A recent Hacker News thread with 96 points and 53 comments discussed Building an (almost) fully self-hosted, sandboxed, agentic software factory. The post describes a local setup that runs AI coding agents inside isolated containers without relying on external APIs.
What It Is
The system combines local LLMs, container orchestration, and agent frameworks to create a closed-loop development environment. Agents generate code, run tests, and iterate inside sandboxes that prevent access to the host system or network.
All components stay on-premise. The stack uses open-source tools for model inference, task routing, and execution isolation.
How It Works
Agents receive tasks through a central orchestrator. Each agent operates inside a Docker container with restricted CPU, memory, and filesystem access. Models run via local inference servers that expose OpenAI-compatible endpoints.
Results feed back into the orchestrator for review or further agent calls. Sandboxing relies on Docker seccomp profiles and network namespaces rather than cloud provider controls.
How to Try It
Install Docker and a local inference server such as Ollama. Pull a coding-capable model and expose it on localhost:11434. Clone the repository linked in the original post and run the provided docker-compose file.
Configure the agent loop with environment variables that point to the local endpoint. Start the orchestrator and submit a simple coding task to verify isolation.
"Core components used"
Pros and Cons
- Keeps all code and prompts inside the local network
- Avoids per-token API costs for high-volume agent runs
- Requires manual maintenance of model updates and container images
- Limited model capability compared with frontier cloud models on complex refactoring tasks
- Sandbox escape surface remains if container runtime is misconfigured
Alternatives and Comparisons
Cloud agent platforms such as Devin or GitHub Copilot Workspace offer higher model intelligence but send code outside the organization. Local stacks using LangGraph or Semantic Kernel provide similar orchestration without the sandbox focus shown in the post.
| Feature | Self-hosted sandbox setup | Cloud agent platforms | Local orchestration only |
|---|---|---|---|
| Data residency | Full | None | Full |
| Sandbox isolation | Docker + seccomp | Provider-managed | Minimal |
| Model choice | Local only | Frontier models | Local only |
| Maintenance overhead | High | Low | Medium |
Who Should Use This
Teams handling proprietary codebases or regulated environments benefit most. Organizations already running local LLMs and comfortable with container security gain a practical path to agentic workflows.
Skip this approach if rapid prototyping with the strongest available models matters more than data control.
Bottom Line
The setup trades model performance for full control and isolation, making it viable for internal tooling where data cannot leave the premises.
Local agent factories remain early-stage experiments, yet the pattern of combining container sandboxing with on-premise inference is likely to spread as more teams prioritize data residency.
Top comments (0)