[](https://huggingface.co/deathlegionteam)
---
## Who we are
We're not building another chat wrapper. Death Legion AI is an autonomous agent platform — the kind that takes a task, breaks it into steps, writes and runs its own code in an isolated sandbox, checks whether the result actually worked, and only comes back to you when it's done or genuinely stuck.
Think Devin, Manus, Capy — same category, our own stack, built to run self-hosted instead of behind someone else's paywall.
If you've used agent tools that confidently return broken output and call it a day, that's exactly the failure mode we built against.
↑ replace this with a real screen recording of the agent loop — GIFs of actual output convert way better than any badge wall
---
## What it actually does
- **Autonomous task loop** — an 11-tool agent loop that plans, executes, checks its own work, and retries on failure instead of silently returning garbage
- **Real sandboxes, not simulations** — every task runs in an isolated Docker environment (self-hosted, no dependency on third-party sandbox APIs), so code execution is real, not roleplayed
- **Streaming everything** — SSE-based streaming means you watch the agent think and act in real time, not stare at a spinner
- **iOS / Android dev modes** — mobile-first sandboxed environments for testing agent output on-device, not just in a browser tab
- **Full state persistence** — Prisma-backed schema tracks every task, tool call, and file the agent touches, so nothing disappears when you close the tab
| Feature | Status |
|---|---|
| Agent loop (11 tools) | ✅ Shipped |
| Self-hosted sandbox microservice | ✅ Shipped |
| SSE streaming UI | ✅ Shipped |
| iOS/Android dev sandbox | 🚧 In progress |
| Multi-agent orchestration | 🧠Designing |
---
## Stack
```
Frontend → Next.js 15, Zustand, SSE
Backend → Node.js, Prisma, SQLite
Sandbox → Docker, deployed on Fly.io / Railway
Agent core → Custom tool-calling loop, 11 tools, retry + verification logic
```
No LangChain. No black-box orchestration frameworks. If something breaks, you can read the loop and know exactly why.
---
## Why self-hosted sandboxes
Most agent platforms rent sandbox execution from a third party and mark it up. We built our own Docker-based sandbox microservice instead — deployable on Fly.io or Railway, no per-execution API tax, and no depending on someone else's uptime for your agent's ability to run code.
It's more work upfront. It's also the difference between "our product" and "a thin wrapper around someone else's infrastructure."
↑ swap for a GIF of a task running end-to-end in the sandbox
---
## Roadmap
- [ ] Multi-agent orchestration (agents that spin up sub-agents for parallel subtasks)
- [ ] Persistent memory across sessions, not just within a single task
- [ ] Public model checkpoints for the planning/verification components
- [ ] Full mobile client (not just dev-mode sandboxes)
Scope keeps growing. That's on purpose — we'd rather ship something ambitious slowly than something small fast.
---
### Models & Spaces
*Browse the repos below for what's actually live right now.*
