--- title: SEED Autonomous Unicorn Company emoji: 🌱 colorFrom: green colorTo: gray sdk: gradio sdk_version: 6.5.1 app_file: app.py python_version: 3.10.13 pinned: false license: apache-2.0 short_description: Autonomous software company --- # SEED — Autonomous Unicorn Company v1 SEED is an autonomous, observable, open-source software company hosted on Hugging Face. Unlike the earlier deterministic demonstration, this version performs real work when its ZeroGPU API endpoint is invoked: - collects current public signals from Hacker News, GitHub Issues and arXiv; - runs an open-weight coding model on ZeroGPU; - coordinates CEO, market, product, technology, development, testing, compliance, growth, resource-control and self-improvement roles; - selects an evidence-backed product direction; - creates product specifications and Python candidate code; - performs Python syntax and security checks; - asks independent Tester and Compliance roles to approve or reject the candidate; - promotes only candidates that pass all gates; - opens new organizational positions when a specific capability gap is found; - persists state, research, generated code, reports and audit events to a Hugging Face Dataset repository; - displays all activity through this Space. ## Important runtime fact ZeroGPU allocates a GPU only when a function decorated with `@spaces.GPU` is called. A sleeping free Space cannot wake itself. This repository deliberately contains **no Hugging Face Scheduled Jobs and requires no prepaid Job credit**. To work while nobody is viewing the Space, configure a free external HTTP wake-up to call the `autonomous_cycle` Gradio API endpoint. The wake-up performs no AI and stores no data. Everything produced by the company remains on Hugging Face. See [`docs/FREE_WAKE_UP.md`](docs/FREE_WAKE_UP.md). ## Required Space secrets and variables Open **Settings → Variables and secrets**. ### Secrets ```text HF_TOKEN= SEED_TRIGGER_SECRET= ``` ### Variables ```text SEED_MEMORY_REPO=Italianhype/seed-company-memory SEED_MODEL_ID=Qwen/Qwen2.5-Coder-1.5B-Instruct SEED_INTERVAL_MINUTES=120 SEED_MEMORY_PRIVATE=false ``` `HF_TOKEN` is required for durable Hub persistence. Without it, the Space still runs but falls back to ephemeral local state. ## Hardware Keep the Space on **ZeroGPU**. `autonomous_cycle()` is decorated with `@spaces.GPU(duration=180)` and executes model inference only while a cycle is running. Dashboard rendering, public research and Hub persistence use CPU. ## Open model The default model is `Qwen/Qwen2.5-Coder-1.5B-Instruct`, licensed Apache 2.0. Change `SEED_MODEL_ID` to another compatible open-weight causal language model after validating its license and ZeroGPU memory requirements. ## API wake-up The Space exposes this named Gradio endpoint: ```text autonomous_cycle(secret: str, force: bool) -> str ``` Example request: ```bash export SEED_SPACE_URL="https://italianhype-seed-autonomous-unicorn.hf.space" export SEED_TRIGGER_SECRET="your-secret" export HF_TOKEN="your-caller-token" ./scripts/trigger_cycle.sh ``` A free HTTP cron can send the equivalent POST every two hours. ## Hugging Face memory structure The Dataset repository configured by `SEED_MEMORY_REPO` is created automatically when the Space token has permission: ```text state/company_state.json artifacts/cycle-0001/research/evidence.json artifacts/cycle-0001/strategy/board_decision.json artifacts/cycle-0001/product/*.py artifacts/cycle-0001/tests/test_candidate.py artifacts/cycle-0001/review/quality_gate.json artifacts/cycle-0001/reports/cycle_summary.md ... ``` Every cycle is a single Hub commit containing the state snapshot and all new artifacts. ## Safety and release boundary This version produces and evaluates real candidate code but does not execute arbitrary generated code and does not overwrite the stable Space runtime. It uses `ast.parse`, Python compilation and a blocked-import policy before independent AI review. Candidate artifacts are persisted for inspection and future isolated testing. The autonomous organization can change its roles and product direction, but it cannot control payments, contracts, tokens or the immutable release boundary. ## Local tests ```bash python -m venv .venv source .venv/bin/activate pip install pydantic pytest huggingface_hub requests pytest -q ``` Local execution uses the deterministic fallback unless `SEED_LOAD_MODEL=true` and compatible CUDA hardware are available. ## Repository structure ```text app.py Gradio dashboard and ZeroGPU API seed_company/engine.py Real multi-role autonomous cycle seed_company/runtime.py Open-weight model runtime and JSON contracts seed_company/research.py Current public evidence collection seed_company/store.py Local atomic + Hugging Face Dataset persistence seed_company/models.py Typed organization and product state seed_company/projections.py Dashboard projections agents/founding_roles.yaml Founding organization and role policy docs/FREE_WAKE_UP.md Free unattended wake-up setup scripts/trigger_cycle.* Reference API clients tests/ Deterministic engine and persistence tests ``` ## License Apache License 2.0.