r-ai

R AI Ecosystem (Meta-Skill)

Overview

This is a meta-skill that helps you choose the right R AI package. The R AI stack consists of 5 specialized packages, each with its own skill.

Install all: install.packages(c("ellmer", "btw", "mcptools", "ragnar", "vitals"))

API keys: Set OPENAI_API_KEY or ANTHROPIC_API_KEY in .Renviron

References

• references/integration-patterns.md - How packages work together in workflows

For per-package API docs, invoke the corresponding package skill (/r-ellmer, /r-btw, /r-mcptools, /r-ragnar, /r-vitals).

When to Use Which Package

Package	Use When	Skill
ellmer	Chat with any LLM from R	/r-ellmer
btw	Provide R context (docs, data) to LLMs	/r-btw
mcptools	Let agents (Claude Code) run R code	/r-mcptools
ragnar	LLM searches your documents (RAG)	/r-ragnar
vitals	Test LLM output quality	/r-vitals

Navigation: When you know which package you need, invoke the specific skill above.

When NOT to Use

• Single one-off API call (use httr2)
• Non-chat ML workflows (use tidymodels)
• Production REST APIs (use plumber)

Key Distinctions

btw vs mcptools:

• btw = Give tools TO ellmer (R → LLM): Chat can read R docs/data
• mcptools = Let agents INTO R (LLM → R): Agents can run R code

ellmer vs ragnar:

• ellmer = Direct chat, limited by context window
• ragnar = Chat searches documents first (RAG)

Common Integration Patterns

Chat + Context Tools

library(ellmer)
library(btw)
chat <- chat_openai()
chat$set_tools(btw_tools())

Chat + RAG

library(ellmer)
library(ragnar)
chat <- chat_openai()
ragnar_register_tool_retrieve(chat, store)

Agent Access to R

library(mcptools)
mcp_session()  # In console, each startup
# Now Claude Code can run R code

Fully Local (Ollama)

library(ellmer)
chat <- chat_ollama(model = "llama3.2")

library(ragnar)
store <- ragnar_store_create("local.duckdb",
  embed = ragnar_embed_ollama(model = "nomic-embed-text"))

RAG Evaluation

library(vitals)
library(ellmer)
library(ragnar)

chat <- chat_openai()
ragnar_register_tool_retrieve(chat, store)

task <- Task$new(
  dataset = test_cases,
  solver = function(input) chat$chat(input, echo = "none"),
  scorer = model_graded_qa()
)
task$run()

Advanced

See references/ for:

• integration-patterns.md: Full cross-package integration examples

For package-specific documentation, invoke the individual package skills above.