Run a Cursor Cloud Agent's commands and file edits in a temporary Render Sandbox. Cursor manages the agent in its cloud; Render provides a separate Linux workspace for each task.
This starter asks the agent to fix a JavaScript program that totals three orders as 30 instead of 42. You retrieve its changes, check them against the original tests in a fresh sandbox, and save the corrected program and report. Choose Python or TypeScript for the local helpers; both run the same JavaScript task.
This starter is a preview. A complete Cursor task and unattended controller recovery still need validation.
Run the example · Prerequisites · Source files
Cursor sends tasks to a named group of machines called a Team Pool. A program on your machine watches the pool and reserves tasks; this is the controller. For each task it starts a worker, the program that executes Cursor's commands, inside a new Render sandbox.
flowchart TB
Setup["1. Your machine<br/>Prepare the task repository<br/>and saved worker image"]
Start["2. Your machine + Cursor cloud<br/>Submit the task;<br/>the controller reserves it"]
Work["3. Render worker sandbox<br/>Run commands and file edits<br/>requested by Cursor"]
Save["4. GitHub<br/>Store the agent's<br/>changed program and report"]
Verify["5. Your machine + fresh sandbox<br/>Download files and<br/>run the original tests"]
Clean["6. Your machine<br/>Delete the agent<br/>and stop its sandboxes"]
Setup --> Start
Start -->|Controller starts worker| Work
Work --> Save
Save --> Verify
Verify --> Clean
Two terminals stay open on your machine: one for the controller, one for the task helpers. Worker sandboxes can reach Cursor and GitHub. The separate verification sandbox has outbound networking blocked. The walkthrough explains setup, expected output, and cleanup in order.
- A Cursor Enterprise team with Team Pools enabled. An admin must enable self-hosted machines, GitHub token minting, and private-worker session tokens.
- The Cursor GitHub App connected to the team and allowed to access the task repository. The submitting user needs access to that repository too.
- A Render workspace with Sandboxes enabled, currently in early access, and its terminal program, the Render command-line interface (CLI), version 2.28.0 or later.
- Python 3.10 or later for both paths, because the shared controller uses it. The TypeScript path also needs Node.js 22 or later to run its helpers, and npm to install their packages.
- Bash or zsh, Git, and the GitHub CLI, signed in with
gh auth login. The walkthrough installs the Cursor CLI.
Prepare these two credentials from the same team. They let programs use Cursor's application programming interface (API):
| Credential and creation page | Purpose | Where it stays |
|---|---|---|
| Service account key | Register the pool, reserve tasks, and obtain worker tokens | First terminal, as CURSOR_API_KEY; CURSOR_RELEASE_API_KEY uses the same value for failed-start cleanup |
| User API key | Submit the task, read its status, and delete the agent | Second terminal, as CURSOR_USER_API_KEY |
The controller needs the service account key; agent login does not supply it. The sandbox receives only a token for its reserved task, never either API key. Render uses render login and your selected workspace for this local walkthrough. A controller hosted on a server instead needs a Render API key.
In a terminal, download the starter and enter the repository. If you use zsh, first run setopt interactivecomments so pasted lines beginning with # are treated as explanations:
# Download the example and its helper programs from GitHub.
git clone https://github.com/render-examples/cursor-cloud-agents-sandboxes.git
# Enter the downloaded folder before opening the walkthrough.
cd cursor-cloud-agents-sandboxesFollow the step-by-step walkthrough. It creates a private task repository, prepares a saved worker image, and runs the controller. Both language paths cover submission, independent verification, and cleanup.
A successful result must pass all four original tests and produce:
{"total":42,"row_count":3}The helpers save results/total.mjs and results/report.json inside your chosen cursor/python or cursor/typescript directory. Keep these downloaded files before cleanup deletes the agent and stops its sandboxes. The cleanup steps also explain how to remove the saved image and optional test repository.
| Files | Purpose |
|---|---|
| Snapshot builder and spawn script | Save the installed worker tools and start one sandbox per reserved task |
| Python helpers and TypeScript helpers | Submit the task, retrieve and verify its output, and clean up |
| Sample program, input, and tests | The deliberately broken program and the original checks |
| Prompt, task note, and verifier | Instructions for the agent and a separate check of its report |
For failures, start with troubleshooting. For a controller that stays available after your laptop disconnects, see running the controller on Render, including the recovery work still needed.