Cursor 3.7 Lets Developers Point at UI and See Every Token Spent
Cursor 3.7 adds point-and-click canvas editing and a live token breakdown report so you can see and fix context bloat in real time
Takeaways
- Design Mode in Canvas: You can now point at and annotate UI elements in agent-generated canvases to guide edits, no text description needed.
- Context Explorer: A new interactive canvas report shows exactly where tokens are going -- system prompt, tools, rules, skills -- so you can diagnose and fix bloat.
- Debug with Agent: A one-click button in the context report opens a fresh conversation to identify and reduce context waste.
- Canvas sharing upgrades: Shared canvases now open full-screen in the browser; agents can embed clickable prompt buttons inside canvases.
- Design Mode also got multi-select and voice queuing in the browser, letting you edit multiple elements at once and queue voice instructions while agents are still working.
- All features are live in Cursor 3.7 with no additional cost beyond existing plans.
Cursor's agents have been able to build interactive dashboards and internal tools inside the IDE for a while now, using a feature called Canvas. But until now, iterating on those artifacts meant typing descriptions into a chat box and hoping the agent understood which element you meant. The 3.7 release changes that in two meaningful ways: you can now point directly at UI elements to edit them, and you can finally see exactly where all your tokens are going.
Point, don't describe
Design Mode is now available in canvases. You can select and annotate UI elements directly in a canvas to guide Cursor's edits, just as you would in the browser. Instead of describing the change in text, you can point to it, provide feedback, and iterate more quickly.
This is the same interaction model that already existed in Cursor's integrated browser for live web apps, now extended to agent-generated canvas artifacts. Design Mode is Cursor's visual editing layer: instead of describing a UI change in text, you point at a live element in the running app and the agent receives its code identity and a screenshot as context. The key detail is what gets sent under the hood -- the agent receives the element's code identity and a screenshot for spatial context, including layout, surrounding elements, and the exact page state.
The 3.7 update also adds multi-select to Design Mode in the browser. Selecting multiple elements together gives the agent their code, surrounding layout, and visual relationships on the page simultaneously, so it can make one match the other, remove repeated content, or adjust a group of components at once. You can also queue edits by voice while an agent is still working. The microphone remains available while an agent is still working, so you can queue the next change by voice before the current one completes.
The token bill, finally itemized
The second major addition is arguably more impactful for teams running complex agents. Context usage -- the total amount of information passed to the model in a single request -- has become a real operational problem. Cursor appears to automatically inject local agent and tool "skills" into the active context window, even when they are unrelated to the current task. This causes context window usage to grow extremely quickly, despite only a few conversation turns and no heavy code analysis. One user on the Cursor forum reported hitting 77% context usage (roughly 230K out of 300K tokens) after minimal work.
The cost implications are real. When the agent works autonomously -- reading files, running terminal commands, iterating -- each background step counts as a separate request. A single agent task can trigger 5 to 15 model calls internally, and at around $0.04 per call, that adds up. And it's not just cost: if your rules take up 25% of the context window, the AI has 25% less space to look at your actual source code. Research shows that when prompts get too long, AI models suffer from "needle-in-a-haystack" issues -- they remember the top and bottom instructions but ignore the ones in the middle.
Cursor can now show your agent's context usage as an interactive report in a canvas. The context explorer breaks down where tokens go across the system prompt, tool definitions, rules, skills, and more. Because it's a canvas, you can ask the agent follow-up questions, and it can customize the report to answer your specific questions. This is the key design decision: the report itself is a live canvas, not a static readout. You can interrogate it conversationally.
Clicking the "Debug with Agent" button embedded in the canvas asks Cursor to identify opportunities to reduce context usage in a new conversation. Starting a new conversation matters here -- it avoids the very token bloat you're trying to diagnose from contaminating the debugging session itself.
What else shipped
Beyond the two headline features, the release includes four smaller canvas improvements:
- Shared canvases can now be opened full-screen in the browser, making them easier to present to others
- Agents can now embed buttons in canvases that run a specific prompt when clicked
- Improved ability for agents to fix canvas type errors automatically
- Better component styling and more chart customization options
The bigger picture
Canvas doesn't exist in isolation. It sits alongside a series of Cursor releases in early 2026 -- including Design Mode, upgraded voice input, Composer 2, and new multi-repo agent capabilities -- that collectively signal a coherent vision: remove friction in human-agent collaboration and make it easier to express intent beyond plain text.
The context explorer is a particularly telling addition. Cursor has been investing heavily in dynamic context management -- developing a "dynamic context discovery" approach to optimize how agents use context windows and token budgets when working on long-running software development tasks. Instead of loading all potentially relevant information upfront, the system enables agents to dynamically pull only the necessary context as needed. The new report gives users visibility into that system for the first time, turning context management from a black box into something you can actively tune.
AI agents are becoming persistent collaborators, not just query-and-response tools. And the interface through which that collaboration happens -- how agents show their work, how developers review and direct it -- is becoming as important as the underlying model quality. The 3.7 canvas update is a direct bet on that thesis: the agent's output is more useful when you can point at it, reshape it, and understand what it cost to produce.
Both features are available now in Cursor 3.7 as part of existing plans. The context explorer can be triggered from the agent's context usage indicator, and Design Mode in canvases works the same way as it does in the integrated browser. The official Design Mode blog post has a detailed walkthrough of the multi-select and voice queuing mechanics.