mcp-discovery

MCP Discovery

When to invoke

BEFORE picking a tool for current-docs research / browser automation / Tauri desktop testing / design extraction / web crawling. Don't hardcode mcp__context7__* in agent procedures — invoke this skill to find the best available tool for the task.

Expert Operating Standard

Follow docs/references/skill-expert-operating-standard.md: start from source of truth, preserve retrieval evidence, apply scope safety, use real producers with runtime receipts for durable delegated outputs, verify before completion claims, and keep confidence below gate when evidence is partial.

Step 0 — Read source of truth

1. Run node <resolved-supervibe-plugin-root>/scripts/discover-mcps.mjs (or rely on SessionStart having done it)
2. Read .supervibe/memory/mcp-registry.json

Decision tree — task → MCP preference

Need current docs / library API?
  Preference order: context7 > ref > WebFetch
Need browser automation / screenshots?
  Preference order: playwright > skip
Need Tauri desktop app testing, native webview state, IPC, windows, logs, or devices?
  Preference order: tauri > playwright (frontend preview only) > skip
Need to extract design from Figma?
  Preference order: figma → if absent, ask user for screenshots
Need to crawl a website?
  Preference order: firecrawl > playwright (one-page) > WebFetch
Need general web search?
  Preference order: firecrawl-search > WebSearch

Capability matrix

Task	Preferred MCP	Fallback	Must report
Current library docs	context7	official docs via web	library id and version if known
Browser interaction	playwright	static scrape or manual instructions	URL, viewport, screenshot path when used
Tauri desktop testing	tauri	Playwright only for browser-hosted frontend preview	session target, available webview/ipc/window/log/device tool groups
Figma/design extraction	figma	screenshot or exported assets	file key, node id, downloaded assets
Web crawl	firecrawl	targeted scrape/search	URL set and extraction format
Search/news	firecrawl-search	web search	recency and source links
Local document parse	firecrawl-parse	direct file read if plain text	path and parser choice
Calendar/mail/drive	installed app connector	ask user to enable/install	connector availability

No-prompt path

If discovery was already run by session startup, do not interrupt the user. Read the registry and choose the best available tool. Ask only when:

• the task requires a connector that is not installed,
• the user must provide a URL, file, or Figma node,
• the fallback would materially reduce quality,
• using a tool would mutate external state.

Output evidence

Always include:

• selected MCP or none,
• fallback reason if no MCP was used,
• exact tool family used,
• for Tauri, available tool groups (webview, ipc, window, logs, devices),
• whether the result is current or cached,
• any limitation that affects confidence.

Procedure

1. Identify the task category (current-docs / browser / desktop-tauri / figma / crawl / search)
2. Look up preference list for that category
3. Call pickMcp(preferenceList) from scripts/lib/mcp-registry.mjs — returns first available, or null
4. If MCP available: use its canonical tools (e.g. mcp__mcp-server-context7__query-docs)
5. If no MCP available: fall back to native tools (WebFetch / WebSearch / Grep / Read)
6. Document choice in agent output: "Used MCP <name>" OR "No suitable MCP, fell back to <native>"

Output contract

Returns:

• { mcp: '<name>' | null, tools: ['<tool1>', '<tool2>'], fallback: '<reason>' | null }
• Cite in agent output

Anti-patterns

• Hardcoding tools: [mcp__context7__*] in agent frontmatter → breaks when user lacks context7
• Assuming Tauri MCP is screenshot-only → misses driver session, webview DOM, IPC monitoring, backend state, logs, windows, setup, and device tools when available
• Using Playwright for native Tauri IPC/window/log verification → only validates a browser preview, not the Rust desktop shell
• Calling MCP tool without checking availability → cryptic error
• Not surfacing fallback choice → user can't tell why agent took longer / had less detail
• Falling back silently → user thinks MCP was used; can't diagnose

Verification

• getRegistry() returns ≥1 MCP OR explicit fallback documented
• Output names which MCP was used (or "no MCP, fell back to X")

Guard rails

• Do not mutate files, provider state, network resources, or external tools unless this skill's procedure and the user approval path allow it.
• Do not skip prerequisites, confidence gates, policy gates, or explicit approval gates.
• Do not claim completion without concrete verification evidence.
• Preserve user-owned content and unrelated worktree changes.

Related

• Tool: scripts/lib/mcp-registry.mjs — registry helpers
• Tool: scripts/discover-mcps.mjs — refresh registry
• Status: npm run supervibe:status shows available MCPs
• Used by: best-practices-researcher / competitive-design-researcher / dependency-researcher / preview-server (for screenshots)