mcp-cli（MCP CLI 搜索）

MCP CLI: On-Demand MCP Server Usage

Use the mcp CLI tool to dynamically discover and invoke MCP server capabilities without pre-configuring them as permanent integrations.

When to Use This Skill

Use this skill when you need to:

• Explore an MCP server's capabilities before deciding to use it
• Make one-off calls to an MCP server without permanent integration
• Access MCP functionality without polluting the context window
• Test or debug MCP servers
• Use MCP servers that aren't pre-configured

Prerequisites

The mcp CLI must be installed at ~/.local/bin/mcp. If not present:

# Clone and build
cd /tmp && git clone --depth 1 https://github.com/f/mcptools.git
cd mcptools && CGO_ENABLED=0 go build -o ~/.local/bin/mcp ./cmd/mcptools

Always ensure PATH includes the binary:

export PATH="$HOME/.local/bin:$PATH"

Discovery Workflow

Step 1: Discover Available Tools

mcp tools <server-command>

Examples:

# Filesystem server
mcp tools npx -y @modelcontextprotocol/server-filesystem /path/to/allow

# Memory/knowledge graph server
mcp tools npx -y @modelcontextprotocol/server-memory

# GitHub server (requires token)
mcp tools docker run -i --rm -e GITHUB_PERSONAL_ACCESS_TOKEN ghcr.io/github/github-mcp-server

# HTTP-based server
mcp tools https://example.com/mcp

Step 2: Discover Resources (if supported)

mcp resources <server-command>

Resources are data sources the server exposes (files, database entries, etc.).

Step 3: Discover Prompts (if supported)

mcp prompts <server-command>

Prompts are pre-defined prompt templates the server provides.

Step 4: Get Detailed Info (JSON format)

# For full schema details including parameter types
mcp tools --format json <server-command>
mcp tools --format pretty <server-command>

Making Tool Calls

Basic Syntax

mcp call <tool_name> --params '<json>' <server-command>

Examples

Read a file:

mcp call read_file --params '{"path": "/tmp/example.txt"}' \
  npx -y @modelcontextprotocol/server-filesystem /tmp

Write a file:

mcp call write_file --params '{"path": "/tmp/test.txt", "content": "Hello world"}' \
  npx -y @modelcontextprotocol/server-filesystem /tmp

List directory:

mcp call list_directory --params '{"path": "/tmp"}' \
  npx -y @modelcontextprotocol/server-filesystem /tmp

Create entities (memory server):

mcp call create_entities --params '{"entities": [{"name": "Project", "entityType": "Software", "observations": ["Uses TypeScript"]}]}' \
  npx -y @modelcontextprotocol/server-memory

Search (memory server):

mcp call search_nodes --params '{"query": "TypeScript"}' \
  npx -y @modelcontextprotocol/server-memory

Complex Parameters

For nested objects and arrays, ensure valid JSON:

mcp call edit_file --params '{
  "path": "/tmp/file.txt",
  "edits": [
    {"oldText": "foo", "newText": "bar"},
    {"oldText": "baz", "newText": "qux"}
  ]
}' npx -y @modelcontextprotocol/server-filesystem /tmp

Output Formats

# Table (default, human-readable)
mcp call <tool> --params '{}' <server>

# JSON (for parsing)
mcp call <tool> --params '{}' -f json <server>

# Pretty JSON (readable JSON)
mcp call <tool> --params '{}' -f pretty <server>

Reading Resources

# List available resources
mcp resources <server-command>

# Read a specific resource
mcp read-resource <resource-uri> <server-command>

# Alternative syntax
mcp call resource:<resource-uri> <server-command>

Using Prompts

# List available prompts
mcp prompts <server-command>

# Get a prompt (may require arguments)
mcp get-prompt <prompt-name> <server-command>

# With parameters
mcp get-prompt <prompt-name> --params '{"arg": "value"}' <server-command>

Server Aliases (for repeated use)

If using a server frequently during a session:

# Create alias
mcp alias add fs npx -y @modelcontextprotocol/server-filesystem /home/user

# Use alias
mcp tools fs
mcp call read_file --params '{"path": "README.md"}' fs

# List aliases
mcp alias list

# Remove when done
mcp alias remove fs

Aliases are stored in ~/.mcpt/aliases.json.

