An R package for interacting with Large Language Models (LLMs) with support for agents, workflows, and MCPs: Get Started

Overview

llmr provides a streamlined interface to LLM providers like Anthropic’s Claude and OpenAI. The package allows you to:

• Create and manage conversation history with LLMs
• Define agents with custom tools that LLMs can use
• Handle structured responses and tool calls
• Integrate with external tools via the mcpr package

Installation

pak::pak("devOpifex/llmr")

Basic Usage

library(llmr)

# Create an agent with ellmer chat object directly (recommended)
agent <- new_agent("assistant", ellmer::chat_anthropic)

# Send request and get response
request(agent, new_message("Explain quantum computing in simple terms"))

# Print the last message
get_last_message(agent)

Creating a Simple Agent

Note that we leverage the tooling from the mcpr package to create the tools, though you may also use the tools from ellmer. This allows seamless integration with MCP (Model Context Protocol) servers.

library(llmr)

# Create a simple agent with ellmer chat object
agent <- new_agent("calculator", ellmer::chat_anthropic)

# Add a calculator tool to the agent
add_tool(
  agent,
  mcpr::new_tool(
    name = "weather",
    description = "Get the weather forecast for a given location",
    input_schema = mcpr::schema(
      properties = mcpr::properties(
        location = mcpr::property_string(
          title = "Location",
          description = "The location for which you want the weather forecast",
          required = TRUE
        )
      )
    ),
    handler = function(params) {
      sprintf("The weather forecast for %s is dark and rainy", params$location)
    }
  )
)

# Make a request to the LLM
request(
  agent,
  new_message("What is the weather in Switzerland?")
)

# Print the response
get_last_message(agent)

Human-in-the-Loop Tool Approval

For sensitive operations, you can require human approval before tools are executed:

library(llmr)

# Create an agent with human approval enabled from the start
agent <- new_agent(
  "assistant", 
  ellmer::chat_anthropic, 
  approval_callback = prompt_human_approval
)

add_tool(
  agent,
  mcpr::new_tool(
    name = "delete_file",
    description = "Delete a file from the system",
    input_schema = mcpr::schema(
      properties = mcpr::properties(
        filepath = mcpr::property_string(
          title = "File Path",
          description = "Path to the file to delete",
          required = TRUE
        )
      )
    ),
    handler = function(params) {
      file.remove(params$filepath)
      sprintf("Deleted file: %s", params$filepath)
    }
  )
)

request(agent, new_message("Please delete the file /tmp/example.txt"))

MCP Integration

See mcpr for how to create and use MCP servers (and clients).

library(llmr)
library(mcpr)

# Create an agent with ellmer chat object
agent <- new_agent("mcp", ellmer::chat_anthropic)

# Create an MCP client that connects to an external calculator service
client <- mcpr::new_client(
  command = "Rscript",
  args = "/path/to/server.R",
  name = "calculator"
)

# Register the MCP with the provider
register_mcp(agent, client)

# Make a request
request(
  agent, 
  new_message("Subtract 5 from 10.")
)

# Print the response
get_last_message(agent)

Workflows

Chain together processing steps and agents into workflows:

library(llmr)

# Define processing functions
add_ten <- function(x) x + 10
multiply_two <- function(x) x * 2

# Create workflow steps
step1 <- step(add_ten)
step2 <- step(multiply_two)

# Chain steps together
workflow <- step1 %->% step2

# Execute workflow
result <- execute(workflow, 5)  # Returns 30: (5 + 10) * 2