Panopticon
Panopticon is an AI behavior control room that sits on top of logi. See in real time how a personalized MCP server or agent RP accesses your data, and step in to control it the moment you need to.
Why you need it
In the LLM era, the number of "AIs that handle my data" grows exponentially. For each AI:
- which tools did it call?
- with what permissions (scope), and what did it do?
- is it behaving differently from usual?
— all of this easily ends up locked in a black box. Panopticon lets the user look directly inside and control where the OAuth tokens that logi already issued are being used.
Panopticon × AI Guard
Panopticon is made of two modules.
| Module | Role |
|---|---|
| Panopticon | The place you watch — activity timeline, usage aggregates, anomaly detection |
| AI Guard | The hand that stops — Rate Limit, Kill Switch, HITL approval, Scope Drift blocking |
The two always work as a pair. When Panopticon detects an anomaly, AI Guard enforces the policy.
Data sovereignty
logi's core principle is "a risk-averse, minimal-retention IdP." Panopticon follows the same principle:
- The AI call history (trace) is an asset of the application the user themselves operates.
- logi only indexes and aggregates it; the meaning of the data is the RP's responsibility.
- Raw traces are deleted after a 90-day retention period; only aggregate data is retained permanently.
It gives you visibility that stays inside your own IdP — without sending telemetry to an external SaaS (Datadog, New Relic, etc.).
Identity and accountability in the age of agentic AI
The government plans to establish an "Agentic AI Safety and Trust Guideline" within this year, covering four risk factors: protecting personal data, governing AI agent permissions, preventing goal drift, and human oversight (announced as part of Korea's Ministry of Science and ICT (MSIT) "Agentic AI Initiative," July 2026).
Panopticon hasn't been certified or approved by any government body, and it doesn't claim to comply with a guideline that hasn't been finalized yet. What it shares with that initiative is the same underlying question — who the AI agent is, what it did, and who is accountable for it — worked out in advance inside your own IdP.
| Risk factor the guideline will cover | How Panopticon responds |
|---|---|
| Protecting personal data | Raw traces are deleted after a 90-day retention period; only aggregate data is retained permanently |
| Governing AI agent permissions | Every tool call records and aggregates the OAuth scope the RP reports as used (RP-reported) |
| Preventing goal drift | The Scope Drift Policy rejects requests for unregistered scopes at /oauth/authorize by default (block → invalid_scope) and records every drift regardless of policy |
| Human oversight | Agent Approval Gate's Ed25519-signed agent identity (agt_) authenticates the request, and the human approves high-risk actions with their own WebAuthn (passkey) signature, backed by a tamper-proof, append-only (WORM) audit log for after-the-fact accountability |
Beta notice
Panopticon is currently in beta.
| Feature | Beta | After GA |
|---|---|---|
| Trace ingestion (100% capture) | ✅ | ✅ |
| Per-application usage aggregates | ✅ | ✅ |
AI Guard — Rate Limit · Scope Drift (block by default) | ✅ | ✅ |
| Live Trace + Activity Timeline (console) | ⏳ Roadmap | ✅ |
| AI Guard — Kill Switch · Panopticon HITL endpoint | ⏳ Roadmap (Agent Approval Gate available as a substitute) | ✅ |
| Anomaly rules (Burst / Off-hour / Failed-auth) | ⏳ Roadmap | ✅ |
| Quota Enforcement | ❌ | ✅ |
| Tier differentiation / billing | ❌ | ✅ |
| LLM-based Intent Analysis | ❌ | 🔬 under review |
During the beta, you can freely use everything up to usage aggregates and policy configuration, and quota hard-blocking does not apply.
Next steps
- Getting Started — enable Panopticon in the console + send your first trace
- Trace API — the
POST /panopticon/tracespec - Policy Configuration — a guide to enabling AI Guard policies