Cursor cloud runs

The Cursor Agent provider supports two run modes:

Mode	Where it runs
Local	In-process via Cursor SDK on your machine. Standard agent run.
Cloud	On Cursor's cloud VMs. Survives Cerevisor restarts, laptop sleep, network drops.

License gate: Trial+/Paid only.

When to use cloud mode

• Long-running workflows (multiple hours).
• Runs where Cerevisor might be closed before completion.
• Runs where laptop sleep would normally interrupt.

For most workflows, local mode is simpler, Cerevisor stays connected and you watch live progress. Cloud mode trades some immediacy for resilience.

Enabling cloud mode

Per workflow: Settings → Workflow → Run mode → Cloud. The mode is only valid when the active provider is Cursor Agent.

Optional knobs:

• Repository targets: which Cursor cloud repos the run should attach to.
• Auto-create PR: automatically open a pull request for any code changes the agents made. Destructive by default: Cerevisor confirms once per workflow before enabling and again the first time a run fires with this set.

What happens when you click Run

1. Cerevisor calls Agent.create({ cloud: { repos, autoCreatePR } }) via the Cursor SDK.
2. Cursor spins up a VM and starts the agent.
3. Cerevisor registers a CloudRunRecord to the cloud-run tracker (electron-store backed, persisted to disk before iterating the stream): so if Cerevisor crashes mid-run, the record survives.
4. Cerevisor streams events back to update the Execution Modal in real time.

You can close the Execution Modal: the run continues on Cursor's VM. The status bar shows a "Cloud run active" indicator.

Surviving Cerevisor restart

If you quit Cerevisor (or it crashes) mid-cloud-run:

1. The cloud VM keeps running.
2. The local record stays on disk under cerevisor-cloud-runs.
3. On next launch, bootstrapCloudResume() scans the records. Records that have been silent >24h are flagged as failed.
4. The Cloud Sessions modal shows the resumable runs.

To reconnect:

• Click the Cloud Sessions icon in the title bar (only visible if any cloud runs exist).
• Find the run you want to reconnect to.
• Click Reconnect.

Cerevisor calls Agent.resume(cursorAgentId) to rehydrate the SDK handles, fetches the latest state, and shows you the Execution Modal again.

If the cloud-side run finished while you were away, the modal shows the final state and offers to fetch artifacts.

Artifacts

After every successful cloud run, Cerevisor downloads any artifacts the run produced to ~/.cerevisor/cloud-artifacts/<sessionPrefix>/<agentPrefix>/.

You can manually re-fetch via the Cloud Sessions modal → Refetch artifacts button.

Artifact paths are sanitized, remote paths with slashes don't recreate as nested local directories.

Cancellation

To cancel an active cloud run:

• Cloud Sessions modal → find the run → Cancel.
• Cerevisor calls Cursor's cancel API; the cloud VM shuts down.

Dismissing finished runs

Completed runs stay in the Cloud Sessions modal until you dismiss them. Click Dismiss on a completed run to remove it from the list (artifacts stay on disk).

What doesn't work in cloud mode (v1)

• File intelligence: mtime tracking doesn't span across the local-to-cloud boundary cleanly.
• Cerevisor's MCP client: Cursor owns its own mcpServers config; cloud-side connectivity is your deploy concern.
• Per-tool permission modal: sandbox/approval is up-front, same as local Cursor and Codex.
• Loop & Enhance: each round spins up a fresh cloud agent. Works, but the inter-round handoff is rougher than local.

Cost

Cursor bills cloud runs against your Cursor subscription. Cerevisor reports $0.00 in run cost (subscription-billed, no per-token visibility).

When NOT to use cloud mode

• Short workflows (< 5 minutes): overhead of VM spin-up isn't worth it.
• Workflows that touch local files extensively: file intelligence isn't reliable across the boundary.
• Workflows where you want live per-tool permission prompts; those don't work in cloud mode.

For local-first workflows that just need to survive Cerevisor closing, local Cursor mode + the cloud-run tracker's resume logic handles most cases gracefully.

Why Cursor and not other providers

Cursor is the only provider Cerevisor supports with first-class cloud execution that the SDK exposes. Anthropic, OpenAI-compatible, and Codex are all local-execution from Cerevisor's perspective; they call hosted APIs, but the agent loop runs on your machine. Cursor's cloud mode runs the entire agent on Cursor's infrastructure, including its own tool calls and sandboxing.