Describe the request
Feature request for product/service
Cursor IDE / Agents Window / CLI
Describe the request
Please expose a supported way to specify the model when programmatically creating a new desktop Agents Window (glass) session.
Today, automation can open a new local Agents Window session via:
cursor /path/to/workspace
When an Agents Window is already open, this correctly routes to a new agent:
vscode:createNewComposer { folderUri }
→ glass.newAgent({ folderUri, source: "cli" })
But the IPC payload only includes folderUri. There is no model field, and cursor --help exposes no --model for this GUI open path.
cursor --model some-model /path/to/workspace
# Warning: 'model' is not in the list of known options, but still passed to Electron/Chromium.
Why this matters
Teams automate multi-session workflows from scripts (design → review → implement loops, parallel agents on different models, CI-like local orchestration).
Right now model selection is inconsistent across surfaces:
| Path | Can specify model? |
|---|---|
| Terminal agent: cursor agent --model | Yes |
| Cloud Agents API: POST /v1/agents with model.id / model.params | Yes |
| Desktop Agents Window via cursor | No |
So local GUI automation cannot pin a model per task, even though Cloud and terminal agent already can.
The capability already exists internally
Agents Window already has a working internal path for “new agent with model”:
Command Palette section: “New Agent with Model”
Command id: newAgentWithModel / glass.newAgentWithModel
Implementation creates a new agent draft, then sets the model on that draft via setModelConfigForComposer(… selectedModels: [{ modelId }] …)
So this is less “build a new feature from scratch” and more expose an existing internal command on a stable external API.
Requested API (any one would unblock us)
Option A — preferred: CLI flag on folder open
cursor /path/to/workspace --model claude-opus-4-7
Pass { folderUri, modelId } (or model) through:
vscode:createNewComposer → glass.newAgent
Option B — stable command / deeplink invocation
Document and support invoking the existing command externally, e.g. a stable way to run glass.newAgentWithModel with a model id, or a cursor:// / deeplink equivalent.
Option C — one-shot create with model + prompt
For scripted desktop spawning, ideally support both model and initial prompt in one call, so automation does not depend on brittle UI paste timing.
Current workarounds (and their gaps)
Cloud Agents API with model — works, but not a local Agents Window session.
cursor agent --model — works, but terminal agent ≠ GUI Agents Window.
Manually set the global default model, then spawn — not per-session; concurrent multi-model jobs stomp each other.
UI automation against the model picker — fragile; not suitable as a supported integration path.
Environment
macOS arm64
Cursor 3.12.30 (also observed on 3.11.x glass routing)
Agents Window (glass) already open
Public cursor CLI shim
Acceptance criteria
A documented, stable way to create a new Agents Window session with an explicit model id.
Model is applied before the first user message is sent.
Works for local/scripted automation without clicking the model picker.
Ideally does not force changing the user’s global default model (per-session override).
Thanks! This would bring desktop Agents Window automation to parity with Cloud API and terminal cursor agent --model.
Operating System (if it applies)
MacOS