fleet-rlm 0.4.0

Recursive Language Models with DSPy + Modal for secure long-context code execution

fleet-rlm

Secure, cloud-sandboxed Recursive Language Models (RLM) with DSPy and Modal.

Allow your LLMs to write code that explores massive datasets or long documents in the cloud, without downloading them locally.

Documentation | Paper | Contributing

---

graph TD
    User[User/Agent] -->|Question| CLI[fleet-rlm CLI]
    CLI -->|Plan| DSPy[DSPy Planner]
    DSPy -->|Generate Code| Modal[Modal Sandbox]
    Modal -->|Execute safely| Cloud[Cloud Environment]
    Cloud -->|Result| Modal
    Modal -->|Answer| User

    style Modal fill:#f9f,stroke:#333,stroke-width:2px
    style DSPy fill:#bbf,stroke:#333,stroke-width:2px

What is this?

fleet-rlm gives your AI agent a secure "computer" in the cloud. Instead of trying to shove 10,000 pages of text into a prompt, the agent writes Python code to:

1. Search and filter data in a remote sandbox (Modal).
2. Read only what matters.
3. Synthesize the answer.

This approach, called Recursive Language Modeling, mimics how humans solve research tasks: we don't memorize the library; we look things up.

Quick Start: Claude Code Integration

1. Install & Initialize

Install the package and register the RLM skills with your local Claude Code agent (~/.claude/).

# Install fleet-rlm
uv pip install fleet-rlm

# Install skills, agents, and prompts to ~/.claude
uv run fleet-rlm init

2. Configure Cloud Runtime

Authenticate with Modal to enable the sandboxed execution environment.

uv run modal setup
uv run modal secret create LITELLM DSPY_LM_MODEL=openai/gpt-4o DSPY_LLM_API_KEY=sk-...

3. Use with Claude

Now your Claude Code agent has "superpowers". You can ask it to perform deep research tasks that require running code.

Example Prompts:

"Use the rlm skill to analyze the latest papers on linear attention mechanisms." "Run the rlm-batch agent to parallelize data extraction for these 50 files."

Available Skills:

• rlm - Core recursive research capability.
• rlm-batch - Parallel processing.
• rlm-memory - Persistent storage.

Standalone Usage

You can also run fleet-rlm directly without Claude Code:

Interactive Chat (TUI) Chat with the RLM agent in your terminal using the OpenTUI interface.

uv run fleet-rlm code-chat --opentui

API Server Start a FastAPI server to expose RLM capabilities over HTTP.

# Dev server with hot reload
uv run fastapi dev src/fleet_rlm/server/main.py

# Production server via CLI
uv run fleet-rlm serve-api

API docs are available at /docs (Swagger) and /scalar (Scalar).

Features

• 🔒 Sandboxed Execution: Code runs in isolated Modal containers, not on your laptop.
• 🧠 DSPy Powered: Uses advanced prompt engineering pipelines for reliable code generation.
• 💬 Interactive TUI: Chat with the agent in your terminal (fleet-rlm code-chat).
• ⚡ Production Ready: Includes a fastapi server and MCP integration for Claude Desktop.

Documentation

• Tutorials: Step-by-step lessons.
• Installation Guide: Detailed setup instructions.
• Skills & Agents: Enhance Claude with RLM capabilities.

Contributing

We welcome contributions! Whether it's reporting a bug, suggesting a feature, or writing code, your input is verified.

1. Check out our Contribution Guide.
2. Fork the repo and create a branch.
3. Run tests with uv run pytest.
4. Submit a Pull Request.

Acknowledgments

This project is built upon the innovative research by Alex L. Zhang (MIT CSAIL), Omar Khattab (Stanford), and Tim Kraska (MIT).

Reference: Recursive Language Models (Zhang, Kraska, Khattab, 2025)

License

This project is licensed under the MIT License - see the LICENSE file for details.