Avirak runs your agents and guards every call they make. Pick a world, then scroll.
Zoom, Slack, GitHub, a calendar, a payments provider or any MCP server. Log in once. Avirak keeps the token.
What starts work, which agents run with which tools, and what waits for a person. Anything you did not mention is refused.
Holds reach your phone, the console or Slack. One tap to approve or refuse. Silence never means yes.
Approve is one hold at a time, by design. Bulk deny only. Every row shows the exact arguments the call was decided on.
A step is offered by the server, claimed under a lease, kept alive by heartbeats and reported back. A stale claim is refused, never applied twice.
Every document is validated against the four compiled articles before it is installed. A document that weakens a human gate is refused with the article quoted.
Loosening a rule needs a typed confirmation phrase and is itself written to the ledger: who, when, before and after.
A write to the ledger that fails refuses the action. Commits carry the model that wrote them. Nothing here can be switched off.
Avirak logs in once and keeps the credential. It performs each allowed call itself. If avirak is down the call fails; it never runs locally with a borrowed key.
Tokens, decisions and the audit trail never leave your company. Nothing to sign up for, no identity provider to adopt. Agents on other machines reach it over mutual TLS.
Agents implement, review and test. Review and QA can never be the same agent that wrote the code.
A front-desk agent triages requests and keeps the day's schedule tidy. A nurse answers every patient message before it leaves.
A planner turns enquiries into itineraries and holds seats. Money moves only when the owner taps approve.
Avirak is Cloudflare for AI agents: every tool call passes through it, allowed, held or refused, and every one is on record.
Tell us which agents you run and which tools you want them to reach. We will set you up and walk through the first workflow with you.