Skip to content

Project notice

Start here — get a useful, reviewable result

Alpha-Factory runs bounded missions against inputs you supply, retains a signed journal, and asks you to inspect the result before approval. Start with the allocation example; it needs no wallet, API key, Docker or model download. The factory guide explains all five mission types and the complete Ascension path. Browser-only examples are available in Decision Studio.

For self-generated reasoning tasks and independent solver review, open the Curriculum Lab.

1. Install one release

Use Python 3.11, 3.12 or 3.13. From one release, download these four assets into the same folder:

  • The alpha_factory_v1-…-py3-none-any.whl wheel.
  • requirements-agent.lock.
  • SHA256SUMS.
  • install_agent.py.

Open a terminal in that folder and run:

python3 install_agent.py --release-dir . --venv .venv-agent

On Windows use python in place of python3. If several Python versions are installed, select a supported interpreter explicitly, for example python3.12.

The installer checks the wheel and dependency lock against the release checksums, creates a new environment, installs the hashed dependencies and verifies the installation. Do not mix assets from different versions. If the environment path already exists, choose a new path; your existing state is retained. For an offline install, add --wheelhouse /path/to/compatible-wheels. Download and extract alpha-agent-v…-operator-guide.zip for all the linked guides in one folder.

Activate the environment on macOS or Linux:

source .venv-agent/bin/activate

On Windows PowerShell:

.\.venv-agent\Scripts\Activate.ps1

2. Run and inspect the example

alpha-agent examples --output my-missions
alpha-agent --home ./agent-state init
alpha-agent --home ./agent-state run my-missions/allocation.json
alpha-agent --home ./agent-state serve

Open http://127.0.0.1:8765 and use the token in agent-state/api.token. Select the returned mission. Inspect its selected projects, budget, risk and totals, then approve or reject the exact result. These example inputs are constructed scenarios; edit their assumptions before using a result for a real decision. Approval records your review and does not execute a trade or make a payment. Stop the console with Ctrl+C.

3. Continue with your own work

  • Edit the copied examples: research, allocation, scheduling, forecasting or code. Code evaluation requires explicit permission and an isolated Docker runtime.
  • Use the factory guide for reviewed native work, FusionPlans and signed delivery.
  • Explore workflow alternatives in the MATS Search Lab, then export a reproducible review bundle.
  • Use the demo walkthrough for all browser experiences and the preserved local experiments.
  • Use operations and recovery before upgrades, backup, restore or provider configuration.
  • Use the Ascension protocol guide for the local-EVM enterprise lifecycle.
  • Inspect the release's release-manifest.json and alpha-agent-v…-validation.zip for exact acceptance results. The source, browser, complete site and unchanged manuscript are separate release assets.

If a command fails, keep the existing environment and journal. Check the JSON error, supported Python version and selected profile; follow the recovery guide using a new destination. Never manually edit signed journal records. The maintained runtime and the undeployed contract reference are distinct from the broader research vision; release readiness defines the verified scope.