rlmy 0.1.4

An interactive RLM (Recursive Language Model) agent built on DSPy

rlmy is pronounced "ar-leh-mee" — like the letters R-L-M out loud ("ar-el-em"), but let them slur together. Think "RLM-y."

What is this?

• An AI coding agent that runs in your terminal
• Uses DSPy's RLM framework — the LLM writes and executes Python code iteratively
• Connects to MCP servers (Slack, internal tools, etc.) as additional tools
• Maintains conversation state across turns (trajectory persistence, including REPL variables)
• Supports cooperative interrupt (Ctrl+C pauses gracefully)

Key Features

• Iterative REPL: LLM writes code, sees output, writes more code — until it solves the problem
• MCP Integration: Connect any MCP-compatible tool server
• Filesystem Tools: Read, write, edit files with safety guards (read-before-write)
• Shell Access: Run shell commands with deny-list safety and approval system
• Trajectory Persistence: Resume sessions where you left off
• Conversation Continuity: Prior context injected into new turns
• Configurable Models: Use any DSPy-compatible LM (Anthropic, Bedrock, OpenAI, Groq, Ollama)

Prerequisites

• Valid LLM credentials (Anthropic API key, AWS profile for Bedrock, etc.)
• Any of these work thanks to LiteLLM:

export OPENAI_API_KEY="sk-proj-..."

export GROQ_API_KEY="gsk_..."

export GEMINI_API_KEY="AQ.Ab8R..."

export AWS_ACCESS_KEY="..."
export AWS_SECRET="..."

Installation

Recommended (installs uv + rlmy in one command):

curl -LsSf https://raw.githubusercontent.com/diego-lima/rlmy/main/setup_install.sh | bash

Alternative (if you already have uv):

uv tool install rlmy

Quick Start

rlmy

• First run asks which AI model to use (model selection wizard)
• Workspaces are created in ~/.config/rlmy/sandboxes/
• Ctrl+C pauses gracefully (doesn't lose work)

To skip the wizard:

export RLM_MAIN_MODEL='bedrock/us.anthropic.claude-sonnet-4-6'
export RLM_SUB_MODEL='bedrock/us.anthropic.claude-sonnet-4-6'
rlmy

Try It Out

If your credentials are set, run rlmy and type:

curl https://calmcode.io/static/data/pokemon.json straight into a variable and teach me something about it.

Watch it fetch the data, explore it with code, and teach you something you didn't know. This is the RLM loop in action: it'll iterate until it has a neat insight.

Configuration

• Priority: env vars > config file > wizard
• Config file: ~/.config/rlmy/config.toml
• Supported model formats: any DSPy model string (e.g., bedrock/us.anthropic.claude-sonnet-4-6, bedrock/us.anthropic.claude-opus-4-6-v1)

MCP Tools (optional)

• Config location: ~/.config/rlmy/mcp_servers.json (refer to mcp_servers.example.json)
• The setup script creates an empty template
• Edit it to connect Slack, internal tools, or any MCP-compatible server
• Agent starts without MCP if config is empty (no crash)

CLI Options

• --sandbox-root PATH: Override sandbox directory (default: ~/.config/rlmy/sandboxes/)
• --cache-path PATH: Override workspace cache file

Architecture

• Built on DSPy's experimental RLM module
• InterruptableRLM: cooperative SIGINT + trajectory injection
• Sandboxed code execution via Deno + Pyodide (WASM)
• Tools are plain Python functions registered with DSPy

License

MIT

Status

• Early release — works well for the author, may have rough edges
• Feedback welcome via GitHub issues

Disclaimer

This is an experimental tool that I vibe-coded over time for my own needs.