# Cursor

> Configure the bundled Cursor SDK with an API key and run sandboxed local-agent sessions.

Canonical URL: https://docs.zuse.sh/providers/cursor



Cursor runs through the SDK bundled with Zuse. It does not require a separate Cursor CLI or use a local Cursor login. Each session runs against the active checkout with the SDK sandbox and auto-review enabled.

## Add an API key [#add-an-api-key]

Open **Settings → Providers → Cursor**, enter a Cursor API key, and save it. The key is required and is stored in the macOS Keychain rather than `settings.json`.

Zuse performs a bounded readiness check against the SDK. Provider status distinguishes:

* **Ready** — the key was verified and the SDK returned a usable model.
* **Needs attention** — the key was saved but could not be verified, commonly because the service was unreachable.
* **Error** — the provider explicitly rejected the key.

An unverified key is not treated as invalid. You can recheck after connectivity recovers; the live session remains authoritative.

## Workflows in Zuse [#workflows-in-zuse]

* Start Cursor sessions in the main checkout or a per-chat Git worktree.
* Stream assistant text, thinking, tool calls, edits, usage, and errors into Zuse's durable timeline.
* Resume the local SDK agent from its stored agent identifier when the SDK still recognizes it.
* Use Plan mode for read-only exploration and switch back to agent mode for implementation.
* Interrupt an active run and continue in the same durable chat.
* Attach images supported by the selected model and reference files from the active checkout.
* Pass compatible user MCP servers from the project's native configuration into the SDK session.
* Use the normal files, terminal, Changes, commit, and pull-request surfaces around the session.

Model choices come from Zuse's Cursor model catalog and are resolved by the bundled SDK. A saved default does not guarantee that an account can use that model; provider errors remain authoritative.

## Sandbox and approvals [#sandbox-and-approvals]

Cursor sessions use the SDK's local sandbox with auto-review enabled. This differs from CLI-backed providers: an action rejected by Cursor auto-review is blocked instead of becoming a Zuse approval prompt. Plan mode still selects the SDK's read-only planning posture.

Review the resulting diff even when a run completes successfully. The sandbox and auto-review reduce risk but do not establish that an edit is correct or belongs in the final commit.

## Limits and recovery [#limits-and-recovery]

Zuse does not discover Cursor-native skills into the shared skills picker. Non-image attachments are retained in the Zuse message timeline but are not forwarded as Cursor SDK image inputs. Provider account limits, model access, and network availability are controlled by Cursor.

If a session fails:

1. Open **Settings → Providers → Cursor** and recheck the saved key.
2. Replace the key if Zuse reports an explicit authentication failure.
3. Verify the network connection if the key remains unverified.
4. Create or restart the session after changing the key or model.

If the stored local agent is stale or missing, Zuse creates a replacement agent and preserves the existing Zuse timeline. Send a concise context summary when the replacement process needs earlier conversational details.