Authentication

HTTP Basic Auth

mcp tools --auth-user "username:password" https://api.example.com/mcp

Bearer Token

mcp tools --auth-header "Bearer your-token-here" https://api.example.com/mcp

Environment Variables (for Docker-based servers)

mcp tools docker run -i --rm \
  -e GITHUB_PERSONAL_ACCESS_TOKEN="$GITHUB_TOKEN" \
  ghcr.io/github/github-mcp-server

Transport Types

Stdio (default for npx/node commands)

mcp tools npx -y @modelcontextprotocol/server-filesystem /tmp

HTTP (auto-detected for http/https URLs)

mcp tools https://example.com/mcp

SSE (Server-Sent Events)

mcp tools http://localhost:3001/sse
# Or explicitly:
mcp tools --transport sse http://localhost:3001

Common MCP Servers

Filesystem

# Allow access to specific directory
mcp tools npx -y @modelcontextprotocol/server-filesystem /path/to/allow

Memory (Knowledge Graph)

mcp tools npx -y @modelcontextprotocol/server-memory

GitHub

export GITHUB_PERSONAL_ACCESS_TOKEN="your-token"
mcp tools docker run -i --rm -e GITHUB_PERSONAL_ACCESS_TOKEN ghcr.io/github/github-mcp-server

Brave Search

export BRAVE_API_KEY="your-key"
mcp tools npx -y @anthropic/mcp-server-brave-search

Puppeteer (Browser Automation)

mcp tools npx -y @anthropic/mcp-server-puppeteer

Best Practices

1. Always Discover First

Before calling tools, run mcp tools to understand what's available and the exact parameter schema.

2. Use JSON Format for Parsing

When you need to process results programmatically:

mcp call <tool> --params '{}' -f json <server> | jq '.field'

3. Validate Parameters

The table output shows parameter signatures. Match them exactly:

• param:str = string
• param:num = number
• param:bool = boolean
• param:str[] = array of strings
• [param:str] = optional parameter

4. Handle Errors Gracefully

Tool calls may fail. Check exit codes and stderr:

if ! result=$(mcp call tool --params '{}' server 2>&1); then
  echo "Error: $result"
fi

5. Use Aliases for Multi-Step Operations

If making several calls to the same server:

mcp alias add tmp-server npx -y @modelcontextprotocol/server-filesystem /tmp
mcp call list_directory --params '{"path": "/tmp"}' tmp-server
mcp call read_file --params '{"path": "/tmp/file.txt"}' tmp-server
mcp alias remove tmp-server

6. Restrict Capabilities with Guard

For safety, limit what tools are accessible:

# Only allow read operations
mcp guard --allow 'tools:read_*,list_*' --deny 'tools:write_*,delete_*' \
  npx -y @modelcontextprotocol/server-filesystem /home

Debugging

View Server Logs

mcp tools --server-logs <server-command>

Check Alias Configuration

cat ~/.mcpt/aliases.json

Verbose Output

Use --format pretty for detailed JSON output to debug parameter issues.

Quick Reference

Example: Complete Workflow

# 1. Discover what's available
mcp tools npx -y @modelcontextprotocol/server-filesystem /home/user/project

# 2. Check for resources
mcp resources npx -y @modelcontextprotocol/server-filesystem /home/user/project

# 3. Create alias for convenience
mcp alias add proj npx -y @modelcontextprotocol/server-filesystem /home/user/project

# 4. Explore directory structure
mcp call directory_tree --params '{"path": "/home/user/project"}' proj

# 5. Read specific files
mcp call read_file --params '{"path": "/home/user/project/README.md"}' proj

# 6. Search for patterns
mcp call search_files --params '{"path": "/home/user/project", "pattern": "**/*.ts"}' proj

# 7. Clean up alias
mcp alias remove proj

Troubleshooting

"command not found: mcp"

Ensure PATH is set: export PATH="$HOME/.local/bin:$PATH"

JSON parse errors

• Escape special characters properly
• Avoid shell expansion issues by using single quotes around JSON
• For complex JSON, write to a temp file and use --params "$(cat params.json)"

Server timeout

Some servers take time to start. The mcp CLI waits for initialization automatically.

Permission denied

For filesystem server, ensure the allowed directory path is correct and accessible.