RLM - Recursive Language Model

RLM - Recursive Language Model

Run the bundled RLM tool to analyze a document by recursively exploring it with LLM-generated Python code in a REPL.

How to run

The RLM Python package is bundled in this plugin. Run it with uv run:

uv run --directory "${CLAUDE_PLUGIN_ROOT}" rlm --context-file <path> --query "<question>" --verbose

This requires no pre-installation — uv resolves dependencies and executes from the bundled source automatically.

Arguments

Parse $ARGUMENTS to determine the context file and query:

• First argument ($0): path to the context file
• Remaining arguments: the query (join them as a single string)

If only one argument is provided, treat it as the context file and use a default query. If no arguments are provided, ask the user which file to analyze and what to ask about it.

Backend options

• Anthropic (default): Requires ANTHROPIC_API_KEY env var to be set.

  uv run --directory "${CLAUDE_PLUGIN_ROOT}" rlm --context-file <path> --query "<question>" --verbose
  

• OpenAI: Requires OPENAI_API_KEY env var to be set.

  uv run --directory "${CLAUDE_PLUGIN_ROOT}" --with openai rlm --backend openai --context-file <path> --query "<question>" --verbose
  

• OpenRouter: Requires OPENROUTER_API_KEY env var to be set.

  uv run --directory "${CLAUDE_PLUGIN_ROOT}" --with openai rlm --backend openrouter --context-file <path> --query "<question>" --verbose
  

• Hugging Face: Requires HF_TOKEN env var to be set.

  uv run --directory "${CLAUDE_PLUGIN_ROOT}" --with openai rlm --backend huggingface --context-file <path> --query "<question>" --verbose
  

• Ollama (local models): Requires Ollama running locally.

  uv run --directory "${CLAUDE_PLUGIN_ROOT}" --with openai rlm --backend ollama --model llama3.2 --context-file <path> --query "<question>" --verbose
  

Additional flags

• --model <name>: Override the LLM model (default: claude-sonnet-4-20250514)
• --sub-rlm-model <name>: Use a different (cheaper) model for sub-RLM calls
• --compact: Use a shorter system prompt
• --max-iterations <n>: Limit REPL iteration count (default: 10)

Steps

1. Confirm the context file exists using the Read or Glob tool
2. Determine the appropriate backend based on available environment (check if ANTHROPIC_API_KEY is set)
3. Run the uv run --directory "${CLAUDE_PLUGIN_ROOT}" rlm command with the appropriate arguments
4. Present the final answer and statistics to the user
5. If the command fails, check the error output and suggest fixes (missing API key, file not found, etc.)