Cursor rule

1---
2description: "Cursor 3.7 MCP optimization: browser Design Mode, canvases, Figma, Cloudflare tools, MCP Apps structured content, and direct action patterns."
3alwaysApply: false
4---
5 
6# Cursor 3.7 MCP Optimization
7 
8Current through **Cursor 3.7** (Jun 2026). Direct guidance for Cursor's MCP and integrated surfaces. Use the right tool for each job.
9 
10## Browser Tools (cursor-ide-browser)
11 
12**Always do before any interaction:**
13```
141. browser_tabs with action: "list" -> check current tabs/URLs
152. browser_snapshot -> get current page structure and refs
16```
17 
18**Interaction pattern:**
19```
201. Identify the target ref from snapshot (use exact ref, not coordinates)
212. For clicks: prefer snapshot refs over mouse_xy coordinates
223. For forms: use browser_fill for existing content, browser_type to append
234. After ANY action that changes page: take fresh snapshot
24```
25 
26**Coordinate clicks ONLY when:**
27- A fresh screenshot confirms the exact visual target
28- No snapshot ref is available
29- The target is visually verifiable (button, link)
30 
31**Never do:**
32- Click without a fresh snapshot first
33- Reuse old screenshot coordinates
34- Assume page state without verification
35 
36## Browser Design Mode (Cursor 3.7)
37 
38Design Mode is a **user-driven** overlay in the integrated browser — agents do not invoke it as a tool. When the user selects elements there, treat those selections as ground truth for UI edits (same discipline as `multimodal-grounded`):
39 
40- **Multi-select:** two or more elements selected together carry code, layout, and visual relationships — use them for parity edits, deduplication, or batch component tweaks.
41- **Voice input:** transcribed voice instructions are user intent; honor them even if the agent is mid-run.
42- **Agent path unchanged:** still snapshot → ref-based interaction for automation; Design Mode complements, not replaces, the MCP browser flow.
43 
44## Canvas (Cursor 3.7)
45 
46Canvases are live React artifacts beside the chat — dashboards, reports, analyses, internal tools. Load the `canvas` skill before creating or editing `.canvas.tsx` files.
47 
48**When to use a canvas:**
49- Structured MCP or tabular deliverables where the data *is* the output (prefer canvas over markdown tables)
50- Interactive reports the user may share or iterate on (context usage breakdowns, audits, timelines)
51 
52**3.7 canvas features:**
53- **Design Mode in canvases** — user can select and annotate UI elements directly; treat annotations like browser Design Mode selections.
54- **Context usage report** — interactive token breakdown (prompt, tools, rules, skills); useful for compression decisions on M3 long-context work.
55- **Embedded prompt buttons** — canvases can run follow-up prompts; keep button text bounded and explicit.
56- **Full-screen share** — shared canvases open in the browser for presentation; not a substitute for app UI verification.
57 
58## Figma Tools (plugin-figma-figma)
59 
60**Design-to-code workflow:**
61```
621. Call get_design_context with fileKey and nodeId
632. Adapt output to project's existing stack/components
643. Do NOT use raw Figma output as final code
65```
66 
67**URL parsing:**
68- `figma.com/design/:fileKey/:fileName?node-id=:nodeId` -> convert "-" to ":" in nodeId
69- `figma.com/board/:fileKey/:fileName` -> FigJam, use get_figjam
70- `figma.com/make/:makeFileKey/:makeFileName` -> use makeFileKey
71 
72**When designer provides screenshots:**
73- Use get_screenshot for visual verification
74- Translate Figma design tokens to project's token system
75- Never copy exact Figma hex values without checking project palette
76- For M3 multimodal parity, the design screenshot is the contract — cite the file path in the report
77 
78## Cloudflare Tools
79 
80**Load these skills before Cloudflare work:**
81- `cloudflare/SKILL.md` — comprehensive platform guidance
82- `wrangler/SKILL.md` — CLI usage and best practices
83- `workers-best-practices/SKILL.md` — production Worker patterns
84 
85**Before deploying:**
86```
871. Read wrangler skill for correct syntax
882. Verify secrets/bindings in wrangler.jsonc
893. Run wrangler dev locally first
90```
91 
92**MCP auth:** If a Cloudflare MCP server has `mcp_auth`, call it before using tools.
93 
94## MCP Apps Structured Content
95 
96Since Cursor 3, **MCP Apps structured content** lets tools return richer structured outputs (typed objects, lists, tables) instead of prose blobs.
97 
98- When a tool returns structured content, **prefer the structured form** over reconstructing from prose.
99- Surface the structured result faithfully in the report — do not paraphrase the shape.
100- A few plugin families (Vercel, Linear, Cloudflare) already return structured outputs of this kind; check the descriptor before assuming a tool returns plain text.
101- If the runtime supports it, render tabular or nested results as a canvas or table, not as a markdown dump.
102 
103## MCP General Rules
104 
105```
1061. Read the tool schema BEFORE calling unfamiliar MCP tools
1072. Use exact parameter names from the schema
1083. If tool unavailable, say so and use next best exposed path
1094. Do not infer hidden parameters or old API versions
1105. If the tool returns structured content, use the structured form rather than the prose form
111```
112 
113## Marketplace plugins
114 
115Plugins installed from the [Cursor Marketplace](https://cursor.com/marketplace) extend agents with MCPs, skills, and related capabilities. Treat them like any other MCP surface: **inventory** what the session exposes, **read descriptors/schemas** before first use, then call the smallest valid tool — same as in `tool-discovery.mdc`.
116 
117**Rollout caveat (Cursor 3.7):** third-party plugin imports now **default to off for Enterprises** when unset, while preserving explicit Admin overrides. If you are working in an Enterprise workspace, confirm the plugin import policy with the Admin before relying on a marketplace plugin.
118 
119## Capability Mapping Checklist
120 
121Before acting, confirm:
122- [ ] Direct native tool exists for this action
123- [ ] MCP tool exists with current schema (and structured-content support if relevant)
124- [ ] Browser interaction confirmed via snapshot
125- [ ] Fallback path identified if primary fails
126