agent does

According to OneCompiler's own description, the agent works across your whole workspace rather than a single file. For one instruction it can:

  • Read existing files to understand the current state
  • Edit files in place
  • Create new files
  • Delete files it no longer needs
  • Clean up after itself as it finishes

A multi-step turn means the agent can carry out several of these actions to complete a single instruction, not just one edit.

A concrete example

The homepage shows a worked example. The instruction given was:

Add a dark mode toggle to the header and remember the choice.

The agent's visible steps were:

  1. Checked the header markup and the existing styles.
  2. Read styles.css.
  3. Edited index.html.
  4. Edited styles.css.
  5. Created script.js.

The result, as described: the header has a toggle button, styles.css switches colours on a dark class, and script.js saves the choice in localStorage so it persists on refresh. The run summary read 4 steps · Edited index.html, styles.css · Created script.js, and the file change summary was 3 files changed +41 −6.

This shows the pattern: one natural-language instruction, multiple coordinated file operations, and a working result you can run immediately.

Checkpoints and rollback

Every turn is a checkpoint. That means after the agent finishes a turn you can:

  • Restore the workspace to before that turn
  • Re-apply a change you previously restored
  • Reject or Apply the pending changes

The example UI shows Reject and Apply controls alongside Restore, plus a Restore option in the run summary. Because edits stream straight into the editor with every changed line marked, you can review exactly what changed before deciding.

How changes appear in the editor

Changes are not applied invisibly. They stream directly into the editor, and each changed line is marked. You watch the agent work step by step rather than waiting for a finished result. The run summary also reports token usage and time (the example shows 2.8k tokens · 5.4s), so you can gauge the cost of a turn.

What you need to use it

  • A browser — the agent runs in the editor, with nothing to install.
  • An open editor workspace with the files you want the agent to work on.
  • A plain-language instruction, or you can type / for commands.

The homepage states the editor is free to use across 110+ languages, and the agent is described as running in your browser. Pricing details beyond that are listed on a separate pricing page, so check there if you need current plan information.

When this approach fits

The agent is a good match when you want to make a scoped change to an existing small project — adding a feature, wiring up a script, adjusting styles — and you want to review each step before keeping it. The checkpoint model means a bad turn is recoverable rather than permanent.

It is less suited to work that depends on local tooling, private dependencies, or a full development environment, since everything runs in the browser editor. For those cases, the agent's workspace is limited to the files in that editor.