rlm

RLM (Recursive Language Model) is an MCP server that allows Claude Code to process massive contexts—up to 10M+ tokens—by loading files as external variables rather than injecting them directly into prompts. Instead of reading large files into context, Claude uses RLM tools behind the scenes to chunk content strategically, run sub-queries on sections, and aggregate results. This approach handles extensive logs, datasets, and codebases without context overflow. You simply ask Claude to analyze a large file, and RLM manages the decomposition automatically. It's designed for operations teams, data analysts, and developers who need to work with files too large for conventional LLM processing.

1. **Prepare Your Dataset**: Save your large file (e.g., JSON, CSV, or text) in a location accessible to Claude Code. Ensure it’s structured for analysis (e.g., rows/columns for tabular data). 2. **Load the Dataset**: Use the rlm command in Claude Code to treat the file as an external variable. Example: ``` rlm --file-path /path/to/large_dataset.json --context-size 10M ``` 3. **Define Your Analysis**: Craft a prompt that specifies the patterns or insights you need. Use placeholders like [SPECIFIC_PATTERNS] to guide the AI. Example: ``` Analyze the dataset for [ANOMALIES] in [TRANSACTION_COLUMN] and highlight [TRENDS] over [TIME_PERIOD]. ``` 4. **Review and Refine**: Examine the AI’s output for accuracy. If the results are too broad, refine your prompt by narrowing the focus (e.g., "Focus only on transactions >$500"). 5. **Export Insights**: Save the AI’s recommendations or summaries to a new file for further action. Use: ``` rlm --export /path/to/output.json ``` **Tip**: For best results, pre-filter your dataset to reduce noise before using rlm. Combine rlm with Claude Code’s built-in tools (e.g., `grep`, `jq`) for initial data cleaning.