rlm

What This Skill Does

The RLM (Recursive Language Models) skill transforms OpenClaw into a deterministic computing environment. Unlike standard LLM interactions that rely on probabilistic text generation—which often leads to "hallucinations" in math or logic—RLM utilizes a recursive execution loop. It breaks down complex requests into subtasks, writes Python code to solve those parts, executes the code, and verifies the output against the original objective. This process repeats until the model confirms the result is accurate. By bridging OpenClaw with the RLM MCP server, you turn your agent into a rigorous research and analysis tool.

Installation

Installation requires setting up the mcporter bridge and the RLM server environment. First, ensure you have npm installed to run npm install -g mcporter. Next, clone the RLM repository to your local machine and create a dedicated directory in your home path (e.g., ~/.claude/mcp-servers/rlm). Follow the setup scripts provided to establish a Python virtual environment and configure the server entry point. Finally, register the server in your ~/.mcp.json configuration file and export your OPENROUTER_API_KEY. Verification is performed using mcporter list and by calling rlm.rlm_status() to ensure the connection is active and authenticated.

Use Cases

RLM is ideal for high-stakes tasks where precision is non-negotiable. Use it for complex mathematical modeling, statistical analysis of large datasets, cryptographic verification, or decomposing multi-step architectural problems. It is particularly powerful for data processing tasks where an agent must clean, filter, and compute aggregate metrics from raw inputs without risk of arithmetic error.

Example Prompts

1. "Analyze the provided CSV file, calculate the moving average of the 'revenue' column over the last 30 days, and generate a summary report of trends."
2. "Find the numerical solution to this complex differential equation and verify the answer by back-substituting into the original expression using Python code."
3. "Decompose this project architecture request into 5 specific technical subtasks and execute the code necessary to validate the feasibility of each step."

Tips & Limitations

RLM excels at logic and calculation but can be slower than standard LLM responses due to the recursive verification loops. Ensure your OpenRouter API key has sufficient credits, as each iteration consumes tokens. If a task is hanging, check your RLM_MAX_ITERATIONS environment variable to increase or limit the depth of the search. Because RLM executes actual Python code on your machine, it should only be used with trusted scripts and in secure environments.