Run Cursor Cloud Agents on Cloudflare through self-hosted machines ↗︎. Each Cursor session assigned to the deployment runs in its own Linux sandbox: a Container that one Durable Object starts for that session.
Cursor hosts the agent loop, inference, and planning. Cloudflare runs commands, file edits, repository operations, and other tools inside infrastructure that you control.
You need:
- A Cursor Enterprise plan with self-hosted machines turned on
- A Cursor team service-account API key with agent scope
- A Cloudflare Workers Paid account with access to Containers and R2
- Node.js 20 ↗︎ or later
- A running Docker ↗︎ daemon for deployment and local development
To create a team pool, set CURSOR_API_KEY in your shell. Then, start a local worker with the Cursor Agent CLI:
CURSOR_API_KEY="$CURSOR_API_KEY" agent worker --pool cloudflare-test startThe command registers the pool and temporarily connects your local machine as a worker. After the pool appears in Cursor, stop the worker with Ctrl+C and run unset CURSOR_API_KEY. Record the pool name for CURSOR_POOL. Keep the local worker stopped while testing the Cloudflare deployment so it does not claim the agent request.
For more information, refer to Cursor team pools ↗︎.
The template deploys a Worker, a Durable Object namespace, a container application, an R2 bucket binding, and a cron trigger.
-
Clone the template and install its dependencies:
git clone https://github.com/anysphere/cloudflare-workers.git cd cloudflare-workers npm install -
Log in to your Cloudflare account:
npx wrangler login -
Create the R2 bucket for optional repository snapshots:
npx wrangler r2 bucket create cursor-pool-worker-snapshotsTo use another bucket name, update
bucket_nameinwrangler.jsonc. -
Store the required Cursor service-account key as a Worker secret:
npx wrangler secret put CURSOR_API_KEYEnter a team service-account key with agent scope. Personal API keys do not work with pool workers.
-
To access private repositories, store your Git credentials as Worker secrets:
npx wrangler secret put GIT_USERNAME npx wrangler secret put GIT_TOKENFor GitHub, set
GIT_USERNAMEtox-access-token. SetGIT_TOKENto a token with access to the repositories that the agents use. -
In
wrangler.jsonc, setvars.CURSOR_POOLto the Cursor team pool name:{ "vars": { "CURSOR_POOL": "default" } } -
Set
containers[].max_instancesto the maximum number of concurrent sessions that the deployment must support. -
Deploy the Worker and container:
npx wrangler deployWrangler builds the image and deploys the Worker, Durable Object, container application, and cron trigger.
A new container image rollout stops running containers. Wait for active Cursor sessions to finish before you deploy an update.
Repository-bound agents route work by Git remote and by team pool name.
-
Go to Cursor Cloud Agents ↗︎.
-
Start an agent and select a repository.
-
Select Self-hosted, then select the name configured in
CURSOR_POOL. -
Wait for Cursor to assign the session to the deployment. The Worker then starts a container for the session. The initial scheduled controller run can take up to five minutes to begin.
The request provides the repository URL. The container restores or clones that repository into $HOME/workspaces/repo-0. It then starts the Cursor worker:
agent worker --worker-dir "$HOME/workspaces/repo-0" --pool "$CURSOR_POOL" start --verboseThe Cursor worker derives its repository label from the Git remote. Do not configure repo= labels manually.
Any-repository agents route work by team pool name. They start with an empty working directory and no Git remote.
-
Go to Cursor Cloud Agents ↗︎.
-
Start an agent and select the Any repo group.
-
Select the team pool name configured in
CURSOR_POOL.
The container creates $HOME/workspaces/repo-0 without a Git remote. The agent or a project hook can clone a repository during the session.
Repository snapshots are an optional cache for repository-bound agents. A snapshot stores the post-clone working tree in R2. An any-repository agent does not use this cache.
-
Store a secret that protects the snapshot routes:
npx wrangler secret put SNAPSHOT_AUTH_TOKEN -
In
wrangler.jsonc, setvars.WORKER_PUBLIC_URLto the deployed Worker URL:{ "vars": { "CURSOR_POOL": "default", "WORKER_PUBLIC_URL": "https://cursor-pool-workers.<ACCOUNT_SUBDOMAIN>.workers.dev" } } -
Deploy the updated configuration:
npx wrangler deploy
A cache miss performs a normal Git clone. It does not prevent the agent from starting.
The template manages one container for each assigned Cursor session.
Your Worker connects to Cursor on a cron trigger and claims sessions. It sends each session to its own Durable Object, which starts and stops a container. The Cursor worker in the container connects to Cursor and runs the session. Every connection to Cursor starts on Cloudflare, so nothing connects in.
A cron trigger runs the Worker every five minutes. Each run lists the sessions that wait for CURSOR_POOL. The run then keeps the server-sent events stream from Cursor open until shortly before the next run, so the Worker also receives sessions that arrive between runs.
The Worker claims each session with a unique worker ID. Cursor then assigns the session only to that Worker. A Durable Object for the session starts one container with the session environment and repository information.
The container stops after five minutes without work, or after WORKER_IDLE_RELEASE_TIMEOUT_SECONDS if you set that variable. The Durable Object also stops the container after eight hours, or after MAX_RUN_LIFETIME_SECONDS if you set that variable.
The container connects out to Cursor, so it needs no inbound port or public IP address. The Worker exposes only its health route and the optional snapshot routes.
Each session runs in its own container and does not share files or processes with other sessions. The container receives the CURSOR_API_KEY service-account key. It also receives GIT_USERNAME, GIT_TOKEN, and SNAPSHOT_AUTH_TOKEN if you set them. Code that runs in a session can read these values. Give GIT_TOKEN access only to the repositories that the agents use.
For more information, refer to Sandbox security.
To stream controller and container logs, run:
npx wrangler tailTo list container instances, run:
npx wrangler containers listTo test one scheduled controller run during local development, start wrangler with scheduled-event testing:
npx wrangler dev --test-scheduledIn another terminal, invoke the scheduled route:
curl "http://localhost:8787/cdn-cgi/local/scheduled?cron=*/5+*+*+*+*"If you change the cron interval, update both triggers.crons in wrangler.jsonc and CONTROLLER_RUN_BUDGET_MS in src/config.ts.
| Symptom | Cause | Resolution |
|---|---|---|
| No sessions are assigned | The cron does not run, the key is missing, or the team pool name does not match | Run npx wrangler tail. Check controller runs, 401 responses, and the configured team pool name. |
The controller returns 401 |
The key is personal or lacks agent scope | Replace CURSOR_API_KEY with a team service-account key that has agent scope. |
| The team pool is absent for a repo | The worker started without repository labels | Select Any repo, or start a repository-bound agent with a configured Git remote. |
| The session is assigned but does not start | The container cannot start, clone the repository, or authenticate | Run npx wrangler containers list and inspect npx wrangler tail. Check capacity and Git secrets. |
The container exits with Error: Container exited with unexpected exit code: 1 and an earlier log reports cursor-agent CLI not found on PATH |
Cloudflare WARP or another TLS-inspecting proxy may have prevented Docker from downloading the Cursor CLI. An unguarded shell pipeline can hide the installation failure and produce an incomplete image. | Run npx wrangler tail and inspect the preceding container logs. If the Cursor CLI is missing, disconnect WARP, clear the Docker build cache, and run npx wrangler deploy again. Or configure Docker to trust the root certificate of your organization. |
| The first repository start is slow | The snapshot cache is empty or not configured | Configure both WORKER_PUBLIC_URL and SNAPSHOT_AUTH_TOKEN, or allow a cold Git clone. |