Start with real work
Merion helps applied AI and research-engineering teams learn which change will improve a real workflow—and preserve the evidence required to defend that decision.
The system begins with genuine work rather than a synthetic benchmark. It captures the starting state, execution context, outcome, human corrections, and economic value of a work episode. Reviewed episodes can then become reproducible evaluation tasks and controlled experiments.
The learning loop
Merion is designed to sit above individual models and agent harnesses. A factory can observe work performed through Codex, OpenCode, a manual workflow, or another integrated harness without forcing a team to replace its execution environment.
What you can do today
- Initialize a local Intelligence Factory in a new folder, an existing directory, or a Git repository.
- Detect local sources and available Codex or OpenCode harnesses.
- Capture a genuine work episode with content-addressed starting-state evidence.
- Close the episode with human outcome review and economic context.
- Turn governed, reviewed episodes into audited evaluation tasks.
- Exchange bounded evidence through ATIF and Harbor adapters.
Start here
Follow the quickstart to link the private-alpha CLI, initialize a factory, and capture your first real-work episode.
Merion is currently a private-alpha, local-first system. Public package installation, hosted execution, autonomous training, and production customer-data ingestion are not available yet.
Last modified on August 21, 2026
