apodex/frontier-agent-demo

FrontierAgent Demo

⭐ This demo is powered by the open-source [FrontierAgent](https://github.com/ApodexAI/FrontierAgent) project — visit GitHub to explore the code, run it locally, or leave a star.

This Space runs the real FrontierAgent react and agent_team workflows — the same native pipelines the project's CLI uses — in your browser.

Give it a task. It plans, searches the web, reads and writes files in a private workspace, and streams its answer back as it works. If it produces a deliverable, you can download it. Follow-up prompts keep the earlier conversation and workspace context, so you can refine the result across multiple turns.

Trying it

Type a task and press Run. Use Enable Agent Team for delegated research; while that mode is running, the intervention box can redirect its coordinator. Good first prompts:

• —Summarise what the ReAct pattern is and save it as `outputs/react.md`.
• —Compare two open-source vector databases and list the trade-offs.
• —Find the current recommended way to pin Python dependencies and write it up.

While it runs you will see queue/running status, each tool the agent starts and finishes, the answer streaming in, and any files it produced.

Stop asks the agent to finish its current step and return whatever it has — it is not a kill switch, so you still get a partial answer. Clear empties your workspace and conversation. New session gives you a fresh, empty one.

What the model is

The agent's reasoning happens on an external OpenAI-compatible endpoint; no model weights run inside this Space. The header shows which model is configured.

Limits and privacy

• —One task runs at a time, so you may briefly queue behind another visitor.
• —A conversation is limited to 10 completed turns. A visitor may start up to 30 runs in a rolling five-hour window, including across new sessions, and the demo as a whole up to 150.
• —Tasks have a turn budget and a wall-clock limit, then land with a best-effort answer.
• —Your session gets its own private directory. Other visitors cannot see or download your files, and a new session starts empty.
• —Sessions are temporary and are cleaned up automatically. Download anything you want to keep.
• —Please don't paste secrets or personal data into a public demo.

What the agent cannot do here

This is a public demo, so the agent runs with a deliberately narrow toolset: it can search the web, fetch pages, and read and write files inside your own session directory. It has no shell, cannot run code, cannot install packages, cannot download arbitrary files, and cannot reach anything outside its own workspace. Office-document deliverables (.docx/.xlsx/.pptx) are not available in this demo; text formats such as Markdown and CSV are.

For the full toolset — including shell access inside a sandbox — run FrontierAgent locally:

uv run frontier-agent --mode react --cwd /path/to/project

Self-hosting this demo

The Space is built from deploy/huggingface/ in the FrontierAgent repository. deploy/huggingface/README.md there is the deployment runbook: build and run commands, the full list of Variables and Secrets, how to point it at your own endpoint, and how to roll back.

Citation

If you find FrontierAgent or this demo useful, please cite Apodex 1.1: Scaling Agentic Intelligence for Complex Work:

@article{apodex2026apodex11,
  title   = {Apodex 1.1: Scaling Agentic Intelligence for Complex Work},
  author  = {{Apodex Team}},
  journal = {arXiv preprint arXiv:2608.23283},
  year    = {2026}
}