Watching agents is solved.
Proving and stopping them isn’t.
The open-source trust boundary for AI agents. We prove what they actually did and stop what no one approved.
Trust, not just telemetry.
Every team can call the same frontier models, and most already watch them — 89% have observability. What they can’t do is prove a run was actually right, or stop an agent before it does something irreversible.
That’s where agents fail in production: the silent success that’s secretly wrong, and the unapproved action no one can undo. Latency charts and token counts sail right past both.
Agenko makes every consequential action cross one open boundary — a receipt it can’t forge, a gate it can’t cross alone — so you can trust what your agents ship.
The problem
- 89% of teams have observability, but only 52% have any evaluation
- Silent failures slip past latency charts every day
- No open-source tool proves what an agent actually did
- No open-source tool stops irreversible actions before they happen
What we believe.
Prove and stop, don't just watch
Seeing is solved. The durable product is proving what agents actually did and stopping the irreversible ones.
Open by default
Apache 2.0, OTEL-native. Self-host in two containers.
Honesty is the product
Measured is measured. Inferred is labeled. No LLM guessing.
Build in public
Pre-launch. Spec and strategy open on GitHub.
Come build the boundary with us.
Read the source, open an issue, or follow the build on GitHub.