The open protocol for agents to work with each other.
Your agents need to reach other agents — to book, buy, coordinate, and collaborate. ARP is the protocol that lets them do it safely. No platform owns the network. No vendor owns your agents.
Agents are starting to talk to each other. Most of it is held together with glue.
Every agent-to-agent interaction today is a custom integration: one-off API keys, one-off auth, no audit trail, no revocation, no way to know who is really on the other end.
The plumbing tax
You can't ship agentic features without rebuilding the same connection plumbing every time — pairing, permissions, logging, revocation, again and again.
Nobody is in control
You can't tell customers their agent is safe on your network when the network is a pile of shared secrets. Keys can't ask what a peer wants, and OAuth wasn't built for delegated agent authority.
All access or none
You can't give your agent real capability without giving it too much. Shared secrets are everything-or-nothing; agents need permissions scoped to the relationship.
Three moves. No glue.
ARP replaces per-partner integrations with three primitives: identity, connections, and policies.
Your agent gets a home
Claim a .agent domain, publish a small identity document, and your agent is addressable on the open internet. No API keys to rotate, no integration per partner.
It connects to other agents
Standard pairing and scoped permissions mean agents work together without custom glue code. You approve the relationship once; a mutually signed connection token binds policy, scopes, and expiry.
You stay in control
Every action runs against a policy you approved. Every decision is logged. Revocation is instant — one click revokes a connection everywhere, and the counterparty is told, not tricked.
Seven layers. One gate. Every request.
Between two agents, every request crosses seven layers gated by a signed policy you control. The policy decision point weighs identity, context, and the scope catalog on every single call.
PERMIT
Obligations applied — spend caps, approval windows, reversibility TTLs — then the audit entry is written and the signed reply goes out.
FORBID
The reason is returned, the audit entry is written, and zero side-effects occur. Denial is a first-class, logged outcome.
CONSENT
High-risk requests pause for the owner. You countersign, the token is mutually bound, and the edge goes active.
Policies compile to Cedar — the same policy language AWS uses for IAM — with ARP extensions: integer cents, epoch-ms, scope obligations. No floats, no ambiguity.
Safety isn't a feature. It's the foundation.
Consent and audit run on every interaction, for every agent, by default — no configuration required, no premium tier to unlock it.
Your keys never leave the browser
Your identity is generated in your browser and never transmitted. Every signed action is verified with the public half only — ARP Cloud keeps no copy of the private key.
Tamper-evident by construction
Every decision lands on a canonicalized, SHA-256 hash-chained audit log — exportable and verifiable offline, without trusting the runtime that produced it.
Low risk flies, high risk asks
Fifty reusable scope templates group into human-readable bundles — scheduling assistant, research agent, procurement agent. Routine reads flow through; spending money always asks.
Nothing leaks across edges
Each connection gets its own memory and tool context. Data shared on one relationship never surfaces on another — by construction, not by trust.
Cloud is convenient. Self-host is sovereign.
Everything is MIT-licensed. The TypeScript reference runtime is the same one that powers ARP Cloud — run it yourself or let us run it for you. Tokens bind to your agent's identity, not the host, so you can migrate without re-pairing.
Managed identity, consent, audit, and observability. Your agent is on the network in minutes.
The same runtime on your own hardware. Full sovereignty, same wire format.
Type-safe TypeScript SDK for custom agents — on the network in about thirty lines.
A contract between frameworks
ARP is not another agent framework — it's the standard they agree on. First-party adapters cover KyberBot, OpenClaw, Hermes, NanoClaw, and LangGraph; an adapter CLI and automated conformance testkit cover the rest.
Built in public
The spec, the scope catalog, the wire formats, and the reference runtime are published and MIT-licensed. No committee, no certification program — spec compliance is a green testkit run.
Free tier. Usage-based after that.
No per-seat fees. No per-integration fees. Early-access pricing while ARP Cloud is in preview.
For tinkerers and prototypes. No credit card.
- → 1 agent
- → 1,000 inbound messages / mo
- → Full audit + revocation
- → Browser-held identity
For teams putting agents on real traffic.
- → Unlimited agents
- → 10,000 inbound messages / mo
- → Advanced policy + obligations
- → Priority support
For compliance and volume.
- → SLA + dedicated support
- → SSO + RBAC
- → Security & privacy review
- → Dedicated runtime region
Give your agent a home. Keep the keys.
Claim a .agent domain, pair with your first peer, and watch every request cross the seven layers with your policy at the gate.