Skip to main content

Model Context Protocol (MCP)

The Model Context Protocol (MCP) enables LLMs to interact with external tools and services through a standardized interface. lc supports MCP servers, allowing you to extend your LLM's capabilities with tools for web access, file manipulation, and more.

Overview​

MCP servers are external processes that provide tools (functions) that LLMs can call during conversations. These tools can:

  • Fetch content from the internet
  • Interact with local files
  • Execute system commands
  • Connect to APIs and databases
  • And much more

Managing MCP Servers​

Adding an MCP Server​

To add a new MCP server to your configuration:

lc mcp add <name> <command> --type <type> [-e KEY=VALUE]...

Parameters:

  • <name>: A unique identifier for the server
  • <command>: The command to start the server or URL for remote servers
  • --type: The server type (stdio, sse, or streamable)
  • -e, --env KEY=VALUE: Set environment variables (can be used multiple times)

Example:

# Add the fetch server for internet access
lc mcp add fetch "uvx mcp-server-fetch" --type stdio

# Add the Playwright server for browser automation
lc mcp add playwright "npx @playwright/mcp@latest" --type stdio

# Add server with environment variables
lc mcp add exa "npx -y exa-mcp-server" --type stdio -e EXA_API_KEY=your_api_key

# Multiple environment variables
lc mcp add custom "uvx custom-mcp" --type stdio -e API_KEY=key -e DEBUG=true

Listing MCP Servers​

View all configured MCP servers:

lc mcp list

Deleting an MCP Server​

Remove an MCP server from your configuration:

lc mcp delete <name>

Working with MCP Servers​

Viewing Available Functions​

To see what functions a server provides:

lc mcp functions <server-name>

Example:

lc mcp functions fetch
# Output:
# Functions: Available functions:
#   • fetch - Fetches a URL from the internet and optionally extracts its contents as markdown.
#     Parameters: max_length, raw, start_index, url

Invoking Functions Directly​

You can test MCP functions directly:

lc mcp invoke <server-name> <function-name> [parameters...]

Example:

# Fetch content from a URL
lc mcp invoke fetch fetch url=http://httpbin.org/json

# With multiple parameters
lc mcp invoke fetch fetch url=http://example.com max_length=1000

Managing Server Connections​

MCP servers are managed automatically by lc's daemon. They start when needed and persist for reuse. If you need to explicitly close a connection:

# Stop a running server connection
lc mcp stop <server-name>

Note: With the new SDK implementation, servers start automatically when you use them with the -t flag or invoke a function. The stop command is only needed if you want to explicitly close a persistent connection.

Using MCP Tools in Prompts​

The real power of MCP comes from using these tools in your LLM conversations. Use the -t or --tools flag to include MCP servers in your prompt:

Direct Prompts with Tools​

# Single tool
lc -t fetch "What's the current weather in San Francisco?"

# Multiple tools (comma-separated)
lc -t fetch,playwright "Navigate to example.com and fetch its content"

Interactive Chat with Tools​

# Start chat with MCP tools
lc chat -m claude-3-opus-20240229 -t fetch

# With multiple tools
lc chat -m gpt-4 -t fetch,playwright

1. mcp-server-fetch​

Provides internet access to fetch web content.

Installation:

lc mcp add fetch "uvx mcp-server-fetch" --type stdio

Usage Example:

lc -t fetch "Get the latest news headlines from a news website"

2. Playwright MCP​

Enables browser automation and web scraping.

Installation: STDIO version

lc mcp add playwright "npx @playwright/mcp@latest" --type stdio

sse version

lc mcp a playwright-sse http://localhost:8931/sse --type sse

streamable version

lc mcp add playwright-mcp http://localhost:8931/mcp --type streamable

Usage Example:

lc -t playwright "Go to https://news.ycombinator.com/ and get me the first post"

output:
The first post on Hacker News is titled **"Show HN: Draw a fish and watch it swim with the others"**. Here are the details:

- **Link**: [Draw a fish](https://drawafish.com)
- **Points**: 610 points
- **Submitted by**: hallak
- **Time ago**: 10 hours ago
- **Comments**: 180 comments

If you need more information or details about other posts, feel free to ask!

3. Context7 MCP​

Provides access to library documentation and code examples.

Installation:

lc mcp add context7 "npx -y @upstash/context7-mcp" --type stdio

Usage Example:

lc -t context7 "Show me React hooks documentation"
lc -t context7 "How to use Express.js middleware"

4. File System MCP​

Allows reading and writing local files.

Installation:

lc mcp add fs "uvx mcp-server-fs" --type stdio

5. Exa search MCP​

Access Exa search capabilities for advanced web searches.

