Mcp Installer

---
name: mcp-installer
description: "Registers, verifies, or imports MCP (Model Context Protocol) servers in LibrAgent or Claude Desktop. Use when users share MCP connection info (npm/npx/PyPI package, MCP server GitHub repo, config JSON, server URL), want a capability that needs a new MCP server (\"I want stock data\"), or want to import/sync settings from Cursor, VS Code, Windsurf, and similar editors. GitHub URL rule: a repo or subpath that contains SKILL.md is a skill — use skill-deployer instead; an MCP server package or MCP config uses this skill; if unsure, inspect the repo before installing. Capability requests (search the catalog with tool__listServers scope=catalog first, not the shell): 데이터 볼 수 있게 해줘, 기능 추가해줘, \"let me see X data\". Triggers: install MCP, add MCP server, connect this, 연결해줘, 깔아줘, 설치해줘, MCP 추가, 서버 등록, 데이터 볼 수 있게 해줘, 기능 추가해줘."
---

# MCP Installer

This skill registers MCP servers into the current LibrAgent session or the user's Claude Desktop configuration. It covers both **direct installation** (from packages or URLs) and **import/sync** (from other AI editors).

## 🚀 Workflow

### Step 0: Determine Input Path

| User intent | Path |
| --- | --- |
| "Import from Cursor", "sync my MCP servers", "what's installed elsewhere?" | **Import** → below |
| npm/PyPI package, MCP server GitHub repo, server URL, or pasted JSON config | **Direct install** → below |
| GitHub URL whose repo or subpath contains `SKILL.md` | Not an MCP server → use **skill-deployer** |
| GitHub URL and you cannot tell | Inspect it first (README, `package.json`/`pyproject.toml`, `SKILL.md`), then pick the path. Ask the user only if it is still unclear. |
| A capability with no package, URL, or config ("I want stock data", "connect Slack") | **Catalog** → below |
| "Install it" with no package, URL, config, or capability | Ask what to install. Do not guess a package. |

### Catalog Workflow

1. **Search the curated catalog** — Call `tool__listServers({"scope":"catalog","query":"<keyword>"})` with a short keyword from the request (e.g. `stock`, `slack`, `search`). Omit `query` to see every preset.
2. **Pick a preset** — If one matches, use it as is: it comes from LibrAgent's official list. If it is marked `registered`, it is already installed; do not register it again (use `tool__updateServer` to change it).
3. **Collect required inputs** — For each entry in `requiredVariables`, replace the matching placeholder (e.g. `YOUR_GITHUB_PAT`) in the preset's `transport`. For a secret (API key, token, password), write `{{secret:NAME}}` (NAME in `UPPER_SNAKE_CASE`, e.g. `Bearer {{secret:GITHUB_PAT}}` or `?exaApiKey={{secret:EXA_API_KEY}}`); LibrAgent asks the user for the value in a secure prompt when you register. Never ask for a secret in chat. Ask in chat only for non-secret values.
4. **Register and verify** — Call `tool__registerServer` with the preset `name`, `description`, `transport` (and `authentication` when present), then run Step 3 of the direct install workflow. If each secret input is a `{{secret:NAME}}` template and every other value stays as the preset gives it, the install needs no user approval. Changing a non-secret value (for example a server URL), the command, args, or URL, adding env/header keys, or passing `authentication` makes it a normal approval-gated install. Presets that track `@latest` run new upstream releases without asking again.
5. **No match** — Only then search the web or npm for a package. Tell the user the source is not curated by LibrAgent before you register it.

### Import Workflow

1. **Identify sources** — Consult [app_paths.md](references/app_paths.md) for config locations.
2. **Scan** — Read detected config files. Resolve environment variables (`%APPDATA%`, `%USERPROFILE%`, `~`) before reading. If a project-level config (e.g., `.cursor/mcp.json`) exists in the current workspace, prioritize it.
3. **Compare** — List all found MCP servers. Use `tool__listServers` to compare with LibrAgent's current registrations. Mark servers as new or already registered.
4. **Select** — Present the list to the user. Ask which servers to import (all or specific ones).
5. **Register** — Use `tool__registerServer` for each selected server. Skip duplicates (same name and command). Each call asks the user for approval (see Step 2 below).
6. **Verify** — Run Step 3 below for each registered server.
7. **(Optional)** — Offer to update `claude_desktop_config.json` as well.

