RLM - Recursive Language Models

Execute tasks with verified code execution via mcporter MCP bridge.

RLM writes and executes Python code iteratively until it produces a verified answer. Unlike direct LLM responses, RLM computations are 100% accurate for calculations.

Prerequisites

1. Install mcporter (MCP bridge)

npm install -g mcporter

2. Install RLM MCP Server

Option A: Clone and setup (recommended)

# Clone RLM project
git clone https://github.com/alexzhang13/rlm.git $HOME/rlm
cd $HOME/rlm
pip install -e .

# Create MCP server directory
mkdir -p $HOME/.claude/mcp-servers/rlm/src

# Download MCP server files
curl -o $HOME/.claude/mcp-servers/rlm/src/server.py \
  https://raw.githubusercontent.com/eesb99/rlm-mcp/main/src/server.py
curl -o $HOME/.claude/mcp-servers/rlm/run_server.sh \
  https://raw.githubusercontent.com/eesb99/rlm-mcp/main/run_server.sh
curl -o $HOME/.claude/mcp-servers/rlm/setup.sh \
  https://raw.githubusercontent.com/eesb99/rlm-mcp/main/setup.sh
curl -o $HOME/.claude/mcp-servers/rlm/requirements.txt \
  https://raw.githubusercontent.com/eesb99/rlm-mcp/main/requirements.txt

# Setup venv and install dependencies
chmod +x $HOME/.claude/mcp-servers/rlm/*.sh
cd $HOME/.claude/mcp-servers/rlm
python3 -m venv venv
venv/bin/pip install -r requirements.txt

Option B: Manual setup

# Create server directory
mkdir -p $HOME/.claude/mcp-servers/rlm/src

# Create venv and install dependencies
cd $HOME/.claude/mcp-servers/rlm
python3 -m venv venv
venv/bin/pip install mcp litellm

# Create run_server.sh
cat > $HOME/.claude/mcp-servers/rlm/run_server.sh << 'EOF'
#!/bin/bash
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
cd "$SCRIPT_DIR"
export PYTHONPATH="$HOME/rlm:$PYTHONPATH"
export RLM_MODEL="${RLM_MODEL:-openrouter/x-ai/grok-code-fast-1}"
export RLM_SUBTASK_MODEL="${RLM_SUBTASK_MODEL:-openrouter/openai/gpt-4o-mini}"
export RLM_MAX_DEPTH="${RLM_MAX_DEPTH:-2}"
export RLM_MAX_ITERATIONS="${RLM_MAX_ITERATIONS:-20}"
exec "$SCRIPT_DIR/venv/bin/python" -m src.server
EOF
chmod +x $HOME/.claude/mcp-servers/rlm/run_server.sh

3. Configure MCP (for Claude Code)

Add to ~/.mcp.json (replace YOUR_HOME with your actual home path, e.g., /Users/john or /home/john):

{
  "mcpServers": {
    "rlm": {
      "command": "bash",
      "args": ["YOUR_HOME/.claude/mcp-servers/rlm/run_server.sh"]
    }
  }
}

Get your home path: echo $HOME

4. Set API Key

RLM requires an OpenRouter API key:

export OPENROUTER_API_KEY="your-key-here"

5. Verify Installation

# Check mcporter sees RLM
mcporter list | grep rlm

# Test RLM
mcporter call 'rlm.rlm_status()'

Available Tools

Quick Commands

Simple calculation:

mcporter call 'rlm.rlm_execute(task: "calculate 127 * 389")'

First N primes:

mcporter call 'rlm.rlm_execute(task: "calculate the first 100 prime numbers")'

Data analysis:

mcporter call 'rlm.rlm_analyze(data: "[23, 45, 67, 89, 12, 34]", question: "what is the mean, median, and standard deviation?")'

Generate code:

mcporter call 'rlm.rlm_code(description: "function to check if a number is prime")'

Complex task (decomposed):

mcporter call 'rlm.rlm_decompose(complex_task: "analyze a $500K portfolio with 60/30/10 allocation, calculate risk metrics and 10-year projection", num_subtasks: 5)'

Check status:

mcporter call 'rlm.rlm_status()'

When to Use RLM

Use RLM for:

• Mathematical calculations requiring precision
• Statistical analysis (mean, std dev, correlations)
• Financial calculations (compound interest, NPV, IRR)
• Algorithm execution (primes, sorting, searching)
• Data transformations and aggregations
• Code generation with verification

Don't use RLM for:

• Simple factual questions (use direct response)
• Creative writing or brainstorming
• Tasks requiring web search or real-time data
• Very simple calculations (2+2)

How It Works

1. You give RLM a task
2. RLM writes Python code to solve it
3. Code executes in sandbox
4. If not complete, RLM iterates
5. Returns verified final answer

Models used:

• Root: grok-code-fast-1 (fast code execution)
• Subtasks: gpt-4o-mini (cheap sub-queries)

Configuration

Environment variables:

Server location: $HOME/.claude/mcp-servers/rlm/

Troubleshooting

"Server offline" or "No module named 'mcp'":

# Reinstall dependencies
cd $HOME/.claude/mcp-servers/rlm
python3 -m venv venv
venv/bin/pip install mcp litellm

"mcporter: command not found":

npm install -g mcporter

"rlm not in mcporter list":

• Check $HOME/.mcp.json exists and has rlm config
• Verify run_server.sh is executable: chmod +x $HOME/.claude/mcp-servers/rlm/run_server.sh

Slow response:

• RLM executes real code, typically 10-30 seconds
• Complex tasks with decomposition take longer

References

• Paper: Recursive Language Models (Zhang, Kraska, Khattab 2025)
• RLM Library: github.com/alexzhang13/rlm
• MCP Server: github.com/eesb99/rlm-mcp
• MCP SDK: modelcontextprotocol.io
• mcporter: mcporter.dev