Agent
Overview
What the Velork agent can do, where it lives in the app, and what it needs before you start.
The Velork agent works on your project from chat. It reads and edits files, runs terminal commands, reads and edits the design canvas, drives the built-in browser, writes plans, and splits large jobs across subagents. You describe the outcome; the agent picks the tools.
Before you start
Chat opens once you are signed in to Velork and have AI available. That can be a connected subscription (Claude Code, Codex, Cursor, or Copilot), your own API key, or Velork's hosted models while your free credits last. See AI providers.
Open a project folder so the agent has files to work on. Without one, type @ in the composer to pick a recent project.
Where the agent lives
Velork has two layouts. Press Cmd + Shift + F to switch between them, or use the Agents / IDE button in the title bar.
- Agents puts chat front and center, with your sessions as tabs or a sidebar and an agents panel beside it for the browser, files, terminal, changes, canvas, spend, and cloud computer.
- IDE is the full editor, with chat in the side bar.
Press Cmd + L to jump to the chat input. In the Agents layout it works from anywhere except a focused browser tab, where it selects the address bar. In the IDE layout it works while focus is in the chat side bar.
What the agent can use
| Area | What it does |
|---|---|
| Files and search | Reads, searches, writes, edits, renames, and deletes files. Edits show as diffs you can review. |
| Terminal | Runs shell commands, including long-running servers in the background, and reads their output. |
| Design canvas | Reads pages, frames, and layers, edits the canvas, exports assets, and works with canvas comments. |
| Browser | Opens pages in the built-in browser, clicks, types, takes screenshots, and reads the console and network. |
| Computer use | Controls your Mac's screen, or a cloud computer that Velork runs for you. |
| Web | Searches the web, fetches pages, and captures live sites onto the canvas as inspiration. |
| Planning | Writes plans, keeps a todo list, asks you questions, and builds Slates for data you need to look at. |
| Subagents | Starts separate agents for independent parts of a job and runs them in parallel. |
| Git and tasks | Commits, pushes, opens pull requests, and creates or runs tasks on the task board. |
| MCP | Calls tools from MCP servers you connect. See MCP servers. |
| Deploy | Publishes your project to a link on velork.dev that anyone can open. |
There is no fixed cap on tool calls. A run continues until the job is done, the agent needs you, or you stop it.
Context the agent gets
Velork passes the agent your open project, the current mode, and anything you attach: files, folders, photos, canvas layers and pages, skills, and elements you pick in the browser or on the canvas.
The agent also reads instructions from:
- Settings → Instructions, added to every new chat
- An
AGENTS.md,CLAUDE.md, orCONTEXT.mdfile at the root of each workspace folder (the first one found, in that order) ~/.config/velork/AGENTS.md,~/.velork/AGENTS.md, or~/.claude/CLAUDE.mdin your home folder, for every project
Share a live link
When your project has something to open, such as a package.json, an index.html, or a Dockerfile at its root, a Share button appears in the title bar of the Agents layout. You can also run Share a Live Link (Deploy) from the command palette, or ask the agent to deploy.
- Deploy publishes the project to a link on velork.dev that anyone can open. With a
velork.deploy.jsonin the project, Velork builds it as that file says; without one, the agent works out how to build and run it. - Update publishes the project as it is now. The link stays the same.
- Change link picks the name in
[name].velork.dev. The old link redirects for 30 days. Choosing a name depends on your plan. - Turn off takes the link down. If a deploy fails, Fix with AI hands the error to the agent.
A deployed project sleeps when nobody uses it and wakes on the next visit.
Stay in control
Each chat runs in a mode, such as Develop for code or Plan for a written plan first. See Modes.
Shell commands ask for approval unless you allow them. Choose how much the agent may run without asking from the + menu in the composer, or set rules in Settings → Permissions. See Permissions.
Each of your messages has a Restore Checkpoint button that rolls files back to before that message. See Chat.