Configuration
This section is the field-by-field reference for everything you can configure in the bundle. For the mechanism that loads and merges configuration — layers, export endpoint, reload paths — see Architecture → Configuration System. Terminology is defined in Core Concepts.
Where to go
| I want to… | Read |
|---|---|
| Set the bundle up in a fresh project | Installation |
| Create or edit an agent | Agents |
| Define named inference providers (byok, github, Ollama, HF, …) | Inference Providers |
| Create a custom skill | Skills |
| Connect an agent to MCP tool servers | MCP Servers |
| Expose project files to an agent for reading | File Staging |
| Configure agent template generation | Agent Templates |
| See the catalog of SDK tools | Tool Security Reference |
| Tune env vars, timeouts, provider tokens | Environment Variables |
| Use a strict OpenAI-compatible provider (Cerebras, Ollama, vLLM, …) | Provider Compatibility |
| Ship preset agents or skills from my own bundle | Shipping Presets |
The merge model at a glance
Every agent and every skill comes from one of three layers. Higher layers override lower layers by name — full-object replacement, no field-level merging.
| Layer (highest wins) | Who manages it | Where it lives |
|---|---|---|
| Studio UI edits | Administrators | SettingsStore (default) or var/config/agent/*.yaml |
| Container config | Project code | pimcore_agent.agents.definitions, auto-prepended YAML |
| Bundle presets | Bundle authors | pimcore_agent.agents.paths / skills.paths in bundles |
In practice: a bundle ships sensible defaults; a project overrides any of them from code; an administrator
tweaks anything further at runtime from the Studio UI without touching code. Upgrading a bundle never clobbers an
administrator's edits. Every entry in the Studio UI carries a bundle or user source badge.
Full mechanics — export endpoint, reload paths, skill materialization — in Architecture → Configuration System.