POLICYBOUND AGENT BUS
WebMCP prototype

PEOPLE + AGENTS, WITH A TRUST BOUNDARY

Let agents collaborate without handing the relay your secrets.

PolicyBound Agent Bus turns a sensitive agent handoff into a policy decision first: who may read it, why, for how long, and how many times. WebMCP gives a browser agent a safe, structured way to ask for that plan.

Try the policy studio See the boundary

This public demo is preview-only. It never accepts a real secret, key, token, or private endpoint.

✦Human / Agentintent
→
✓Policy Enginebrowser-mediated
▣Encrypted Relayciphertext only
→
◎Agent Runtimekey stays isolated
⌁ readable only inside an authorized runtime
01Purpose-boundEvery preview states why access is needed.
02Time-boundedShort-lived expiry is part of the envelope.
03Key-isolatedThe model never receives a private key.

THE IDEA

A secret bus with a narrow public surface.

Traditional agent tools make it easy to call an API. Sensitive workflows need one more question: can this agent use the result for this purpose? PolicyBound Agent Bus keeps that decision explicit.

A

WebMCP at the edge

The page exposes typed tools instead of asking an agent to guess at buttons, fields, or hidden state.

document.modelContext.registerTool()
B

Policy before payload

The first action creates a redacted envelope plan. A relay can route ciphertext without learning the message.

purpose · expiry · max_reads
C

Runtime owns the key

Decryption belongs to an authorized agent runtime and its secure key store, never to the prompt or the relay.

plaintext_exposed_to_relay: false

LIVE POLICY STUDIO

Ask an agent to prepare a safe handoff.

Checking WebMCP…
01
Define the boundaryThese values become authenticated policy metadata.

No network request is made. The preview uses demo identities and synthetic metadata.

02Envelope plan
WAITING
⌁

Run the policy tool to see a redacted, machine-readable result.

⌁Demo mode · ciphertext and keys are never created here

FOR AGENTS

Three small tools. One clear contract.

In a WebMCP-capable browser, an agent can discover the tools below and call the policy planner with structured arguments. The human UI and the agent surface share the same validation path.

◇
inspect_security_boundaryExplain what the relay can and cannot see.
ready
◇
list_demo_agentsReturn public demo identities and capabilities.
ready
◇
create_policy_bound_envelopeValidate purpose, expiry, reads, and forwarding.
ready