mcp-to-skill（MCP TO 技能）

MCP-to-Skill Converter

Convert MCP servers into on-demand skills. MCP tool schemas sit in the system prompt on every turn (~500-2000 tokens per tool, regardless of whether they're used). Skills cost zero tokens until loaded via view. For a typical setup with 4-5 MCP servers exposing 20-40 tools, this reclaims 10,000-30,000 tokens of context per turn.

This matters because that's 10-30% of the context window burned before the conversation even starts — and it compounds: every turn re-injects the full schema.

Decision Framework: Convert vs. Keep

Not every MCP should become a skill. Apply this heuristic:

Convert when the MCP wraps a REST API (use curl/web_fetch), wraps a CLI tool (gh, aws, gcloud — invoke directly), implements a reasoning/planning pattern (capture as methodology), or when you use fewer than half its tools regularly.

Keep as MCP when it maintains persistent server-side state (DB connections, WebSocket sessions), handles binary protocols or streaming, provides real-time event subscriptions, or is tiny (1-2 tools, under 500 tokens — negligible overhead).

Hybrid approach — convert the stateless tools to a skill, keep stateful ones as a slimmed-down MCP. This is often the sweet spot for large MCP servers.

Conversion Workflow

Proceed through 5 phases. Present findings at each phase boundary and wait for user confirmation before continuing. The user knows their usage patterns better than any analysis can infer — lean on their input.

Phase 1: Discovery

Acquire the MCP's tool definitions. Try these sources in order:

1. Active session tools — Inspect tools visible in the current conversation. Ask the user to identify which tools belong to the target MCP. This is the most reliable source because you see the exact schema consuming context.
2. MCP config file — Parse the user's MCP configuration:

- Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json - Cursor: .cursor/mcp.json in the project root - Claude Code: ~/.claude/settings.json or project .mcp.json - Config files give server names and connection details, not tool schemas.

1. MCP server source code — If the user points to a repo or local path, look for tool definitions: FastMCP @mcp.tool() decorators, SDK server.setRequestHandler, or similar patterns. Extract name, description, parameter schemas, return types.
2. Package registry — For published MCPs: npm info <pkg> or pip show <pkg>, then fetch the README or source to find tool definitions.
3. User-provided schema — Ask the user to paste or upload tool definitions.

Produce a structured inventory for each tool:

Tool: tool_name
Description: what it does
Parameters: param list with types
Returns: return type/shape
Estimated tokens: rough schema size

Present this and ask: "Are these all the tools? Did I miss any?"

Phase 2: Classification

Classify each tool along two dimensions. This classification drives the entire replacement strategy, so getting it right matters.

Replacement category:

Usage frequency — Ask the user directly:

Present a classification table and ask: "Does this look right? Which tools do you actually use regularly?"

Flag any STATEFUL tools explicitly — these are the ones that may not convert cleanly, and the user should understand the trade-off.

Phase 3: Replacement Strategy

For each tool marked ESSENTIAL or NICE_TO_HAVE, design the concrete replacement.

Read references/replacement-patterns.md — it contains detailed patterns for each category: REST API wrappers, CLI mappings, logic patterns, file operations, stateful workarounds, composite workflows, auth patterns, and output parsing.

For each tool, determine:

• The exact command (curl, CLI, or methodology) that replaces it
• How MCP tool parameters map to command arguments
• How to parse the output into a useful format
• Common error cases and their fixes

Also identify multi-tool workflows — sequences of tools the user commonly chains. These become "Common Workflows" sections in the generated skill, which is where skills often provide more value than the MCP because workflows make the multi-step pattern explicit rather than relying on the agent to discover it.

Ask the user:

• "What CLI tools are available in your environment?"
• "Are there common sequences where you use multiple tools together?"
• "How do you handle authentication?" (env vars, config files, OAuth tokens)

If the target environment is unclear, read references/environment-guide.md for environment-specific constraints (Claude.ai vs Claude Code vs Cursor vs API).

Phase 4: Generation

Generate the complete skill package.

Read references/skill-template.md for the output template, sizing guide, frontmatter checklist, and quality checklist.

The generated skill structure:

skill-name/
  SKILL.md
    Frontmatter (name, description with aggressive triggers)
    Quick Reference table (old tool name to new command mapping)
    Prerequisites (CLI tools, env vars, auth setup)
    Core Operations (one subsection per essential tool)
    Common Workflows (multi-step patterns)
    Error Handling and Troubleshooting
  references/                        (only if SKILL.md exceeds ~400 lines)
    api-reference.md                 (overflow for complex tool replacements)

Generation rules — these exist to ensure the generated skill actually triggers and works correctly in practice:

• Frontmatter description must be pushy. Include original MCP tool names as trigger phrases, the service name, action verbs, and explicit "Use this skill when..." language. Skills undertrigger by default; compensate with a broad net.
• Quick Reference table at the top. Users and agents scan this first.
• Each Core Operation shows what it replaces, the replacement command, parameter mapping, a concrete example, and error handling.
• Keep SKILL.md under 500 lines. Move detailed patterns to references/ if needed.
• Auth via env vars or config, never hardcoded. Generated skills must not embed credentials.
• Prerequisites include install commands for every required CLI tool.

Phase 5: Validation

After generating, validate the skill and estimate savings.

Run the token estimation using scripts/estimate_tokens.py:

python3 scripts/estimate_tokens.py --mcp-tools TOOL_COUNT --avg-schema-chars AVG_CHARS

This shows before/after token savings per turn and across a typical conversation.

Opus 4.7 note: Input tokens run 1.0–1.35× Opus 4.6 for the same text due to a tokenizer update. Treat pre-4.7 baselines as a lower bound — actual savings on Opus 4.7 may be larger than the estimator reports.

Generate 2-3 test scenarios — realistic prompts that would trigger the new skill and show the replacement commands in action. Present them to the user.

Migration checklist:

• Generated skill reviewed and any edits applied
• Required CLI tools installed and authenticated
• Skill placed in the target skills directory
• MCP removed from configuration
• Test scenarios validated

Validate the generated skill structure:

Verify the generated skill directory contains a valid SKILL.md with frontmatter (name, description), and that all file references in the body resolve to existing files within the skill directory.

Present the complete package to the user. Offer to iterate on any section.

Limitations

Conversions succeed best for stateless tools and REST/CLI wrappers. Inherent constraints:

• Stateful tools don't convert cleanly — MCPs maintaining persistent connections, sessions, or real-time subscriptions should stay as MCPs. See "Keep as MCP when" in the Decision Framework.
• Binary protocols and streaming — If the MCP handles binary data or WebSocket streams, conversion requires additional infrastructure outside Claude's scope.
• API fabrication risk — Replacement strategy only works if the underlying API or CLI is known. Unknown APIs must be researched first; guessing produces broken skills.
• Auth complexity — Conversions with multi-step OAuth or credential management are possible but require explicit env var and config setup. See Phase 3 for auth patterns.
• Partial coverage acceptable — A 90% conversion with one stateful MCP remaining is preferable to a broken attempt at 100%. Acknowledge trade-offs honestly.

Constraints

These exist to prevent common failure modes in generated skills:

• Never fabricate API endpoints. If the underlying API is unknown, ask the user or research the MCP's source code. Guessing at URLs produces broken skills.
• Acknowledge limitations honestly. A partial conversion (80% of tools) with one remaining small MCP is better than a broken skill that claims full coverage.
• Test mentally before presenting. Trace through each replacement command: would it actually work? Does curl need specific headers? Does the CLI tool require auth setup?
• Preserve error handling quality. Many MCPs provide helpful error messages. The generated skill should include equivalent troubleshooting guidance.