# AI Providers

Every control under Settings → AI Providers, including default and per-feature models, providers, model limits, and advanced request settings.

Canonical URL: https://docs.talkbar.app/settings/ai-providers


**Settings → AI Providers** decides which AI model each TalkBar feature uses and where it runs. It has four inner tabs: **AI Features**, **Providers**, **Models** and **Advanced**. For the concepts behind these controls, see [How TalkBar uses AI](/ai/how-talkbar-uses-ai) and [Providers and models](/ai/providers-and-models).




## Page header

These controls sit above the inner tabs and stay visible on every tab.

| Setting | What it does | Default |
| --- | --- | --- |
| **Setup AI** | Opens the guided AI setup, which walks you through providers, local models and feature assignments. | — |
| Issue count (for example **1 issue**) | Appears only when an AI feature is failing. Click it to jump to the problem. | — |
| **Private AI Mode** | Keeps all AI on this computer and turns off any external providers you connected. Shown only after you add a cloud or CLI provider. You can't change it while a session is in progress. See [Private AI Mode](/ai/private-ai-mode). | Off |

When an AI feature fails, a red card appears under the header with the affected provider and model and the latest error. **Retry** tests the connection again. **Clear this error** removes the error for that provider or model.

If an Ollama provider uses a host name, TalkBar can't tell whether it's on this computer, so it asks. Choose **Treat as local** or **Keep as remote**. Trusted hosts count as local in Private AI Mode; otherwise the feature shows a "Not local" warning, but nothing is blocked. You can change this later with **Treat host as local** when you edit the provider.

## AI Features

The **AI Features** tab (search entry: **AI Features**) shows which model powers each feature.

| Setting | What it does | Default |
| --- | --- | --- |
| **Test** | Sends a short test request for every enabled feature and lists the result, the model used and the response time. | — |
| **Quick Apply Wizard** | Opens the AI setup on its feature step, where you can assign models to features in one pass. | — |

### Default model

**Default model** is used by every AI feature you haven't customized. It has two dropdowns. The one in use right now is marked **Active**.

| Setting | What it does | Default |
| --- | --- | --- |
| **Standard** | The model features use while Private AI Mode is off. Any model from any provider. | Qwen3.5 4B (Built-in AI) |
| **Private AI** | The model features use while Private AI Mode is on. Only models that run on this computer are listed. Disabled until you add a cloud or CLI provider, because until then all AI already runs locally. | Qwen3.5 4B (Built-in AI) |

If you haven't added a cloud or CLI provider yet, a **Set up a provider** button takes you to the **Providers** tab.

### Per-feature models

**Per-feature models** lets individual features use their own model instead of the default. Click **Customize** to show the controls (and **Hide** to collapse them). A **Standard** / **Private AI** switch picks which set of assignments you're editing. The two sets are kept separately.

Features are grouped. Each group row has an on/off switch, a model dropdown and a **Reset** button. **Show features** expands a group so you can set each feature on its own.

| Group | Features in the group | Default |
| --- | --- | --- |
| **Chat & Ask AI** | **Chat replies**, **Ask AI** (note text from a prompt) | On, uses the default model |
| **Summaries** | **Voice summaries**, **Timeline** (live timeline bullets) | On, uses the default model |
| **Dictation & voice commands** | **Dictation polish**, **Command interpretation** | On, uses the default model |
| **Knowledge & titles** | **Knowledge extraction**, **Title generation** | On, uses the default model |

| Control | What it does | Default |
| --- | --- | --- |
| Group or feature switch | Turns that AI feature on or off. | On |
| Model dropdown | Picks a model for the group or feature. **Use default** follows the default model. Only models that can do the job are listed. | **Use default** |
| **Reset** | Returns the group or feature to the default model. | — |

A badge next to each feature shows its status: **Ready**, **Disabled**, **Invalid** or **Needs config**. **Customized** or **Inherited** tells you whether a group has its own model. **Model not downloaded** means the chosen local model still needs to be downloaded on **Settings → Private Local AI**.

In the Private AI set, **Dictation polish** starts out assigned to the small Qwen3 0.6B model instead of the default.

