FrontierAgent
🧩 FrontierAgent, our agent framework, open-sourced alongside it — native command-line TUI, ReAct and Agent Team modes, one command on macOS and Linux, no preinstall, no hard Docker dependency.
What is it?
What it is
FrontierAgent is an open-source agent runtime, terminal product, and evaluation suite from ApodexAI for long-horizon research and file-based work. It ships as a Python TUI (`frontier-agent`) with two native workflows — a stateful ReAct agent and an Agent Team coordinator that dispatches work to parallel sub-agents — and the same workflow engine powers the benchmark harness used to evaluate Apodex models.
Why it exists
The project provides a reusable, open-source foundation for building and benchmarking agentic AI systems on complex, file-grounded tasks. It separates the framework, tools, workflows, and evaluation layer so each can be reused independently, and pairs the runtime with a benchmark suite (BrowseComp, FrontierSearchBench, FrontierChallenge, GDPval, etc.) and reproducible Apodex-1.1 results. The repository is also offered alongside a hosted, OpenAI-compatible Apodex-1.1 endpoint so users can run the agent without self-hosting a model.
Who should use it
Developers, researchers, and teams building agentic AI applications who need to orchestrate complex multi-step tasks involving research, file manipulation, and autonomous decision-making
Who should avoid it
Users without programming experience, individuals seeking simple scripting solutions, or projects requiring basic command-line tools without agent orchestration capabilities
Stars over time
No change stars in the last 1 day
How it works
A quick walkthrough in plain English
How FrontierAgent works
Step 1 of 3
You interact with it
Open FrontierAgent, send a request, or connect it to your stack.
Features
Advantages
- Modular architecture allows independent reuse of framework, tools, workflows, and evaluation
- Live TUI sidebar provides real-time task board, activity, and deliverables visibility
- Sandboxed security model with path policies fail-closed for safety
- Asynchronous intervention queues inputs without interrupting active runs
- Checkpointing and revert capabilities enable reliable session recovery
- Supports both ReAct and Agent Team modes for different task scales
- Docker and native deployment options across macOS, Linux, and GPU containers
- Extensive benchmark support with progress inspection and rerun individual failures
- Transparent deliverables with on-disk trace, checkpoint, and trajectory files
- OpenAPI-compatible model endpoint flexibility, no self-hosting required to start
- Community channels (Discord, X, Hugging Face) and detailed documentation
- Performance improvements shown in Apodex-1.1 vs prior benchmarks
Disadvantages
- GPU/SGLang setup requires matching driver/CUDA versions; mismatches cause opaque errors
- Scientific and document packages are optional; agent installs only task-specific dependencies
- Docker GPU setups require nested daemons; native SGLang avoids but adds complexity
- Windows support limited to WSL2; no native Windows binary
- Requires OpenAI-compatible model endpoint; no local model hosting included out-of-box
- Approval gates on mutating operations can slow interactive workflow
- Some benchmarks require external scorers and separate isolation (e.g., FrontierSearchBench)
- Learning curve steep due to modular architecture and many configuration options
- Benchmark evaluation requires downloaded datasets per official guide
- Concurrence management needed for Agent Team; start with --concurrency 1
- Web research tools need SERPER_API_KEY and JINA_API_KEY for full functionality
- Limited-time free API offer on Apodex platform; not permanent free tier
Installation
docker
cp .env.example .env docker compose run --rm agent
native
git clone https://github.com/ApodexAI/FrontierAgent.git cd FrontierAgent uv sync --python 3.12 --extra dev cp .env.example .env # Add OPENAI_API_KEY, OPENAI_BASE_URL, OPENAI_MODEL to .env uv run frontier-agent --mode react --cwd /path/to/project uv run frontier-agent --mode agent_team --cwd /path/to/project
compose
cp .env.example .env docker compose run --rm agent
FAQ
What are the two native workflows in FrontierAgent?
FrontierAgent ships two native workflows: ReAct, which uses one stateful agent for research, file work, and command execution in a task-scoped sandbox, and Agent Team, where a coordinator maintains a task board, delegates independent work to parallel sub-agents, collects their reports, and synthesizes the final result.
How do I install and start FrontierAgent?
Clone the repo, run 'uv sync --python 3.12 --extra dev', copy .env.example to .env, add your OpenAI-compatible endpoint details, then launch with 'uv run frontier-agent --mode react --cwd /path/to/project' for ReAct or 'uv run frontier-agent --mode agent_team --cwd /path/to/project' for Agent Team.
What is the filesystem and security model?
The sandbox uses three paths: /inputs (read-only for supplied documents), /workspace (read-write for source and scratch work), and /outputs (controlled read-write for persistent deliverables). File and shell tools share this sandbox, with interactive sessions requiring approval on writes and mutations journaled for /revert recovery.
How do I run benchmarks with FrontierAgent?
After syncing with --extra eval --extra sandbox --extra document-readers and downloading datasets, run a smoke test with 'uv run python -m benchmarks.public.runner.run_subprocess --benchmark browsecomp --pipeline stateful-react-agent --profile default --limit 1 --concurrency 1 --out ./results/smoke'. See docs/eval.md for full operator reference.
Can I use Docker or local SGLang models?
Yes. Pre-built linux/amd64 and linux/arm64 images are available via 'docker compose run --rm agent'. For local models, Docker SGLang on Linux NVIDIA hosts and native SGLang without nested Docker are supported, with configuration details under config/sglang/.
Featured in Videos
YouTube tutorials and walkthroughs for FrontierAgent
Alternatives
Similar projects ranked by category, topics, and text overlap.
