Give a self hosted LLM API access to write and run code inside its own sandboxed workspace.
Offload long installs or batch jobs to the background so the conversation is not blocked while they run.
Turn a completed task into a reusable tool that future conversations can search for and call.
Build a small interactive web app, like a two player game, that the model updates in real time.
| kkarsyline/curwe | 1038lab/comfyui-agnes-ai | adityaarsharma/wordpress-malware-removal | |
|---|---|---|---|
| Stars | 18 | 18 | 18 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | easy | moderate |
| Complexity | 4/5 | 2/5 | 3/5 |
| Audience | developer | developer | developer |
Figures from each repo's GitHub metadata at analysis time.
Needs Docker plus the NET_ADMIN capability, and has no built in authentication so a reverse proxy is required for remote access.
This repository is a workspace gateway that gives a self hosted AI model agent like abilities similar to Claude Code or Codex, but through any OpenAI or Anthropic compatible API and running on your own server. The model gets a dedicated folder called /workspace where it can execute shell commands, read, write, edit, and patch files, view diffs, and list directories. Long running jobs, like installing large dependencies or running a batch process for many minutes, can be pushed to the background so the chat is not blocked. When the job finishes, an event is sent back into the conversation so the model can pick up where it left off. The model can also turn a task it has completed into a reusable named tool and store it in a tools folder, so future conversations can search for and call that tool without having to relearn how to do the task. New tools are discovered at runtime rather than added to the fixed list sent with every request, which keeps prompt costs from growing as more tools are registered. Files written to an artifacts folder automatically become downloadable or viewable links, so the model can hand over reports, charts, zip files, or small static websites. The project also supports building small interactive web applications: the model writes both the frontend and the backend, which runs locally behind the gateway, and actions taken in the browser are sent back to the model so it can respond and update the page. This allows simple turn based interactions, such as the included five in a row game example. On security, if a user shares a credential in chat, the model only ever sees a placeholder token for it. The gateway resolves the real value into the environment of a spawned subprocess, keeping it out of the model's context, command strings, and logs. Shell commands run under a low privilege sandboxed user, and the gateway itself has no built in authentication, so exposing it publicly requires putting an authenticated reverse proxy in front of it. The project is written in Python and released under the MIT license.
A self hosted gateway that gives any compatible LLM agent style abilities: running shell commands, editing files, working in the background, building reusable tools, and creating small web apps.
Mainly Python. The stack also includes Python, Docker, Shell.
Use freely for any purpose, including commercial use, as long as you keep the copyright notice.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
double-check against the repo, no cap.