The Model Registry is where you bring the LLMs your systems call — models served from connected platforms — under governance, so you can observe their usage and apply policies to them.
Overview
Don't confuse the Model Registry with the Model Library: the Model Library (under AI Agents) is the workspace catalog of models you use to build agents on UnifyApps, while the Model Registry governs external models your apps already call.
Each row shows the model's Model name, Platform, Status, live telemetry — Last span (when it last recorded activity), Activity (its span count), Cost, Incidents, Policies — plus Workspace (in the all-workspaces view), Owner, and Last updated. Models arrive synced from Discovery or imported manually; only registered models are listed. The row menu offers Delete; clicking a row opens the model.
Empty state: "No models registered" — "Add models to bring them under governance — synced from discovery or imported manually."
Adding Models
Add Model — "Select the platform this model is served from" — is a three-step drawer:
Pick the platform the model runs on (the model-capable platforms — AWS Bedrock, Anthropic, OpenAI, Snowflake).
Select a connection — a connection is always required for models. Pick an existing, active connection, or use New to create one here (you need permission to create connections for that app). There's no manual name-entry path for models.
Import models — "Pick one or more models from this connection to import." Search ("Search models"), tick rows (each shows just the platform icon and the model's id, like
gpt-4o— that id becomes its registry name, and there's no rename), or use "Select all (N)" / "Clear all" (they act on the rows matching your search). The button reads "Import N models" once something's ticked.
On success: "Models imported successfully".
Note: In practice OpenAI is the platform that reliably lists importable models — other platforms can show "No models found — This connection returned no importable models." Re-importing a model you already registered is a silent no-op (the success message still shows). If the model list can't be fetched you'll see "Couldn't load models for this connection". Switching platform or connection clears your ticks.
Model Statuses
Status | Meaning |
|---|---|
Active | Registered and being observed. |
Inactive | Registered, but the platform hasn't confirmed data for it yet. |
Unlike agents, models have no Setup tab and no "Setup Pending" stage. A manually imported model shows Active immediately — before any traffic exists, so Active is not proof data is flowing; check its Activity/Last span columns or its Spans view for that. A model swept in by Discovery starts Inactive until its ingestion runs.
When a Model Shows Activity (Spans)
A model's usage is collected from the platform side. Today that collection is implemented for OpenAI models: the platform checks the connection roughly every 15 minutes for new stored completions of that exact model and records one span per completion (so only traffic your application stores with OpenAI is visible, up to a batch per check).
Note: Models registered from other platforms can sit Active with no spans — that's a platform-coverage limit, not a setup mistake on your side. Span latency for these collected completions always reads 0 ms (the provider doesn't expose durations), and per-span cost isn't shown. For OpenAI, the provider only returns completions your application asked it to store (the API's storage option) — un-stored traffic is invisible to this collection.
A Model's Detail Page
Five tabs — models have no Monitoring tab and no Setup tab. The header shows when the model's data last came in: a model connected through a platform connection shows Last synced with a Sync now button; otherwise it shows Last span.
Tab | What it shows |
|---|---|
Overview | Model Activity strip (Requests, Total cost, Error rate, Latency P90) over a volume chart windowed by Last 7 / 14 / 28 / 90 days ("No activity in this period" when quiet); an Incidents card (Total/Critical/High/Medium/Low, most recent few, "No policies yet"); a Model details card (Connection type, Status, Owner, Workspace, Department, Description). |
Observability | The model's Spans only (no sessions or traces), with an All / Evaluated toggle: status, span name, kind, trace, latency, tokens, started. "No Spans Found" when quiet. |
Evaluation | Online evaluation of the model's traffic — evaluators that score one model span at a time. Use Add evaluator to bind one; the gear opens Evaluation Settings (how often spans are swept for scoring). Empty: "Bind an evaluator to start scoring this model's traffic." |
Incidents | "Incidents flagged for this model by its governance policies"; empty: "Flagged and blocked policy violations for this model will appear here." |
Policies | "Governance policies currently applied to this model", read-only; clicking one opens the policy. |
Why Does My Model's Cost Read $0?
The Cost figures are computed from token counts priced against the workspace's model catalog — the Model Library under AI Agents, which holds each model's per-token pricing (not this Model Registry). If the model (or its exact provider/name combination) isn't in that catalog, its tokens are priced at nothing and the cost reads $0 even while activity flows. The same applies to an agent's cost on its Overview.
Note: These derived figures are estimates for governance, not a billing statement — expect them to differ from the provider's invoice.
Removing a Model
From its row menu, Delete (with a confirmation). This removes the model from governance — its policies stop covering it and it leaves the registry.
Note: Deleting is row-level only; there's no delete on the detail page, and no edit/rename anywhere.