LLM Model

Subdirectory: llm_models/ · Body: must be empty Declares a model in the catalog agents can pick as an engine, matching what you’d otherwise configure at Settings → Models. See Models for the concepts behind each field. Models are always tenant-scoped - there is no org field.

Frontmatter

Field	Type	Required	Default	Description
`provider`	string	Yes	-	The owning Provider’s `key`.
`key`	string	Yes	-	Display name shown in Command Center and referenced by agents when selecting an engine.
`model_id`	string	Yes	-	Must match the model’s deployment name on the provider exactly.
`context_length`	integer	Yes	-	Maximum tokens the model accepts.
`enabled`	boolean	No	`true`	Only enabled models appear as selectable engines.
`input_modalities`	array	No	-	`text` and/or `image`. Include `image` for vision-capable models.
`output_modalities`	array	No	-	`text` and/or `image`.
`provider_options`	object	No	-	Provider-specific passthrough options.
`config`	object	No	-	Sampling parameters - see Sampling config.

Sampling config

All fields under config are optional:

Key	Type	Description
`temperature`	number	Sampling temperature
`topP`	number	Nucleus sampling threshold
`topK`	number	Top-K sampling cutoff
`presencePenalty`	number	Penalize tokens already present
`frequencyPenalty`	number	Penalize tokens by frequency
`maxOutputTokens`	number	Cap on generated tokens
`seed`	number	Deterministic sampling seed
`stopSequences`	string[]	Sequences that stop generation
`reasoning`	string	Thinking-effort level: `provider-default`, `none`, `minimal`, `low`, `medium`, `high`, `xhigh` - see Thinking effort

Fields NOT supported

id, tenant_id, provider_id, created_at, updated_at - system-set
org - models are always tenant-wide
Pricing plans and performance-test results - configured in Command Center, not part of the CLI schema

Example

llm_models/gpt-5.mdx

---
provider: openai
key: gpt-5
model_id: gpt-5
context_length: 400000
enabled: true
input_modalities: [text, image]
output_modalities: [text]
config:
  reasoning: high
  maxOutputTokens: 32000
---

Gotchas

Body must be empty. Everything is frontmatter.
provider is a name, resolved at sync time. Referencing a provider declared in the same directory works even before that provider exists on the platform yet.
A model assigned to any agent can’t be pruned. Reassign those agents (or archive them) first, or sync with enabled: false instead - see Deletion rule.

CLI

Deploy

Configure

Frontmatter

Sampling config

Fields NOT supported

Example

Gotchas

​Frontmatter

​Sampling config

​Fields NOT supported

​Example

​Gotchas

Frontmatter

Sampling config

Fields NOT supported

Example

Gotchas