barycenters (Python)

Govern any AI agent in 3 lines. Ships in shadow mode — drop it into your LangChain / CrewAI / LangGraph loop and see what /admit would have blocked. Zero dependencies.

pip install barycenters

Drop it in (shadow mode — zero risk)

from barycenters import Barycenters

bary = Barycenters(api_key="bary_...", namespace="acme/eng")
# mode defaults to "shadow": it NEVER blocks. It only records what /admit would have refused.

def deploy():
    d = bary.admit("deploy_prod", metadata={"env": "production"})
    # shadow: d never raises. d.would_block tells you what WOULD happen once you enforce.
    do_the_deploy()
    bary.receipt(d.token, output_hash="sha256:...")

Or the decorator — the actual 3 lines:

@bary.boundary("delete_prod_db")
def wipe():
    ...

Run your agent for a day. Then read the one thing that sells it to your team:

print(bary.shadow_report()["summary"])
# → "2 of 17 action(s) would have been blocked by /admit"

That report is the whole pitch: you learn exactly which agent actions your policy would have stopped — before a single one is blocked. No risk, no downtime, no config to trust.

Honesty contract

• Shadow, no endpoint wired: every decision is an honest AWAITING (NO_ENDPOINT) — never a fabricated pass. The shadow report says so.
• Enforce mode: a refusal raises RefusedError; an unreachable endpoint fails closed (raises too) — ungoverned execution is never silently allowed.
• Zero dependencies: stdlib urllib only; transport is injectable for tests.

License

AGPL-3.0-or-later. A commercial license is available from Barycenters Systems LLC.