Full metadata observability for every AI session — model, tokens, cost, latency, and anomalies. No prompts. No responses. No payload logging by default.
Metadata fields sent per session. Never prompt or response text.
tsISO-8601 timestamp of session start / endmodelModel identifier (e.g. gpt-4o, claude-3-5-sonnet)tokens_inPrompt tokens consumed this sessiontokens_outCompletion tokens generated this sessionlatency_msWall-clock ms from first token to lastcost_usdCalculated cost using live pricing tableuser_idHashed identity — never raw email in payloadproject_idRepository / workspace identifiersession_idUnique UUID per session for drill-downerror_codeHTTP / model error code if session failed{
"session_id": "sess_a4f92c1e",
"ts": "2026-05-29T14:32:01Z",
"model": "claude-3-5-sonnet-20241022",
"tokens_in": 12401,
"tokens_out": 3892,
"latency_ms": 820,
"cost_usd": 0.1843,
"user_id": "u_8f3e… ← hashed",
"project_id": "proj_forg-dashboard",
"error_code": null
}Real-time ingestion — sessions appear within 50 ms of completion.
12.3M tokens across all models — last 7 days.
FORG surfaces behavioural anomalies automatically — no rules to write.
user: eve@co.com — $340 spend in the last 2 hours. Session loop suspected.
Model not in team allowlist. 3 sessions fired before block rule applied.
47 sessions logged between 00:00–05:00 UTC. Review for unattended scripts.
Every role sees exactly what they should — no more, no less.
HTTP 422 — protecting individual privacy at the API level.Your data stays as long as you need it.
FORG speaks the protocols your infra team already trusts.
Install the adapter in 60 seconds. Your first session appears in the dashboard within minutes — full metadata visibility, zero prompt logging.