product · registry
Registry
The inventory. Every agent, tool, and session your organisation runs — and seven anomaly rules watching them.
Most teams can name the two engineers who know which agents exist and what they call. Registry makes that knowledge a system: agents, tools, sessions, and permissions, captured by AGT, searchable by the people who answer for them — plus, on Business and up, the AI services your logs show but nobody registered, found by Radar.
the raw material
What an event looks like
After the wrapper tokenises PII and adds the signature envelope, this is what lands on /ingest and what you see in Registry → Events.
The PII: tokens are deterministic within a session — the same email resolves to the same token without ever revealing the value to us.
// Event identity:
// What the agent did:
// Parameters (PII tokenised):
// Integrity proof (3-layer):
behavioural detection
Seven rules. Running on every event.
Sub-second evaluation, deduplicated per rule and agent so one issue doesn't page you thirty times. Custom rules ship Q1 2027.
what's inside
Six views, one inventory.
events
Every tool call and message, searchable, with AGT's policy decision and envelope status.
agents
Last seen, event counts, trust score, key fingerprint, delegation history.
tools
Per-tool risk tier, owner, classification mix, destination breakdown.
sessions
A multi-step run reconstructed end-to-end by AGT trace ID.
permissions
Read-only view of AGT policy state; editing lives in Shield.
anomalies
The seven rules above, with a review queue.
wytness radar · discovery
Discovered from the logs you already keep.
Radar reads identity sign-ins, proxy exports and vendor usage reports for the fingerprints of AI services, and lands what it finds here as unreviewed rows. No SDK, no change to any agent: a read-only container beside the logs, on the same chassis as the Collector.
identity
Entra sign-in logs — the service principals and apps that authenticated to AI services. The strongest owner attribution.
proxy
Zscaler, Netskope or any egress export landed in Log Analytics — the traffic nobody registered.
vendor usage
OpenAI and Anthropic org-admin usage reports — every key and project on the account.
A finding arrives with a Radar badge and walks the same approval queue as every other agent. It records nothing until you instrument it, never counts against your agent cap, and a low-confidence link is shown as a suggestion you confirm — never promoted on its own. Business and up.
the honest pitch
- what a finding proves
- → something ran, from where, how often
- what it cannot prove
- → what it did — only the SDK and the Collector produce evidence
- what leaves your side
- → matched hits and day counts; raw logs and credentials never do
- coverage
- → bounded by the sources you connect; the blind spots are shown, not buried
availability
What each tier gets
| Starter | Growth | Business | Enterprise | |
|---|---|---|---|---|
| Events · Agents · Tools · Sessions · Permissions | ✓ | ✓ | ✓ | ✓ |
| Anomaly engine (7 rules) | Full | Full | Full | Full |
| Custom anomaly rules | — | — | API only | API + UI (Q1 2027) |
| Trust score history | 30 days | 90 days | 7 years | Custom |
| Radar discovery (identity · proxy · vendor usage logs) | — | — | ✓ | ✓ |
See every agent. Catch what shouldn't be there.
Starter covers the full inventory and the full seven-rule anomaly engine.