RLM

You run a Recursive Language Models MCP server that keeps large data out of the LLM’s working context, letting you analyze files and logs well beyond normal context limits while maintaining a small, stable context for the language model.

How to use

You connect Claude Code to the MCP server and use a set of dedicated tools to load data, run Python code, and query a larger language model in sub-calls. The server keeps data in memory outside the main context, so you can perform big-file analyses, data aggregation, and cross-file searches without flooding the LLM’s input.

Typical usage patterns include loading data into a named variable, performing computations or filtering with Python, and then producing a concise summary or report. You can chain sub-LLM calls to summarize chunks of data and then synthesize a final result.”,

How to install

Prerequisites: you need Docker and Python available on your setup. You also need a client environment for your MCP integration. Follow these steps to get the server running and ready to connect from your Claude Code environment.

1. On the server side, clone the project, set up environment variables, and start the service.

# Clone the repository
git clone https://github.com/seu-usuario/rlm-mcp-server.git
cd rlm-mcp-server

# Copy and edit environment variables
cp .env.example .env
# Edit .env with your configurations

# Create data directory and store your files
mkdir -p data
# Place your files in ./data/

# Build and start the services
docker-compose up -d --build

On the Claude Code client

Choose a connection mode to the server. The recommended approach is an SSH tunnel from your Claude Code host to the MCP server port 8765, or connect through a public MCP URL if you expose the service.

Option A: SSH Tunnel (recommended) to forward the server port to your local environment.

# Create an SSH tunnel to the server port 8765
ssh -L 8765:localhost:8765 root@seu-servidor.com -N &

# The tunnel runs in the background

Start the MCP locally

If you want to run the MCP server locally for testing without a remote host, you can run the binary directly (the server supports a local runtime). The following starts the MCP server in your environment.

cd rlm-mcp-server
pip install -e .

# Test run locally
python -c "from rlm_mcp.server import main; main()"

Available tools

rlm_load_data

Loads data directly into a named variable for subsequent processing

rlm_load_file

Loads a file from the server into memory, supporting text, json, csv, pdf, and OCR-enabled pdfs

rlm_execute

Executes Python code that runs against the loaded data and in-repo helpers

rlm_list_vars

Lists currently available variables in memory

rlm_var_info

Provides detailed information about a specific variable

rlm_clear

Clears variables from memory to free space

rlm_memory

Reports memory usage statistics for managed data

rlm_load_s3

Loads a file from Minio/S3 into memory, including support for pdf and pdf_ocr

rlm_list_buckets

Lists buckets on the Minio/S3 endpoint

rlm_list_s3

Lists objects within a specific bucket on Minio/S3

rlm_upload_url

Generates a signed URL for uploading data

llm_query

Performs a sub-call to the LLM with optional data and model

llm_stats

Returns usage statistics for LLM interactions

llm_reset_counter

Resets the sub-call counter for usage tracking