using-mcp

name: using-mcp description: Discover, connect to, and invoke tools on MCP (Model Context Protocol) servers. CRITICAL - When the user message contains [@mcp:...] mention, you MUST load this skill first to use MCP tools correctly.

name-cn: MCP 工具调用技能 description-cn: 发现并调用 MCP（Model Context Protocol）服务器上的工具。关键规则：当用户消息包含 [@mcp:...] 引用时，必须首先加载此技能。

MCP Tools Calling Skill

Use this skill whenever you need to talk to an MCP server: list servers, connect to one, inspect a tool's schema, or invoke a tool.

How it works

The six MCP capabilities are exposed as Code Mode tools (mcp_*). They are NOT directly callable as standalone tool calls. You must invoke them through run_sdk_snippet and sdk.tool.call:

run_sdk_snippet(python_code="""
from sdk.tool import tool

result = tool.call('mcp_list_servers', {})
print(result.content)
""")

tool.call(name, params) returns a Result with:

• result.ok — boolean, check this first
• result.content — complete information for reasoning and next-step decisions

Available MCP tools

Standard workflow

Follow this order. Skipping steps will cause failures because tool names and parameters cannot be guessed.

1. mcp_list_servers          → pick the server, read its `status`
   ├── status == 'connected'    → go to step 3
   └── status == 'disconnected' → go to step 2
2. mcp_connect_server        → connect; receive the real tool list
3. mcp_get_tool_schema       → fetch input schema(s)
4. mcp_call_tool             → invoke with parameters that match the schema

When the user wants to register a brand-new MCP server, run mcp_add_server first, then proceed from step 1.

Rules

1. NEVER fabricate server names, tool names, or parameter names. Always derive them from the previous step's result.content.
2. ALWAYS call mcp_get_tool_schema before mcp_call_tool unless you already saw the schema in this conversation.
3. NEVER call mcp_* tools as standalone tool calls. They only work inside run_sdk_snippet via sdk.tool.call.
4. ALWAYS check result.ok before proceeding. Errors should be surfaced to the user, not silently retried.

Environment variables in reusable MCP configs

MCP configs can be written as reusable templates in skills. Put sensitive values behind ${VAR_NAME} placeholders, and ask the user to save their own value with env-manager before registering or connecting the MCP server.

Supported placeholder locations:

• url, including query parameters: https://mcp.example.com/sse?key=${EXAMPLE_API_KEY}
• headers values: {"x-api-key": "${EXAMPLE_API_KEY}"}
• token
• stdio env values: {"EXAMPLE_API_KEY": "${EXAMPLE_API_KEY}"}

Do not put personal API keys directly in skill docs or MCP config templates. If a referenced env variable is missing or unusable, connecting the MCP server fails with the missing variable name. The secret value is not written into the persisted MCP config.

Only the fields above are resolved. Do not use ${VAR_NAME} in name, description, command, or args.

End-to-end example

User asks to call a tool on some MCP server. The flow below uses angle- bracket placeholders (<server_name>, <tool_name>, <workspace_dir>, etc.); replace every <...> with the real value from your context or from the previous step before running the snippet.

run_sdk_snippet(python_code="""
from sdk.tool import tool

# 1. Discover servers
servers = tool.call('mcp_list_servers', {})
print(servers.content)
# Parse the server name from content, e.g. "<server_name> (status=connected, N tool(s))"

# 2. Connect if needed
connect = tool.call('mcp_connect_server', {'server_name': '<server_name>'})
if not connect.ok:
    raise SystemExit(f'Connect failed: {connect.content}')
print(connect.content)

# 3. Inspect schema
schema = tool.call('mcp_get_tool_schema', {
    'server_name': '<server_name>',
    'tool_name': '<tool_name>',
})
print(schema.content)

# 4. Call the tool (tool_params is a JSON string matching the schema)
result = tool.call('mcp_call_tool', {
    'server_name': '<server_name>',
    'tool_name': '<tool_name>',
    'tool_params': '{"key": "value"}',
})
print(result.content)
""")

Saving the call result to a file (output_file_path)

mcp_call_tool accepts an optional output_file_path to persist the result as a JSON file in the workspace, instead of consuming it only in your reasoning context.

When to set it:

• The result is likely to be large (for example a SQL query without WHERE or LIMIT that may return thousands of rows, fetching a full article body, or exporting a long list).
• The user expects a concrete file deliverable (a report, dataset dump, attachment-like artifact).
• A downstream step needs a stable on-disk path to open later.

