Company
ApexClaw
ApexClaw builds the layer that decides whether an AI agent's action is authorized — before it executes — and produces evidence of what happened that a reader can verify without us.
Why this exists
Most AI safety work addresses what a model says. The harder problem arriving now is what an agent does: moves money, changes a record, actuates equipment, reaches a tool that can act. Those effects are not recoverable by editing a response.
The industry's answer so far has been observability — watch closely, log thoroughly, alert quickly. All three describe a decision that has already been made. ApexClaw takes the decision itself: identity, authority, deterministic policy, human approval and hard limits applied in the same transaction that commits the action.
The second half is evidence. A log written by the component that performed the action, signed with the same key, attests to very little. Records here are hash-linked, signed with separated keys, and checkable offline by a verifier that needs no database, no network and no installation of the product.
How we work
Acceptance gates before code
Each capability has a written acceptance contract before it is implemented, so a passing result cannot be produced by moving the target after seeing the output.
Verdicts, not adjectives
Capabilities are published as proven, candidate or planned. The weak ones stay visible; removing them would make the strong ones less believable.
Limitations attached to claims
Every stated limitation sits on the same card as the capability it constrains, because a limitation on another page is a limitation nobody reads.
Founder
ApexClaw was founded by Julian Joseph, who chairs the W3C Agent Conformance and Benchmarking Community Group. Participation in a community group is not endorsement: the W3C does not certify ApexClaw and does not endorse it.