**Schema notes:** VS Code Copilot uses `servers` as the root key; Zed uses `context_servers`. Normalize entries before registration.

### Direct Install Workflow

#### 1. Analyze Information

Identify the source and transport details.

- **npx/npm**: Use `npx -y <package>` for registration.
- **GitHub**: Determine if it's TypeScript (needs build) or Python (can use `uvx`).
- **Claude Config**: Extract command, args, and env variables.

#### 2. Register in LibrAgent

Use the `tool__registerServer` tool to add the MCP server. Never hand-edit LibrAgent config files or the database.

`tool__registerServer` (except an unchanged catalog preset) and `tool__updateServer` always pause for the user's approval, even in YOLO mode, because a stdio server runs an arbitrary command. The saved server records its source (`catalog` preset or `external`); the success message shows it. Before you call it, tell the user in one line what will run (command and args, or URL) and which env variables it needs. If the user rejects it, stop and ask what to change; do not retry the same call.

**Example Task:**
"Register the 'everything' MCP server using npx."

**Execution:**
```json
{
  "name": "everything",
  "description": "MCP server with all capabilities",
  "transport": {
    "type": "stdio",
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-everything"]
  }
}
```

#### 3. Verify (required)

Installation is not done until the server is verified.

1. Call `tool__verifyServer` with the server name.
2. On success, report the tools it exposes, then go to Step 3a. Saving a server does not make it callable in any session yet.
3. On failure, report the real error and the next step:
   - `command not found` / missing `node`, `npx`, `uv`, `python` → run **setup-wizard** or install the runtime, then call `tool__updateServer` for the saved server, and then verify.
   - Missing or wrong API key / env variable → call `tool__updateServer` with the secret as `{{secret:NAME}}` and `reenterSecrets: ["NAME"]`, so the user enters it again in the secure prompt. Do not ask for it in chat.
   - Package not found or crashes on start → show the error output; do not report success.

#### 3a. Offer to attach (required after a successful verify)

1. Ask: "Attach `<server>` to this assistant so it can use these tools?" Use the agent id from `## Agent Runtime Identity`.
2. **Yes** → read the assistant's current `externalMcpServers` with `agent__listAgents(type: "configs")`, then call `agent__updateAgent(id: "<agent id>", externalMcpServers: [...current ids, "<server id>"])`. The list is replaced, not merged, so always include the existing ids. Then tell the user: "Start a new session to use the `<server>` tools. The current session cannot load them."
3. **No** → tell the user the server is saved but not attached, and that they can attach it later to any assistant (by name) in Settings or by asking you.

#### 4. Claude Desktop (Optional)

If the user specifically asks to install it in Claude Desktop:

1. Locate `claude_desktop_config.json` (see [app_paths.md](references/app_paths.md)).
2. Read the current configuration.
3. Append the new server entry.
4. Save the file and notify the user to restart Claude Desktop.

## 🛠 Guidelines

- **Duplicate Prevention**: Do not register servers that already exist in LibrAgent with the same name and command.
- **Naming**: Use lowercase slugs (e.g., `github`, `brave-search`).
- **Safety**: Verify the source before execution.
- **Secrets**: Write every secret env value, header value, or URL query value as `{{secret:NAME}}`; the user enters it in a secure prompt and it is stored outside the config. Never ask for a secret in chat. If the user sends a secret in chat anyway, do not repeat it in your reply or in any tool call; use a `{{secret:NAME}}` template and let the prompt collect it.
- **Dependencies**: Use [setup-wizard](../setup-wizard/SKILL.md) if `node`, `npm`, or `uv` are missing.
- **Path Resolution**: Resolve environment variables before reading config files on any platform.

## Reference

- [installation.md](references/installation.md): npm, GitHub, and Python install patterns
- [app_paths.md](references/app_paths.md): Config paths for Cursor, VS Code, Windsurf, and other editors