Models / 02.4 / 6

Capabilities

Capabilities describe what a model can do. Scirix uses them to decide which tasks, called roles, a provider is allowed to handle. Setting them correctly matters: a provider missing a required capability is simply not offered for that role.

Where you set them🔗

A preset: capabilities are set automatically by the preset. You do not need to do anything.
Custom: open the Advanced settings disclosure and check the capabilities there.
Your own hardware (Part C): the Capabilities checkboxes in the provider form. The default is chat_completion only. Check more if your model supports them.

Available capabilities🔗

Capability	Meaning
`chat_completion`	The model can hold a conversation and generate text. The baseline for almost everything.
`tool_use`	The model can call tools and functions. Required for all agentic work.
`vision`	The model can understand images.
`json_mode`	The model can be forced to output valid JSON.
`streaming`	The model can stream its output token by token.
`reasoning`	The model supports extended reasoning. Advertised for information; reasoning effort is configured separately per provider.
`embedding`	The model produces embeddings, vector representations of text. Required to build the search index.
`embedding_batching`	The model accepts batched embedding requests, which speeds up indexing.

There is one additional operator-level option, strict_tool_calling, which forces tool-call arguments to match their declared JSON schema exactly. It is an opt-in enforced at the gateway and is not part of the standard checkbox list.

Which roles need which capabilities🔗

Each Scirix role requires certain capabilities. A provider or group can only be assigned to a role if it advertises everything that role needs.

Role	Required capabilities
Coding Agent	`chat_completion` + `tool_use`
Design Agent	`chat_completion` + `tool_use`
Single-Agent	`chat_completion` + `tool_use`
Task Manager	`chat_completion` + `tool_use`
Code Review	`chat_completion` + `tool_use`
Audit Remediation	`chat_completion` + `tool_use`
Goal Shaping	`chat_completion` + `tool_use`
Instant Assistant	`chat_completion`
Batch Summarization	`chat_completion`
Namespace Worker	`chat_completion`
Light Utility	`chat_completion`
Vision Analysis	`chat_completion` + `vision`
Doc/Code Audit	`chat_completion` + `json_mode`
Convention Audit	`chat_completion` + `json_mode`
Embedding	`embedding`

Practical guidance🔗

If you want your model to do real agent work (coding, review, audits), it must have tool_use checked, and the model must genuinely support tool calling.
If you want it to serve audits, it also needs json_mode.
If you want it to build the index, it needs embedding. See The embedding index.
Only check capabilities your model actually has. Checking tool_use on a model that cannot call tools causes failures at run time, not at setup time. In practice this looks like tasks that start and then fail with model errors. See Troubleshooting.