# Model providers and the AI switch

Connect Anthropic, OpenAI or a compatible endpoint, choose the model each AI feature uses, price your models and turn every AI call off in one place.

Every AI feature in Tracix, from Brain to Auto-Fix, runs on a model connection you add. This page shows how to connect a provider, choose which model each feature uses, set prices so spend caps work, and switch AI off for the whole workspace.

## How AI features find a model

A model connection is a credential in your workspace: an Anthropic, OpenAI or compatible API key, with an optional base URL. Tracix lists the models each connection serves, so a new model is available the day its provider starts listing it.

A feature picks its model in this order:

1. The model set in the feature's own settings, for the features that have them (Brain, Auto-Fix, workflows and agents, security agents and the SNMP profile generator).
2. The model set for that feature under **Per feature** in AI Settings.
3. The **Preferred model** in AI Settings.

If you never choose a connection, Tracix uses the first Anthropic API credential the person can use. Tracix does not provide a model of its own: until someone adds a credential, AI features have nothing to call.

## Supported providers

| Provider                                                                                         | Credential type                | What you enter                                                      | Model list                                                             | Prices                                                       |
| ------------------------------------------------------------------------------------------------ | ------------------------------ | ------------------------------------------------------------------- | ---------------------------------------------------------------------- | ------------------------------------------------------------ |
| Anthropic                                                                                        | **Anthropic API**              | API key                                                             | Live from Anthropic, with each model's context window and output limit | Built-in table for Claude models                             |
| Anthropic-compatible gateway or proxy                                                            | **Anthropic API**              | API key and the gateway's base URL                                  | Live from the gateway                                                  | Built-in table for Claude models                             |
| OpenAI                                                                                           | **OpenAI API**                 | API key                                                             | Live, chat models only (limits not reported)                           | Set by an admin                                              |
| OpenAI-compatible server, such as Ollama, vLLM, LM Studio, LiteLLM, OpenRouter, Groq or Together | **OpenAI-compatible endpoint** | Base URL (usually ending in `/v1`); API key if the server needs one | Live when the server lists its models; otherwise type the model ID     | Read from OpenRouter and Together; otherwise set by an admin |

> **Your endpoint must be reachable from your workspace**
>
> Your workspace calls the provider from Tracix's cloud, so a model server you run must be reachable from there over the internet. Tracix blocks base URLs that resolve to private, loopback or link-local addresses. If a server does not list its models, type the model ID instead.

## Connect a provider

**1. Create the credential**

Open **Credentials** in the console sidebar. In the **New Credential** panel, pick the **Type** (**Anthropic API**, **OpenAI API** or **OpenAI-compatible endpoint**), give it a **Name** and fill in the API key or base URL. A new credential is private to you: select **Share with the team** if other people's AI features should use it. Then select **Save credential**. Secret values are encrypted at rest and never shown again. Creating credentials needs permission to manage them, which admins have.

**2. Choose it as the preferred model**

Open **AI & Automation › AI Settings** and select the **Models** tab. Under **Preferred model**, pick the connection, then pick a model from its list or type a model ID.

**3. Test the model**

Select **Test**. Tracix runs one short Brain turn on the model and shows how long it took to answer, or the provider's error.

**4. Save**

Select **Save models**. Every AI feature without a model of its own now uses this one.

## Choose a model per feature

Under **Per feature**, give a feature its own model, for example a smaller model for drafting SQL. A feature left empty uses the preferred model.

| Feature                                     | Where it appears  |
| ------------------------------------------- | ----------------- |
| Natural-language queries                    | Metrics Explorer  |
| Draft a monitor from a description          | Monitors          |
| Draft alert SQL                             | Alerts            |
| Alert sweep findings                        | Alerts            |
| Draft an anomaly job                        | Anomaly detection |
| Generate a log pipeline                     | Log pipelines     |
| Natural language to config, and vendor docs | Config Management |
| Workflow planning in Brain chat             | Tracix Brain      |
| Recommendation summaries                    | Databases         |

These features keep their model in their own settings, and AI Settings links to each one under **Set in the feature's own settings**:

* **Brain chat**: the Brain profile, on the **Brain** tab of AI Settings. See [Brain](https://tracix.io/docs/ai/brain).
* **Auto-Fix**: **Auto-Fix › Policies**. See [Workflows, agents and Auto-Fix](https://tracix.io/docs/ai/workflows).
* **Workflows and agents**: the **Chat Model** of each AI Agent node.
* **Security agents**: **Security › Agent Governance** can pin a model for each agent.
* **SNMP profile generator**: its own author and reviewer models.

## Add a fallback model

Under **Fallback model**, pick a model for when the main one is rate limited, overloaded or not answering. Brain retries the main model briefly, then answers with the fallback and says so in the reply. A fallback on a different connection keeps Brain working through one provider's outage.

## Set prices for spend caps

Spend caps and usage reports need a price for every model. Tracix reads prices from OpenRouter and Together, and prices Claude models from a built-in table. For any other model, add the price yourself.

Under **Prices**, select **Add price** and enter the input, output and cached input price in US dollars per million tokens for a connection. Leave the model empty to price every model on that connection that has no row of its own. Your price wins over a listed or built-in price, so use it for negotiated rates and for models you run yourself.

> **Unpriced is not free**
>
> A model with no price counts as unpriced, not as costing nothing. Add a price for every model you use, so spend caps such as Brain's daily budget can do their job.

## Turn AI off for the workspace

The **Usage** tab holds the master switch, **AI usage across Tracix**. It is on in a new workspace. When you turn it off, no model or embedding provider is called anywhere in the console until you turn it back on.

**1. Open the Usage tab**

Open **AI & Automation › AI Settings**. The **Usage** tab opens first and shows whether AI is **ON** or **OFF**.

**2. Turn it off**

Select **Turn off AI usage**, then confirm with **Turn off AI**. Changing the switch requires an admin.

| Area               | What happens while AI is off                                                                                                                    |
| ------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------- |
| Workflows & Agents | AI Agent and chain steps are skipped. Runs finish as partial, with a note on the step.                                                          |
| Brain & Auto-Fix   | Brain stops answering, and Auto-Fix starts no fix a model would write. Reverts need no model, and fixes already awaiting approval keep waiting. |
| Natural language   | Metrics NL→SQL, Config Management NL→config and log pipeline generation reply that AI is off.                                                   |
| Recommendations    | Database recommendations fall back to their summary without AI.                                                                                 |
| Embeddings         | Remote embedding providers are blocked; the local embedder keeps working.                                                                       |
| Anomaly detection  | Not affected. Anomaly jobs are statistical and make no model calls.                                                                             |

## Limit what agents can do

Two more tabs in AI Settings apply to every agent in the workspace:

* **Tools** lists the tools agents can use, grouped by category, such as code execution, web search, web fetch and Brain's own tools. Clear a tool's checkbox to take it away from every agent, or use **Enable all** and **Disable all**.
* **Skills** holds instructions agents can load, written as a `SKILL.md`. Tracix includes a few, such as incident triage and safe code changes. Select **Create skill**, or **Import** one from a URL or pasted text, and switch each on or off.

Only admins, and custom roles with permission to manage AI settings, can change these tabs. Everyone else sees them read-only.

## Next steps

- [Ask Brain a question](https://tracix.io/docs/ai/brain): What Brain reads, what it sends to your model and how it asks before changes

- [Build a workflow](https://tracix.io/docs/ai/workflows): AI agents, approvals and Auto-Fix

- [Users and roles](https://tracix.io/docs/admin/users-and-roles): Decide who can manage AI settings