empirica runs inside your AI coding tool as a plugin and a command-line tool. It keeps a record of what the AI learns and does, and it holds the AI to a simple rule: understand before you act. Everything stays on your machine.
Think before acting, enforced
Before the AI can change a file or run a command, it has to show it understands the work. A gate enforces this, so it is the system holding the line, not the AI's good intentions.
A record you can audit
Findings, decisions, and dead-ends are logged with their confidence as the work happens. You can look back and see what the AI knew, assumed, or got wrong, at any point.
Confidence checked against reality
After each task, the AI's view of how it did is compared against real evidence: did the tests pass, did the goals get met. The gap between belief and outcome is the signal that keeps it honest over time.
How the work is structured
empirica gives a piece of work a shape. A plan is broken into transactions, small measured chunks, each carrying its own goals and the tasks that deliver them. Inside every one, the AI runs the same short loop before anything ships.
A plan, broken into measured chunks
Work is decomposed into small, self-contained units. Each opens with a declared intent and closes with a measurement, so progress is something you can see in pieces, not one big leap of faith.
Goals and tasks, not vibes
Each chunk names its goals and splits them into concrete tasks. A task closes with evidence, a commit or a passing test or a file, so "done" means something checkable, not a feeling.
Investigate, then act, gated on grounded confidence
Inside each chunk the AI investigates first: it reads, searches, and records what it learns. It can only move to changing things once its confidence is grounded in what it found, not in a guess. That gate is the Sentinel.
Measured at the close
When the chunk is done, the AI re-checks its own state and the change is graded against real evidence. The gap between what it expected and what happened is what sharpens its calibration next time.
empirica works on one machine, for one practitioner. Sharing what it learns across projects and teammates is the job of the hosted product, not the free framework. The framework is the foundation. The shared layer is built on top.
Get started
empirica runs inside Claude Code and compatible terminal AI tools.
# Install the framework and CLI
pip install empirica
# Set up the plugin, hooks, and the gate for Claude Code
empirica setup-claude-codeAlso available via Homebrew and Chocolatey. See the repository for current versions and full setup.