When to leave it empty:

• You only need the data for your own reasoning. Small results are returned inline; oversized results are auto-persisted in the background to avoid flooding the context, so you do not need to manage that yourself.

Rules:

1. The path MUST be absolute. Relative paths are rejected and the call silently falls back to the runtime directory.
2. Prefer a location inside the current workspace so the file is easy for the user to inspect and accept as a deliverable. The workspace path is already available in your context — just prepend it directly, e.g. <workspace_dir>/data/<server>/<tool>.json. Do not rebuild it from os.getcwd() or hard-code a path from another machine.
3. The file MUST be JSON. Pick a meaningful filename and a tidy directory layout, for example reports/<topic>/<name>.json or data/<server>/<tool>.json.

Example — deliver the call result as a workspace file. Replace every <...> with real values: <workspace_dir> is the workspace path already visible in your context, and <server_name> / <tool_name> come from the previous discovery steps:

run_sdk_snippet(python_code="""
from sdk.tool import tool

result = tool.call('mcp_call_tool', {
    'server_name': '<server_name>',
    'tool_name': '<tool_name>',
    'tool_params': '{"key": "value"}',
    'output_file_path': '<workspace_dir>/data/<server_name>/<tool_name>.json',
})
print(result.content)
""")

Adding a new server

run_sdk_snippet(python_code="""
from sdk.tool import tool

# stdio server (npx-based MCP).
# `args` MUST be a real list, not a space-separated string.
add = tool.call('mcp_add_server', {
    'name': 'my-fs-server',
    'server_type': 'stdio',
    'command': 'npx',
    'args': ['-y', '@modelcontextprotocol/server-filesystem', '/tmp'],
    'label_name': 'Filesystem',
})
print(add.content)

# stdio server with an API key injected into the subprocess environment.
# First save EXAMPLE_API_KEY with set_env, then register this template.
add = tool.call('mcp_add_server', {
    'name': 'example-stdio-auth',
    'server_type': 'stdio',
    'command': 'npx',
    'args': ['-y', '@example/mcp-server'],
    'env': {'EXAMPLE_API_KEY': '${EXAMPLE_API_KEY}'},
    'label_name': 'Example Stdio Auth',
})
print(add.content)

# http server example
add = tool.call('mcp_add_server', {
    'name': 'my-api-server',
    'server_type': 'http',
    'url': 'http://localhost:3000/mcp',
    'label_name': 'Custom API',
})
print(add.content)

# http server with an API key stored through env-manager.
# First save EXAMPLE_API_KEY with set_env, then register this template.
add = tool.call('mcp_add_server', {
    'name': 'example-header-auth',
    'server_type': 'http',
    'url': 'https://mcp.example.com/http',
    'headers': {'x-api-key': '${EXAMPLE_API_KEY}'},
    'label_name': 'Example Header Auth',
})
print(add.content)

# URL query auth example for a reusable skill template.
add = tool.call('mcp_add_server', {
    'name': 'example-query-auth',
    'server_type': 'http',
    'url': 'https://mcp.example.com/sse?key=${EXAMPLE_API_KEY}',
    'label_name': 'Example Query Auth',
})
print(add.content)
""")

After mcp_add_server, the server status is disconnected. Run mcp_connect_server (or just call any tool on it — connection happens on demand) before invoking its tools.

Common pitfalls

• Passing tool_params to mcp_call_tool as a Python dict. It MUST be a JSON object string (e.g. '{"key": "value"}'); pass '{}' when no parameters are needed.
• Calling mcp_call_tool without first checking mcp_get_tool_schema and supplying random parameter names. Always look at the schema first.
• Passing args to mcp_add_server as a single string like '-y @pkg /tmp'. It MUST be a list of strings.
• Treating mcp_list_tools without server_name as a way to "discover" tools on disconnected servers. It only returns tools from already-connected servers; use mcp_list_servers + mcp_connect_server to see disconnected ones.
• Trying to call mcp_call_tool directly as a tool call (without run_sdk_snippet). It will be rejected because these tools are Code Mode only.
• Passing output_file_path as a relative path like 'reports/data.json'. It is rejected and the call silently falls back to the runtime directory, so the user never sees the file. Always build an absolute path, and prefer placing it under the current workspace.