Installation:

lc mcp add exa "npx -y exa-mcp-server" --type stdio -e "EXA_API_KEY=your_exa_api_key"

Usage Example:

lc -t exa "Search for the latest AI research papers"

output:
Thinking... [EXA-MCP-DEBUG] [web_search_exa-1754081391442-9zlic] [web_search_exa] Starting search for query: "latest AI research papers"
[EXA-MCP-DEBUG] [web_search_exa-1754081391442-9zlic] [web_search_exa] Sending request to Exa API
[EXA-MCP-DEBUG] [web_search_exa-1754081391442-9zlic] [web_search_exa] Received response from Exa API
[EXA-MCP-DEBUG] [web_search_exa-1754081391442-9zlic] [web_search_exa] Found 5 results
[EXA-MCP-DEBUG] [web_search_exa-1754081391442-9zlic] [web_search_exa] Successfully completed request
Here are some of the latest resources and research papers related to AI:

1. **[AI Research Papers - Arize AI](https://arize.com/ai-research-papers/)**
   - Explore the latest technical papers with the Arize Community, which includes trending AI research and resources to stay updated with breakthroughs in AI.

2. **[Publications - Google DeepMind](https://deepmind.google/research/publications/)**
   - Discover the latest AI breakthroughs and updates from Google DeepMind. You can find a selection of recent research on complex and interesting challenges in AI.

3. **[Noteworthy AI Research Papers of 2024 (Part One) - Ahead of AI](https://magazine.sebastianraschka.com/p/ai-research-papers-2024-part-1)**
   - This article highlights significant AI research papers from January to June 2024, focusing on various topics, particularly LLM research.

4. **[Where To Find The Latest AI Research? Top 7 Sources to Stay Updated](https://learnprompting.org/blog/resources_latest_research_papers?srsltid=AfmBOoqNahYsL1aeuJH3ZA15JHlm26BU1agUyhpYqpFNvFSUOm2zWM6Q)**
   - This article discusses different journals, conferences, platforms, and communities for finding the latest AI research papers, along with strategies for staying informed.

5. **[Artificial Intelligence - arXiv](https://arxiv.org/list/cs.AI/recent)**
   - Access the most recent AI research papers published on arXiv. This resource provides a comprehensive list of the latest entries in AI research.

Feel free to explore these links for more detailed insights into the latest advancements in AI research!

Configuration​

MCP server configurations are stored in ~/Library/Application Support/lc/mcp.toml (macOS) or the equivalent directory on other platforms.

Example configuration:

[servers.fetch]
name = "fetch"
server_type = "Stdio"
command_or_url = "uvx mcp-server-fetch"

[servers.fetch.env]

[servers.playwright]
name = "playwright"
server_type = "Stdio"
command_or_url = "npx @playwright/mcp@latest"

[servers.playwright.env]

[servers.kagi]
name = "kagi"
server_type = "Stdio"
command_or_url = "uvx kagimcp"

[servers.kagi.env]
KAGI_API_KEY = "your_api_key_here"

Environment Variables​

MCP servers can require environment variables for configuration, such as API keys or debug flags. These are stored securely in the configuration file and passed to the server when it starts.

Security Note: Be careful not to commit your mcp.toml file to version control if it contains sensitive API keys.

Best Practices​

  1. Tool Selection: Only include the tools you need for a specific task to avoid confusion
  2. Testing: Use lc mcp invoke to test tools before using them in prompts
  3. Error Handling: Some tools may fail due to network issues or permissions - always have a fallback plan
  4. Security: Be cautious with tools that can execute commands or access sensitive data

Troubleshooting​

Connection Issues​

If an MCP server fails to connect:

  1. Check if the command is installed (uvx, npx, etc.)
  2. Verify network connectivity for remote servers
  3. Check environment variables are set correctly if required
  4. Check server logs in ~/Library/Application Support/lc/

HTTPS/SSL Issues​

Some MCP servers may have issues with HTTPS connections. Try:

  • Using HTTP URLs when possible
  • Checking proxy settings
  • Updating the MCP server to the latest version

Server Not Responding​

If a server stops responding:

# Stop the server connection
lc mcp stop <server-name>

# The server will restart automatically when you use it again
lc mcp invoke <server-name> <function-name> <args>

Examples​

Fetching and Analyzing Web Content​

lc -t fetch "Analyze the content structure of https://example.com and summarize the main sections"

Research Assistant​

lc -t fetch "Research the latest developments in quantum computing and provide a summary with sources"

Multi-Tool Automation​

lc -t fetch,fs "Fetch the latest documentation from https://docs.example.com and save it to a local file"

Advanced Usage​

Custom MCP Servers​

You can create your own MCP servers. See the MCP SDK documentation for details.

Remote MCP Servers​

MCP supports Server-Sent Events (SSE) for remote servers:

lc mcp add remote-server "http://localhost:8080/sse" --type sse

This allows you to host MCP servers on remote machines or in the cloud.