@mgong/mcp-to-skills

中文文档 · English

Discover every MCP server already configured in Claude Code, connect to them, and turn each one into a Skill that Claude or Cursor can pick up automatically.

For every server you get:

• SKILL.md with frontmatter, server description, server instructions, and a per-tool reference table
• tools/<tool>.json — the raw JSON Schema of every tool (optional)
• mcp.json — a backup of the original MCP entry, so you can safely delete it from ~/.claude.json

Supported transports: stdio, Streamable HTTP, SSE.

Supported sources (merged automatically):

• Global — top-level mcpServers in ~/.claude.json (use --config to point elsewhere)
• Project (cwd) — <cwd>/.mcp.json (Claude Code project format) and <cwd>/.cursor/mcp.json (Cursor project format)

The legacy projects[*].mcpServers map inside ~/.claude.json is not read by default; pass --include-projects if you still want it.

---

Quick start (no install)

# 1. List MCP servers found on this machine and probe their tools
npx @mgong/mcp-to-skills list

# 2. Generate ./skills/<server>/SKILL.md for every reachable server
npx @mgong/mcp-to-skills sync

# 3. Install the generated skills into Claude / Cursor (interactive picker)
npx @mgong/mcp-to-skills install

Or install globally once:

npm i -g @mgong/mcp-to-skills
mcp-to-skills sync
mcp-to-skills install

After sync your tree looks like this:

skills/
  context7/
    SKILL.md                 # primary entry the model loads
    mcp.json                 # backup of the original MCP entry
    tools/
      resolve-library-id.json
      query-docs.json
  mcp-backup.json            # aggregate backup, shaped like ~/.claude.json

---

Commands

Command	What it does
list	Probe every server, print connectivity + tool list
sync	Probe servers and write SKILL.md + config backups
install	Copy ./skills/<server>/ into a target skills directory
setup	One-shot interactive wizard: pick MCPs → generate → pick skills → pick install target

Run mcp-to-skills --help for the full list of flags.

Common flags

# Pick a different output directory
mcp-to-skills sync --out ~/.claude/skills

# Only the global config (skip <cwd>/.mcp.json and <cwd>/.cursor/mcp.json)
mcp-to-skills list --no-local

# Only project-level configs in the current directory (skip ~/.claude.json)
mcp-to-skills list --no-global

# Scan a specific directory for project-level configs
mcp-to-skills list --cwd /path/to/some/project

# Also include the legacy projects[*].mcpServers map from ~/.claude.json
mcp-to-skills list --include-projects

# Don't emit per-tool JSON schemas
mcp-to-skills sync --no-schemas

# Don't write config backups (mcp.json / mcp-backup.json)
mcp-to-skills sync --no-backup

# Keep raw env / headers values in backups (NOT recommended; default is to redact)
mcp-to-skills sync --no-redact-env

# Bigger timeout for slow-starting stdio servers
mcp-to-skills list --timeout 30000

# Use a different global config file
mcp-to-skills list --config /path/to/some.claude.json

# More parallel probes
mcp-to-skills list --concurrency 8

---

install — copy skills into Claude / Cursor

sync only writes into ./skills/. To make Claude Code or Cursor actually load them, copy the directory into the client's skills folder. install is an interactive helper for that step (built on @clack/prompts).

mcp-to-skills install

◆  Where do you want to install the skills?
│  ● Claude Code (global)    /Users/you/.claude/skills
│  ○ Claude Code (project)   .claude/skills
│  ○ Cursor (global)         /Users/you/.cursor/skills
│  ○ Custom path…            enter your own

• Custom path… opens a text input (~/..., relative or absolute paths supported)
• If the target already has a same-named skill, it asks before overwriting (default: skip)
• Copies the whole skill folder: SKILL.md + mcp.json + tools/

Non-interactive / CI:

# Take the default (~/.claude/skills) without prompting
mcp-to-skills install --yes

# Explicit target (also skips the menu)
mcp-to-skills install --to ~/.cursor/skills

# Replace existing same-named skills
mcp-to-skills install --to ~/.claude/skills --overwrite

# Just print the actions without writing anything
mcp-to-skills install --dry-run

# Use a different source folder (default ./skills)
mcp-to-skills install --from ./generated --to ~/.claude/skills

---

Use as a library

import { discoverAll, loadClaudeCodeConfig, loadLocalMcpConfigs, writeSkill } from '@mgong/mcp-to-skills'

const servers = [
  ...await loadClaudeCodeConfig(), // global ~/.claude.json (top-level only)
  ...await loadLocalMcpConfigs({ cwd: process.cwd() }), // <cwd>/.mcp.json + <cwd>/.cursor/mcp.json
]
const results = await discoverAll(servers, { timeoutMs: 20_000 })

for (const r of results.filter(r => r.ok && r.tools.length > 5))
  await writeSkill(r, { outDir: './big-skills' })

Public exports:

