Cursor rule

1---
2description: "MiniMax M3 + Cursor 3.7 tool discovery: runtime inventory, canvases, Design Mode inputs, schema-first MCP use, MCP Apps structured content, capability mapping, and safe fallbacks."
3alwaysApply: false
4---
5 
6# Tool Discovery (Cursor 3.7 + M3)
7 
8Current through **Cursor 3.7** (Jun 2026). Use this rule when the runtime surface is unfamiliar, tool-heavy, or likely to differ across environments.
9 
10## Discovery Order
11 
12Inventory the current surface in this order:
13 
141. direct tools exposed in the prompt (`Read`, `Grep`, `Task`, `Await`, etc. — names vary by surface; the live list wins)
152. browser, canvas, and other IDE-native surfaces exposed in the prompt
16   - **Browser MCP** (`cursor-ide-browser`): snapshot-first automation; read the server's tool descriptors
17   - **Canvas** (`.canvas.tsx`): confirm the session supports canvases before promising one; load the `canvas` skill before authoring
18   - **Design Mode** (3.7): user-driven in browser or canvas — not an agent tool; when the user has selected or annotated elements, treat that as the primary input for UI edits
193. **Cursor Marketplace** (or team marketplace) plugins already installed — same schema-first rules as other MCPs; discover tools and resources from each plugin's descriptors before calling
204. MCP tools and resources with their current schemas (including project-configured servers under the session's MCP descriptor path)
215. web docs only when discovery or versions still remain unclear
22 
23For delegation, read the live `Task` schema for available `subagent_type` values and whether nesting or `run_in_background` applies (Multitask Mode expects background runs).
24 
25## Schema-First Use
26 
27- Read the current schema or descriptor before calling unfamiliar MCP tools.
28- Match the task step to the smallest tool that can honestly do it.
29- Do not infer hidden parameters or old wrapper names from memory.
30- **MCP Apps structured content**: when a tool returns structured content, prefer the structured form over reconstructing from prose. When the structured data *is* the deliverable, route it to a **canvas** rather than a markdown table in chat.
31 
32## Capability Mapping
33 
34Before acting, translate the task into:
35 
36- what must be read
37- what must be changed
38- what must be verified
39- which currently exposed tool best serves each step
40 
41For visual-fidelity work, also translate into: which image/frame must be read, which **Design Mode** selections the user provided (if any), and which post-change frame must be re-read after editing.
42 
43For analytical or tabular deliverables, also translate into: whether a **canvas** is the right surface (audit, metrics breakdown, MCP query results) vs. inline chat prose.
44 
45## Discovery Loop
46 
47```text
481. Inventory current tools
492. Read the schema for unfamiliar options
503. Choose the smallest viable tool
514. Try the narrowest valid call
525. Verify the result
536. Escalate or fall back only if needed
54```
55 
56## Safe Fallbacks
57 
58- If a tool is unavailable, say so and choose the next best exposed path.
59- If discovery fails, simplify the task step and re-check the current surface.
60- Do not promise a tool-based deliverable until the path is confirmed.
61 
62## Anti-Patterns
63 
64- assuming a tool exists because it existed in another environment
65- using shell as the first choice when a direct tool is exposed
66- calling MCP tools without reading current schemas
67- hiding tool uncertainty behind confident prose
68- ignoring MCP Apps structured content and dumping it as prose or markdown tables when a canvas is available
69- promising canvas, Design Mode, browser, or nested `Task` behavior before confirming the current session exposes it
70