Govern the action, not the session
Each semantic action is authorized against attenuated authority in real time. A valid session is not a blank check.
Private beta · AI Action Control Plane
RaksHex evaluates consequential AI actions before execution, checks delegated authority and policy, releases brokered credentials only after an ALLOW, and records every decision in a tamper-evident Action Ledger.
Scoped evaluation for engineering and security teams. No self-serve production claims.
Agent identified
finance-support-prod
Authority checked
parent scope ≤ $50
Decision reached
DENY — exceeds limit
Ledger recorded
hash-chained entry
Evaluate the semantic action before execution.
Keep child authority equal to or narrower than its parent.
Release credentials only after an enforceable ALLOW.
Write the decision to a tamper-evident Action Ledger.
Drag the handle to compare governing who has a session against governing what each individual action is actually allowed to do.
Parent-to-child attenuation — a child authority can never exceed its parent's scope.
A DENY blocks the credential itself — the secret never reaches a denied caller.
Every decision is recorded tamper-evidently for audit and dispute resolution.
A valid API key or session token proves who is calling — not whether this specific action, right now, should be allowed.
Every semantic action is evaluated against the caller's actual delegated authority, not just whether the session is valid.
Enforcement happens at the credential broker, so a blocked action can't fall back to a raw API key that still works.
The hash-chained Action Ledger gives you a tamper-evident record of every authorization decision, not a log you have to trust.
The credential broker mediates every brokered request. A DENY blocks the secret from ever reaching the caller — enforcement lives at the same layer as the credential, not in a dashboard someone has to watch.
The Agent Firewall core, piece by piece: how an action gets authorized, delegated, enforced, and recorded.
Policy is written against what an agent is doing, not the HTTP call underneath it.
The same semantic action and delegated-authority model applies regardless of which LLM provider or coding agent is making the call.
Action Authorization & Policy Engine
One system of record
Point tools solve isolated pieces. RaksHex connects authorization, credential mediation, and evidence so the same answer holds up in an engineering review, a finance review, and an audit.
Each semantic action is authorized against attenuated authority in real time. A valid session is not a blank check.
The credential broker mediates brokered requests. A DENY blocks the secret from reaching the caller.
Evaluate provider and model rules, budgets, tool policies, PII redaction, injection signals, and kill switches before execution.
Adoption without theater
We do not publish invented customer logos, quotations, or comparative scores. Pilot teams receive a scoped rollout plan, permission map, test cases, and an exportable audit trail for their own environment.
Open the same evidence-first evaluation prompt in the AI tool you already use.
Founder
Founder & Builder, RaksHex
Building RaksHex to make autonomous AI agents controllable across actions, credentials, permissions, and spend before execution — with evidence teams can inspect afterward. RaksHex is currently a private-beta, pre-incorporation startup focused on technical validation and early design partners.
See what's new in RaksHex
Published customer legal, privacy, and security documentation
May 2026Waitlist system with email confirmation
April 2026Credential broker kill-switch controls added
Private beta FAQ
No certification theater and no magic claims. Start with the enforcement boundary, the authority model, and the evidence RaksHex leaves behind.
Read the Agent Firewall docsRaksHex is an AI Action Control Plane. Its Agent Firewall evaluates semantic actions against delegated authority and policy before execution, mediates brokered credentials, and writes decisions to a hash-chained Action Ledger.
Start with one agent, one provider, and one action that matters. Prove ALLOW, DENY, credential mediation, and the resulting audit evidence before expanding the rollout.