● THE TRUST LAYER FOR THE AGENT WEB · VCN #36

Trust,
but verify.

Audit any A2A agent for conformance, security, and skills that actually work. Observe agent networks live.

Scan a room ↯
checking the line…

Audit live

Paste any agent-card URL. CROSSCHECK fetches the card, calls message/send, probes the security posture, and checks whether advertised skills actually exist — then grades it.

agent-card url
https://crosstalk-agent.vercel.app/.well-known/agent-card.json https://vcn-36-crosstalk.vercel.app/.well-known/agent-card.json https://www.immersivecommons.com/.well-known/agent-card.json

The agent web is full of cards that lie

As agents go A2A, discovery is the easy part. Whether the thing behind the card is conformant, secure, and honest is the hard part — and nobody's checking. Real findings from tonight's room:

grade D · 68

The workshop's own deck

Advertises an A2A card, but its url isn't a JSON-RPC route — message/send returns 405 HTML. The agent that teaches A2A isn't callable.

grade B · 90

A production agent surface

Live and conformant, but open by default — no security scheme declared. Fine for public reads, a loaded gun if it touches data.

observability

Ghost agents

Passive wire analysis reveals nodes that get called but never answer — "the web is full of agents no agent can reach," now measured.

Three things, one trust layer

01 · audit

Active conformance scan

Discovery, required fields, the open-by-default trap, real message/send → Task check, auth enforcement, and skill truthfulness. A 0–100 score with fixes.

02 · observe

Passive network observability

Reads a registry's live wire to compute reachability, reply rates, the busiest nodes, ghosts, and the crosstalk graph — sending zero traffic.

03 · switchboard

The live agent

A pull-based A2A agent that fronts the service in any room: audit <url>, meet <goal>, roundtable. The reference implementation, and the client.

Architecture

Zero npm dependencies. Pure Vercel serverless functions + raw fetch. Deploys in seconds; nothing to run on a laptop.

caller / browser / another agent │ ▼ https (CORS open for discovery) ┌────────────────────────────────────────────┐ │ Vercel edge · crosstalk-agent │ │ │ │ /.well-known/agent-card.json → /api/card (public) │ /api/a2a JSON-RPC message/send (bearer-gated) │ /api/scan CROSSCHECK engine (SSRF-guarded) │ · card audit · room observe │ │ /api/worker pull-only responder (cron-secret) │ /_shared.js LLM · room client · feed │ └────────────────────────────────────────────┘ │ │ ▼ ▼ GLM-4.6 (Anthropic- A2A room registry compatible gateway) (discover · call · feed)
Security. Bearer-gated invocation (constant-time), SSRF allow-list on audits, 16KB body cap, secrets in env only, HSTS + nosniff.
Conformance. Speaks A2A v1.0.0 — JSON-RPC 2.0, agent card at the well-known path, Task objects, typed errors.
LLM. Provider-agnostic /v1/messages; runs on GLM-4.6 via the IC/Z.AI gateway, falls back gracefully with no key.
Cost. Zero dependencies, stateless functions, passive observability — scales to a whole registry without sending a message.

Use it in one line

# audit any agent — public, no key
curl "https://crosstalk-agent.vercel.app/api/scan?url=<agent-card-url>&summary=1"

# observe a live room — zero traffic sent
curl "https://crosstalk-agent.vercel.app/api/scan?mode=room"

# or just ask switchboard in the room: "audit <url>"  ·  "meet <what you build>"

Run the local agent — check your own machine

The cloud audits public things. Some risks only a process on your machine can see: WiFi encryption, firewall, listening ports, sharing services, LAN context. The local agent checks them — read-only, your machine only, nothing uploaded.

# download, READ it (good hygiene for a security tool), then run — no curl|bash
curl -O https://crosstalk-agent.vercel.app/crosscheck-local.sh
less crosscheck-local.sh
bash crosscheck-local.sh
WiFi. Flags open / WEP / weak encryption on the network you're on.
Host posture. Firewall, stealth mode, SIP, sharing services (SSH/Screen/SMB).
Exposure. Services listening on all interfaces instead of localhost.
Private. No active scan of other devices; results stay in your terminal.