In development
LLM Remote Assistant
Natural language in, sandboxed Python out.
An experiment in giving a language model real hands. Instead of a fixed catalogue of commands, the model writes Python for the request at hand and a three-layer safety architecture decides whether that code is allowed to run. A mobile client is the command surface; a lightweight agent on the machine does the work.
- Python
- Flutter
- Go
- Rust
- WebSocket
- Sandboxing
- LLM
Genuinely dynamic execution
There is no list of supported actions to outgrow. The model interprets the request and writes an executable script for it, which means unusual multi-step tasks are just as expressible as common ones. The limit is what Python can do, not what someone thought to implement.
Three layers of safety, because one is not enough
Running generated code is the whole risk. First a deliberately restrictive system prompt shapes what gets written; then a second safety model reviews the script before anything runs; finally execution happens inside an isolated sandbox with no access to sensitive paths or unexpected network destinations. Any one layer can fail without the system becoming dangerous.
Split by design, not by accident
A cross-platform mobile client is the command centre and a small agent runs on the host machine. Keeping them separate is what makes the thing usable from anywhere, and it also means the trust boundary between "what the user asked for" and "what actually executes" is a real network hop rather than a function call.
Other side projects