How AI Agents Work Inside a CoCalc Project

Updated

AI agents in CoCalc run inside the same persistent Linux project that holds your files, software, and running services, rather than in a separate sandbox. That means an agent can open, edit, and save the actual project files your collaborators already use, and you can review or take over its work afterward. CoCalc supports an integrated Codex agent, Claude Code (an experimental preview on sites that enable it), and other command-line agents run from project terminals.

The core idea: agents work where the project lives

A CoCalc project is a durable shared Linux environment. It keeps notebooks, terminals, code, documents, services, discussion, history, and recovery in one place. Agents are not given a copy of your work in an isolated container — they operate on the project itself.

This has three practical consequences:

  • Direct file access. The agent reads and writes the same files you and your collaborators see. There is no export/import step between "agent workspace" and "real project."
  • Access to tools and services. The software installed in the project and any running services are available to the agent, because it runs in that environment.
  • Shared state. Because the project is persistent, the agent's changes remain in place and are subject to the project's history and recovery features.

How the agent options differ

Option How it runs Notes
Integrated Codex agent Built into CoCalc Use it directly within the project
Claude Code Experimental preview Only on sites that enable it; works with your personal Claude Pro or Max subscription
Other command-line agents Run in project terminals Any CLI agent you can launch in a terminal

The choice is mostly about which agent you prefer, not about where the work happens — all three operate against the same project.

What collaboration looks like

Collaborators see edits live. When an agent saves changes, others in the project can open what it saved, or take over the work themselves. This is the main difference from a standalone agent sandbox: there is no handoff step, because the agent was already working in the shared environment.

A typical flow looks like this:

  1. An agent (Codex, Claude Code, or a terminal-based agent) makes changes in the project.
  2. Collaborators watch the edits appear live.
  3. Anyone can open the saved result and continue the work manually.

How this compares with an agent sandbox

An isolated agent sandbox gives the agent its own environment, separate from your real project. That can be useful for containment, but it means the agent's output has to be moved into your project before it's usable, and the agent can't touch your installed tools or running services.

In CoCalc, the agent works in the project's actual environment. The trade-off is the reverse: you get direct access to real files and services, and you rely on the project's history and recovery features to manage changes rather than on the sandbox boundary.

Where to go next

  • To try it, start on CoCalc.ai and open a project.
  • For agent-specific workflows, see the agent workflows page.
  • For the sandbox comparison, see the compare-with-agent-sandboxes page.
  • For deployment choices — hosted CoCalc.ai, local or single-VM, or a customer-operated private deployment — review the product paths.
  • For cost and licensing details, see Pricing and licensing; for help, see support and sales.

CoCalc is developed by SageMath, Inc., and has offered collaborative computing online since 2013.