Agent Skills › AjayIrkal23/agentic-mercy-10x › mcp-usage-standards

mcp-usage-standards

GitHub

规范 MCP 工具选择策略，依据场景匹配 context7、fetch 等工具。遵循最小证据原则，优先本地资源，禁止传输密钥，确保验证纪律与正确性。

skills/mcp-usage-standards/SKILL.md AjayIrkal23/agentic-mercy-10x

触发场景

需要选择外部 MCP 工具时涉及证据范围界定或外部验证策略时工具选择影响设计、调试或代码库分析时

安装

npx skills add AjayIrkal23/agentic-mercy-10x --skill mcp-usage-standards -g -y

更多选项

不安装直接使用

npx skills use AjayIrkal23/agentic-mercy-10x@mcp-usage-standards

指定 Agent (Claude Code)

npx skills add AjayIrkal23/agentic-mercy-10x --skill mcp-usage-standards -a claude-code -g -y

安装 repo 全部 skill

npx skills add AjayIrkal23/agentic-mercy-10x --all -g -y

预览 repo 内 skill

npx skills add AjayIrkal23/agentic-mercy-10x --list

SKILL.md

Frontmatter

{
    "name": "mcp-usage-standards",
    "schema": 1,
    "category": "general",
    "surfaces": [
        "general"
    ],
    "triggers": {
        "paths": [],
        "intents": [
            "general"
        ],
        "keywords": [
            "affects",
            "analysis",
            "avoid",
            "choose",
            "debugging",
            "design",
            "disciplined",
            "evidence",
            "external",
            "implementation",
            "materially",
            "mcp",
            "narrow",
            "repo",
            "retrieval",
            "right",
            "scope",
            "selection",
            "standards",
            "strategy",
            "tool",
            "unnecessary",
            "usage",
            "verification"
        ]
    },
    "platforms": [
        "linux",
        "darwin",
        "windows"
    ],
    "token-cost": 805,
    "description": "ALWAYS invoke when MCP selection, evidence scope, or external verification strategy materially affects design, debugging, implementation, or repo analysis. MUST use to choose the right MCP, narrow the evidence scope, keep evidence retrieval disciplined, and avoid unnecessary tool usage.",
    "disable-model-invocation": false
}

MCP Usage Standards

Overview

MCP routing for this machine is driven by ~/.claude/settings.json. Load this skill when tool choice materially affects correctness, evidence, or verification — not for routine single-file edits with obvious local answers.

Canonical quick table: ~/.claude/rules/user-mcp-inventory.mdc.

Active MCP registry (`~/.claude/settings.json`)

Server	Use when	Skip when
`context7`	Official/current docs for libraries, frameworks, SDKs, APIs, CLIs, cloud APIs	Answer is purely in-repo; no external doc uncertainty
`sequential-thinking`	Complex decomposition, ambiguous tradeoffs, high-risk reasoning	Straightforward edits or shallow questions
`fetch`	Static HTML/markdown-ish pages; URLs where no JS interaction is needed	SPA state, clicks, auth flows, or rendered behavior required
`memory`	Durable stash/recall genuinely worth carrying across threads	Secrets, chatter, ephemeral task state
`browser-tools-mcp`	That server’s console/network/DevTools-oriented evidence is explicitly needed	Same check is cheaper via `fetch` or repo-only
`playwright`	Real browser automation: flows, clicks, accessibility snapshots / UI probes (`@playwright/mcp`)	Static fetch or local code read suffices
`markdownify`	Convert HTML/PDF/etc. → markdown for ingest (respect server path/env rules)	No conversion workflow
`graphify`	Query knowledge graph backed by `graphify-out/graph.json`	`graph.json` missing — build graph first (`graphify` / wiki); MCP will fail

Browser MCP triage: fetch (static) → playwright (drive the page) → browser-tools-mcp when you need that toolchain’s diagnostics; don’t stack two browsers for one trivial question.

Selection rules

Prefer repo files, project docs, AGENTS.md / linkage files before MCP for codebase questions.
Prefer context7 over ad-hoc web answers for upstream library correctness.
Prefer fetch before playwright when HTML is effectively static.
Prefer playwright before ad-hoc shell curl | grep UI checks when deterministic browser actions matter.
Use memory sparingly — never secrets.
Use markdownify only when conversion is part of the task scope.
Use sequential-thinking only when reasoning cost justifies serialized steps.

Non-negotiables

Pick MCP intentionally; smallest tool wins.
Minimum evidence retrieval; summarize what changed your mind.
No secrets/tokens/credentials on the wire or in memory entries.
If an MCP fails, change strategy once — state unverified assumptions.

Workflow

Name what is unknown vs what local source should prove.
Choose one MCP tier (fetch vs playwright vs docs vs thinking).
Narrow the prompt before calling tools.
Record what evidence confirmed or ruled out.

See references/full-guide.md for per-server detail and appendix on optional / not-installed MCPs.

Completion checklist

Correct MCP tier chosen.
No redundant browser/doc calls.
No secrets leaked.
Findings summarized.

版本历史

581d130 当前 2026-07-19 09:12

元信息

文件数: 0
版本: fc73590
Hash: 0bfee940
收录时间: 2026-07-19 09:12

mcp-usage-standards