🧠 AI
The engine that runs agents — and the app that maintains them.
One entry. Installing this gives a deployment its chat runtime and the catalogs where agents, models, providers, skills and tiers are administered. They ship in one assembly because they are one thing: an engine nobody can configure is not a feature.
What ships
- The agent runtime — threads, rounds, delegation, tool calling, harnesses, token accounting.
- The maintenance app — the Agents, Skills, Providers, Models and Tiers catalogs, plus the agent overview and details views.
- The content types —
Agent,Skill,Harness,ModelProvider,Thread,ThreadMessage,TokenUsage— served from Postgres or in-memory depending on what this deployment configures.
What does NOT ship here
Model providers are their own entries. OpenAI, Anthropic, AzureFoundry, ClaudeCode, Copilot and WebSearch each install separately, so a deployment carries only the ones it uses and drops one by uninstalling rather than by editing a list. This entry is the engine they plug into.
Credential administration stays in Providers & Models, and per-user settings stay in AI Settings. Both read this engine's node types; neither is folded in here, deliberately — a page that renders a key form has different rules from one every signed-in user opens.