OpenFARS

English | 简体中文

A local-first, multi-agent research OS for ideas beyond the average.

OpenFARS routes each research stage to the model best suited to it, searches a quality-diverse idea frontier, asks humans only for high-value decisions, runs real experiments locally or on remote GPUs, and turns verified evidence into figures, a paper, media packages and a reviewable open-science release.

direction → literature → exploration → critique → task → plan
          → experiment ⇄ evaluation → figures → paper → podcast → video → release

What is different

• Not best-of-N brainstorming: causal divergence operators, heterogeneous model families, nearest-literature checks, blind judges and a quality-diversity archive.
• Human gradients, bounded context: decision packets expose finalists, falsifiers, anomalies and disagreement—not full transcripts.
• DeepSeek Harness inside: the experimenter uses a durable Harness session with plugin composition, sandboxed tools, lifecycle events and workspace permissions.
• Real research infrastructure: SSH GPU execution, append-only traces, failed-run retention, deterministic plots and evidence-locked citations.
• Local 3D WebUI: a lightweight office shows all 13 agents walking, working and handing off tasks alongside the SSE event stream, artifacts and human approvals; secrets never enter the browser.
• Safe one-click openness: checksums, cards, RO-Crate and explicit, identity-verified publishing to GitHub, Hugging Face and ModelScope.

Quickstart

Install Miniconda first. The commands below create an isolated Python 3.11 environment (OpenFARS supports Python 3.10+).

conda create -n openfars python=3.11 -y
conda activate openfars

git clone https://github.com/open-fars/openfars.git
cd openfars
python -m pip install -e .

# Zero-key demo: completes the workflow but never simulates experimental evidence.
openfars web --config examples/offline.yaml

Open http://127.0.0.1:8765. The lightweight 3D office puts 13 clickable agents, the task graph, live events and handoffs, artifacts, model routes and human decisions in one view. Appearance and movement are adjustable; without WebGL, the same desks remain available in 2D.

For frontier routes:

python -m pip install -e '.[models,harness,publish]'
cp openfars.yaml openfars.local.yaml
export OPENAI_API_KEY=...
export ANTHROPIC_API_KEY=...
export GEMINI_API_KEY=...
export DEEPSEEK_API_KEY=...
openfars doctor --config openfars.local.yaml
openfars web --config openfars.local.yaml

Keys are read from environment variables only. Never place tokens in YAML. OpenFARS uses LiteLLM's unified adapter for 100+ hosted and local providers, while keeping model choice independent per agent. See provider configuration, including Ark, OpenRouter, vLLM and Ollama examples.

Default model team

Defaults are an evidence-based snapshot dated 2026-08-15, not permanent winners.

Role	Default
director, librarian, task_designer, critic, evaluator	GPT‑5.6 Sol
explorer	Claude Opus 4.8 + GPT/Gemini/DeepSeek model pool
planner, writer, podcaster	Claude Opus 4.8
experimenter	DeepSeek‑V4‑Pro through DeepSeek Harness
visualizer	GPT‑5.6 Sol + deterministic renderer
video_producer	Gemini 3.6 Flash
publisher	GPT‑5.6 Luna + deterministic permission checks

The detailed current-best design, alternatives and promotion benchmarks for all 13 roles are in docs/AGENTS.md. Refresh task-specific leaderboard snapshots with:

openfars models-refresh --config openfars.local.yaml --force

Snapshots are advisory and never silently rewrite model routes. The WebUI refreshes stale subscriptions in the background; CLI-only users can run the command above explicitly.

Podcast and video agents produce evidence-linked, reviewable source packages. Optional renderers use shell-free argument lists; their logs stay out of the release bundle.

Remote GPU experiments

Keep the private key on the controlling machine and reference its path through an environment variable:

compute:
  targets:
    gpu-lab:
      host: 121.89.85.xxx
      user: root
      port: 32430
      identity_file_env: OPENFARS_SSH_KEY
      workdir: /data/dingrui/code/openfars
      output_dir: /data/dingrui/output
      datasets_dir: /data/dingrui/datasets
      models_dir: /data/dingrui/models

export OPENFARS_SSH_KEY=/Users/xxx/.ssh/id_ed25519_xxx
openfars remote-probe gpu-lab --config openfars.local.yaml

OpenFARS calls system ssh/rsync without reading or uploading private-key bytes. It clears stale remote results before each run and never advances after a failed command. See the remote GPU example and sandbox design.

Run, review, publish

openfars run --config openfars.local.yaml --topic "your broad direction"
openfars status <project-id> --config openfars.local.yaml
openfars decide <project-id> idea --approve --select <idea-id> --feedback "..."
# Or move the whole frontier; prior candidates and feedback remain archived.
openfars decide <project-id> idea --revise --feedback "seek a cheaper causal falsifier"

# Local bundle only; no external write.
openfars bundle <project-id> --config openfars.local.yaml

# Explicit external authorization; authenticated identities are verified first.
openfars publish <project-id> --github --confirm --config openfars.local.yaml

GitHub publication is restricted to authenticated account Dingrui-Wang; the repository owner is open-fars. Every external destination requires explicit user-provided permission.

Design

• Architecture and DeepSeek Harness mapping
• Agent research and model routing
• Model/API providers
• Auto-research gaps, OpenClaw lessons and product thesis

python -m pip install -e '.[dev]'
pytest
ruff check src tests scripts run.py
# With the Harness extra installed; uses only a loopback fake provider.
python scripts/harness_smoke.py

MIT licensed. AI-generated research artifacts must disclose AI assistance and remain subject to human scientific, safety, privacy and license review.