A private AI workspace that runs on your computer and opens in your browser.
Local RLM Cowork combines a GGUF model, llama.cpp, Recursive Language Models, and Open WebUI. Release bundles include the application runtime and model server, so normal installation does not require Python, Command Prompt, PowerShell, Terminal, pip, virtual environments, or API configuration.
You need only:
- A Local RLM Cowork release bundle for your operating system.
- A
.ggufmodel that fits your computer.
Download and fully extract one of these release files from the repository's Releases page:
local-rlm-cowork-windows-x64.ziplocal-rlm-cowork-macos-x64.ziplocal-rlm-cowork-linux-x64.tar.gz
The Continuous build prerelease is rebuilt automatically from the latest successful main commit. Numbered v* releases are permanent snapshots intended for normal distribution.
Do not run the launcher from inside the archive.
Double-click:
launchers\windows\Start Local RLM Cowork.vbs
Open:
launchers/macos/Local RLM Cowork.app
Development builds are not signed or notarized. macOS may require approval under System Settings → Privacy & Security.
Double-click:
launchers/linux/Start Local RLM Cowork.desktop
Some desktops require Allow Launching once.
Press Open Local RLM in the small launcher window. The bundled application opens the control screen at http://127.0.0.1:7860.
In One-click local stack:
- Enter or paste the path to your
.ggufmodel. - Leave the recommended settings unchanged for the first run.
- Press Start complete stack.
The release already contains a matching llama-server; its path is selected automatically and hidden from the normal setup form.
Your browser
→ Open WebUI workspace
→ Local RLM proxy
→ bundled llama-server
→ your GGUF model
Default local addresses:
Control screen: http://127.0.0.1:7860
AI workspace: http://127.0.0.1:3000
RLM proxy: http://127.0.0.1:8000
llama-server: http://127.0.0.1:8080/v1
Everything binds to this computer only by default.
Each platform archive contains:
- A relocatable CPython runtime
- All Python packages required by the proxy, Gradio, and Open WebUI
- A platform-matched llama.cpp release
- An offline wheelhouse for repair
- Native no-terminal launchers
- A manifest recording exact Python, llama.cpp, application, platform, and commit versions
The repository also publishes the proxy-only container image to GitHub Container Registry:
ghcr.io/rpatel622/recursive-llm-proxy:latest
The container package is for server and developer deployments. Desktop users should download the operating-system bundles from Releases instead.
Use Repair installation in the launcher to reinstall the bundled application from the local wheelhouse without downloading dependencies. Repair preserves models, Open WebUI data, conversations, knowledge collections, and logs.
Release bundles are immutable. To update, download and extract a newer bundle. Persistent user data remains under the normal user-data directory.
Source installation, CLI startup, OpenAI-compatible API use, context routing, workstreams, and development instructions remain available in the documentation:
- Bundled installation and release production
- Browser cowork interface
- Proxy API reference
- Rolling ingestion
- Slot and workstream routing
- Architecture
- Developer and CLI reference
This is an accessibility-focused local AI project under active development. The application is intended for one trusted local operator. Workspaces and slots organize context; they are not security boundaries.
This fork is based on grishahq/recursive-llm, implementing the Recursive Language Models approach described by Alex L. Zhang, Tim Kraska, and Omar Khattab.
MIT. See LICENSE.