## Providers

The **Providers** tab (search entry: **Providers**) lists every AI provider TalkBar can use. **Built-in AI** is always there and runs on this computer.

| Setting | What it does | Default |
| --- | --- | --- |
| **Set up via wizard** | Opens the guided AI setup on its providers step. | — |
| **Add Provider** | Opens **Add AI Provider**, where you connect a cloud API or a command-line AI tool. | — |

### Each provider

Each provider card shows its type (**API / BYOK**, **Installed CLI** or **Built-in (on-device)**), how many models it has, and warnings such as **No API key**, **Invalid key** or **Validation failed**.

| Control | What it does | Default |
| --- | --- | --- |
| **Test** | Checks that the provider answers. | — |
| **Manage downloads** | Built-in AI only. Opens **Settings → Private Local AI** to download local models. | — |
| **Edit** | Opens **Edit AI Provider** to change the display name, endpoint, API key and, for CLI tools, prompt controls. **Validate Provider** tests the connection; **Save Changes** saves it. | — |
| **Remove** | Deletes the provider and its API key. Features that used it lose their model. | — |
| On/off switch | Turns the provider on or off. Built-in AI can't be turned off. | On |

### Adding a provider

In **Add AI Provider**, choose a **Provider Type**:

- **Cloud API**: pick a provider and enter your API key. The list includes Anthropic, OpenAI, Google Gemini, Ollama (Local), Groq, xAI (Grok), DeepSeek, Cohere, Fireworks AI, Together AI, OpenRouter and AWS Bedrock. **Custom...** lets you enter any compatible endpoint.
- **CLI Tool**: pick an AI command-line tool that's installed on this computer. **Scan** detects installed tools; **Use** fills in a detected one. You can also set the **Executable path** yourself.

**Advanced settings...** shows extra fields such as **Display name** and **Endpoint override (optional)**. Then click **Connect & Select Models**, tick the models you want and add them.

For command-line tools, the edit screen also has **CLI prompt controls**, including **Enable CLI parallel execution**. When it's on, long summaries send up to 3 requests to the tool at once. Turn it off if the tool hits its plan limits. Default: On.

### Local CLI Tools

The **Local CLI Tools** card detects AI command-line tools installed on this computer. Supported tools are Claude Code, Codex, Agy, Grok CLI and OpenCode (marked **Experimental**). Click **Scan** (or **Rescan**) to check again. Each tool shows **Detected** or **Not found**.

## Models

The **Models** tab (search entry: **Models**) lists the models attached to your providers.

| Setting | What it does | Default |
| --- | --- | --- |
| **Add Model** | Attaches a model to an existing provider. Choose the **Provider**, click **Get list** to load its available models (or type a **Remote model key**), set a **Display name** and the **Capabilities** (**Text**, **Vision**, **Structured**), then click **Create Model**. | — |

Each model card shows which features use it and has these controls:

| Control | What it does | Default |
| --- | --- | --- |
| **Test** | Sends a short request and shows how long the model took to answer. | — |
| **Chat budgets** | Built-in models only. Opens **Chat context and generation**. | — |
| **Generation limits** | Built-in models only. Opens **Summary generation limits**. | — |
| **Manage** | Built-in models only. Opens **Settings → Private Local AI** to download or delete the model. | — |
| **Edit** | Other models. Changes the display name, **Context limit** and capabilities. The edit screen also has the **Chat context and generation** and **Summary generation limits** sections. | — |
| **Remove** | Other models. Deletes the model. | — |
| On/off switch | Turns the model on or off. | On |

### Chat context and generation

Sets how much room a model gets for chat. Values are in tokens. The panel first shows read-only figures, such as **Effective chat context**, **Chat working input** and **Reserved generation**. Changes apply to your next message. **Reset to Auto** clears your values.

| Setting | What it does | Default |
| --- | --- | --- |
| **Chat input ceiling** | Caps how much input (sources and conversation) a chat sends. **Custom** lets you enter a **Custom token limit**. | Auto |
| **Chat generation ceiling** | Caps how long an answer can be, including reasoning where the provider counts it. A small ceiling can cut an answer short. | Auto |
| **Ollama context allocation** | Ollama models only. How much context TalkBar asks Ollama to load. Auto requests up to 16K tokens within the model's capacity. **Custom allocation** uses more memory. | Auto |
| **Declared model capacity** | Shown when TalkBar can't verify the model's capacity. Enter the model's context size to confirm it. Leave it blank to use discovered limits or a safe fallback. | Blank (Auto) |

For providers that manage their own limits, the panel shows **Provider-managed limits** instead of these fields.