Export	Purpose
loadClaudeCodeConfig(opts?)	Parse ~/.claude.json (top-level mcpServers only by default; pass includeProjects: true for projects[*])
loadLocalMcpConfigs({ cwd? })	Parse <cwd>/.mcp.json and <cwd>/.cursor/mcp.json (deduped via realpath)
discoverServer(server, opts?)	Connect to a single server and list its tools
discoverAll(servers, opts?)	Probe many servers in parallel
renderSkill(result)	Render a DiscoveryResult into a SKILL.md string
writeSkill(result, { outDir })	Write SKILL.md and tools/*.json to disk
writeServerBackup(result, outDir, { redactEnv? })	Write <outDir>/<name>/mcp.json
writeAggregateBackup(results, outDir, { configPath?, redactEnv? })	Write <outDir>/mcp-backup.json
sanitizeEntry(entry, { redactEnv? })	Clone an MCP entry, optionally redacting secrets
listSkillDirs(from)	List subdirectories of from that contain a SKILL.md
installSkills(from, to, { overwrite?, dryRun? })	Copy <from>/<server>/ into <to>/<server>/

---

Backup & restore

sync writes the original MCP server config alongside the generated skill so you can safely delete the mcpServers field from ~/.claude.json afterwards. Two files are produced:

File	Purpose
<outDir>/<server>/mcp.json	A single server's entry (name / source / entry, plus lastError if the connection failed)
<outDir>/mcp-backup.json	Everything aggregated and grouped to mirror ~/.claude.json (mcpServers + projects[*].mcpServers)

Restore everything back to ~/.claude.json

cp ~/.claude.json ~/.claude.json.bak

jq -s '.[0] * { mcpServers: .[1].mcpServers }' \
  ~/.claude.json skills/mcp-backup.json > ~/.claude.json.next \
  && mv ~/.claude.json.next ~/.claude.json

If you also want to merge projects.*.mcpServers:

jq -s '
  .[0]
  * { mcpServers: .[1].mcpServers }
  * { projects: ((.[0].projects // {}) * (.[1].projects // {})) }
' ~/.claude.json skills/mcp-backup.json > ~/.claude.json.next \
  && mv ~/.claude.json.next ~/.claude.json

Restore a single server

jq --slurpfile s skills/context7/mcp.json \
  '.mcpServers["\($s[0].name)"] = $s[0].entry' \
  ~/.claude.json > ~/.claude.json.next \
  && mv ~/.claude.json.next ~/.claude.json

Backups created with --no-redact-env contain raw secrets (env values and HTTP headers). Do not commit them to a public repo.

---

Generated SKILL.md

---
name: context7
description: "Use this server to fetch current documentation whenever the user asks about a library, framework, SDK, API, CLI tool, or cloud service... Use when the user wants to call any of the context7 tools: `resolve-library-id`, `query-docs`."
---

# context7

> Source: claude-code:/Users/you/.claude.json
> Server: `Context7` v2.1.8

## Server instructions

Use this server to fetch current documentation whenever the user asks about a library...

## Tools

### `resolve-library-id`

Resolves a package/product name to a Context7-compatible library ID...

**Input schema**

| Name          | Type   | Required | Description                                |
| ------------- | ------ | -------- | ------------------------------------------ |
| `query`       | string | yes      | The question or task you need help with... |
| `libraryName` | string | yes      | Library name to search for...              |

---

Troubleshooting

Symptom	Likely cause
FAIL (transport: ...)	Wrong / missing fields on the URL or command
FAIL (... timed out)	Server starts slowly — retry with --timeout 30000
FAIL (spawn xxx ENOENT)	The stdio command is not on PATH. Use an absolute path
FAIL (401 / 403)	The HTTP server needs auth headers (headers in the entry)
No MCP servers found	Neither ~/.claude.json nor <cwd>/.mcp.json / <cwd>/.cursor/mcp.json defines any server. Pass --include-projects for the legacy projects[*].mcpServers map, or cd into a project that has a local .mcp.json

To see a stdio server's own stderr, edit src/discover.ts and switch stderr: 'pipe' to 'inherit'.

---

Security

• Backups (mcp.json and mcp-backup.json) redact env and headers values to "***" by default
• Pass --no-redact-env only when you fully control the destination
• The skills/ directory is in the project's .gitignore. Never commit it
• This tool reads ~/.claude.json, spawns the MCP server processes you already have configured (stdio), or makes outbound HTTP requests to URLs you already trust. It does not phone home

Found a security issue? Please report privately to the email in package.json (or open a draft security advisory on GitHub) instead of filing a public issue.

---

Contributing

pnpm install
pnpm typecheck   # tsc --noEmit
pnpm lint        # antfu eslint
pnpm test        # vitest
pnpm build       # produces dist/cli.js + dist/index.js + .d.ts

PRs welcome. Please:

1. Add a test for new behavior (tests/*.test.ts)
2. Run pnpm lint:fix && pnpm test && pnpm typecheck before opening the PR
3. Update CHANGELOG.md under "Unreleased"

---

Roadmap

• sync hash compare — skip writing identical files (better git diff)
• sync --prune — delete skill directories whose source server is gone
• install --prune — mirror deletes at the install target
• Use server instructions first paragraph as description frontmatter (fallback: tool descriptions + tool names)
• install interactive command
• Pull configs from more clients: Cursor (global / project), Claude Desktop, …

---

License

MIT © Michael Gong