llamaR

A CLI-first AI coding agent, written in R.

Built on the Model Context Protocol (MCP). Connects to LLMs (Anthropic, OpenAI, Ollama) and gives them tools to act.

The llamar CLI is the human-facing agent. The llamaR package is the engine and MCP server.

---

In Spanish, llamar (pronounced “Ya Mar”) means “to call.”

---

Why R?

The same power as moltbot or Claude Code, in a language you already know.

• Readable — R users can inspect exactly what the agent does
• Extensible — add tools by writing R functions
• Tinyverse — minimize dependencies, maximize stability
• Isomorphic — skills work in llamaR and openclaw

---

What is llamaR?

1. llamar (CLI agent) An interactive terminal agent that connects to LLMs and uses tools to act.
2. MCP server A standalone server exposing tools for any MCP client (Claude Desktop, Claude Code, etc.).

It can run:

• as a CLI agent for humans
• as a tool provider for other MCP clients
• as an MCP client connecting to other servers

---

Relationship to mcptools

llamaR draws inspiration from mcptools.

The ecosystems have parallel structure:

Role	Posit (tidyverse)	cornyverse
LLM API client	ellmer	llm.api
Context tools	btw	fyi
MCP bridge	mcptools	llamaR

Different philosophies:

mcptools integrates R into the broader MCP ecosystem—Claude Desktop, VS Code Copilot, Positron, shiny apps via ellmer. It’s polished, on CRAN, and backed by Posit.

llamaR is a standalone agent runtime. It ships its own CLI, handles LLM connections internally, and doesn’t require external clients.

If you want to…	Consider
Use R from Claude Desktop / VS Code / Positron	mcptools
Run an AI agent from your terminal	llamaR

AI and CLI coding agents enable bespoke software. This is our take.

---

Tools

llamaR exposes tools to LLMs:

Tool	Description
bash	Run shell commands
read_file	Read file contents
write_file	Write or create files
edit_file	Surgical file edits
list_files	List directory contents
git	Status, diff, commit, log
run_r	Execute R code in a persistent session
r_help	Query R documentation via fyi
installed_packages	List and search R packages
web_search	Search the web via Tavily

Plus any tools from connected MCP servers.

Web Search

Web search is powered by Tavily, an AI-native search engine with a free tier. Set TAVILY_API_KEY in ~/.Renviron to enable.

---

Skills

Skills are portable agent extensions—Markdown files that teach the LLM how to use shell commands.

---
name: weather
description: Get current weather (no API key required)
---

# Weather

```bash
curl -s "wttr.in/London?format=3"

Drop a `SKILL.md` file in `~/.llamar/skills/weather/` and it's immediately available.

### Isomorphic with openclaw

llamaR uses the same skill format as [openclaw](https://github.com/mariozechner/openclaw). Skills are portable between R (llamaR) and TypeScript (openclaw) agent runtimes.

```bash
# Use openclaw skills in llamaR
ln -s ~/.openclaw/skills ~/.llamar/skills

Why this matters:

• Write once, run anywhere — same skill works in both systems
• No lock-in — skills are Markdown, not code
• Ecosystem access — R users get the openclaw skill library
• Community leverage — contributions benefit both communities

Tested openclaw skills:

Skill	Description
github	GitHub CLI (gh) for issues, PRs, runs
weather	Weather via wttr.in
tmux	Remote-control tmux sessions

See docs/isomorphism.md for our full interoperability approach.

R Skills

R runs from shell via littler (r) or Rscript:

---
name: r-eval
description: Execute R code
---

# R Evaluation

```bash
r -e 'summary(lm(mpg ~ wt, mtcars))'

Shell-based R is **stateless**—each call is a fresh session. For **stateful** R (persistent variables across calls), use llamaR's built-in `run_r` tool. openclaw users can access stateful R by connecting to llamaR as an MCP server.

See [docs/skills.md](docs/skills.md) for the full specification.

-----

## Architecture

### As a CLI Agent

```bash
llamar

> summarize this CSV
> fix the bug in app.R
> commit with a good message
> refactor this into smaller functions

As an MCP Server

For Claude Desktop:

{
  "mcpServers": {
    "llamaR": {
      "command": "Rscript",
      "args": ["-e", "llamaR::serve()"]
    }
  }
}

As an MCP Client

web_tools <- mcp_connect(port = 7851)
db_tools  <- mcp_connect(port = 7852)

all_tools <- c(
  mcp_tools_for_api(web_tools),
  mcp_tools_for_api(db_tools)
)

Chain tools from multiple servers.

---

Installation

System Requirements

• R (>= 4.4.0)
• littler — fast R scripting frontend (recommended)

# Ubuntu/Debian
sudo apt install littler

# Or from R
install.packages("littler")

R Package

# Install llamaR (not yet on CRAN)
remotes::install_github("cornball-ai/llamaR")

# Install the CLI to ~/bin
llamaR::install_cli()

# Add ~/bin to PATH if needed
export PATH="$HOME/bin:$PATH"

Required R Packages

# Core dependencies (on CRAN)
install.packages(c("curl", "jsonlite"))

# LLM provider abstraction (not on CRAN)
remotes::install_github("cornball-ai/llm.api")

API Keys

Set in ~/.Renviron:

ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
TAVILY_API_KEY=tvly-...   # Optional, for web search

Voice Mode (Optional)

Voice mode requires additional packages and services:

# R packages
remotes::install_github("cornball-ai/stt.api")  # Speech-to-text
remotes::install_github("cornball-ai/tts.api")  # Text-to-speech

# System packages (Ubuntu)
sudo apt install sox alsa-utils ffmpeg

Voice services must be running:

• STT: whisper or compatible API on port 4123
• TTS: qwen3-tts-api or compatible API on port 7812

Memory Index (Optional)

For searchable conversation history:

install.packages("duckdb")
install.packages("digest")

# Import Claude Code history
llamaR::memory_import_claude()

# Search
llamaR::memory_search_fts("your query")

---

Quick Start

llamar

Or from R:

llamaR::run()

---

Platform Support

Platform	Status
Linux	Fully supported
macOS	Expected to work
Windows	Partial (stdin handling pending)

---

Design Philosophy

• CLI-first
• tinyverse: minimal dependencies
• model-agnostic (Ollama, Anthropic, OpenAI)
• MCP-native
• composable
• hackable

---

Status

Experimental. Interfaces may change.

---

License

MIT