### Summary generation limits

Sets output limits for session summaries. Changes apply to the next summary. **Reset to Auto** clears your values.

| Setting | What it does | Default |
| --- | --- | --- |
| **Extraction output limit** | Caps the output of each extraction step of a long summary. Auto adapts to segment size, sections and language. | Auto |
| **Final output limit** | Caps the length of the final summary. Auto depends on the summary length. | Auto |
| **Local summary context** | Built-in models only. The context window used for summaries. Choose Auto or a fixed size. | Auto |
| **Reasoning budget** | Built-in models that support it. Caps tokens spent on reasoning. Leave blank for Auto. | Blank (Auto) |

For providers that manage their own limits, **Output limits** and **Reasoning budget** show **Provider-managed**.

## Advanced

The **Advanced** tab holds diagnostics and settings that apply to all AI requests.

### Diagnostics

| Setting | What it does | Default |
| --- | --- | --- |
| **Ollama (Local) — GPU memory** | Shown only when Ollama is installed or running. Lists what Ollama holds in GPU memory. **Unload now** frees the GPU for the built-in model. | — |
| **AI Usage Overview** | A **Feature routing** table that shows each feature, the provider and model it uses, and whether it can run. **Refresh** reloads it. | — |
| **Diagnostics** | Shows each provider's health, how many features are ready, its last validation and recent errors. **Clear issue history** removes past errors. | — |

### Request Settings

| Setting | What it does | Default |
| --- | --- | --- |
| **Redact URLs** | Replaces full web addresses in content with placeholders before it's sent to an AI provider, so tokens and private links don't leave your device. | On |
| **Log requests** | Saves a local copy of every AI request and response for troubleshooting. Logs stay on your device. | Off |

### AI time limits on this computer

Gives built-in and local Ollama models more time on slower computers before a request counts as timed out. Cloud models and command-line tools are not affected.

| Setting | What it does | Default |
| --- | --- | --- |
| **Auto** | Sets the limits from the measured speed of your computer. Until a built-in model has run, Auto doubles the limits. Under the control, TalkBar shows the measured speed for each model. | Selected |
| **Standard** | Keeps the default limits. | — |
| **Longer** | Doubles the limits. | — |
| **Longest** | Allows four times as long. | — |

### Local AI request queue

Runs built-in and Ollama requests one after another, so tasks don't compete for the same GPU.

| Setting | What it does | Default |
| --- | --- | --- |
| **Queue built-in and Ollama** | Turns the queue on. Cloud providers and command-line tools always run outside it. | On |
| **Concurrent local requests** | How many local requests run at the same time (1–8). Use 1 to fully serialize GPU work. | 1 |
| **Pending local requests** | How many requests can wait in line (up to 512). | 32 |

### Response Language

| Setting | What it does | Default |
| --- | --- | --- |
| **Response Language** | The language AI writes in, including chat replies. A chat can override it with **Reply language** in its **Answer using** menu. In Auto, chat replies follow the language of your question. Ask AI in notes keeps the language of the selected text or note and uses this setting only for an empty note. **Custom...** lets you type any language (up to 30 characters). | **Auto (Application UI Language)** |

### Custom AI Instructions

| Setting | What it does | Default |
| --- | --- | --- |
| **Custom AI Instructions** | Extra instructions added to every AI feature, up to 300 characters. They add to the built-in prompts; they don't replace them. | Empty |

### Voice

A shortcut card. **Open Voice settings** takes you to [Settings → Voice](/settings/voice), where recording, transcription, dictation and speaker labels live.





If your organization manages TalkBar, some of these controls may be locked and show a notice. Locked settings keep working as your administrator set them up.


## Related

- [Private Local AI settings](/settings/private-local-ai)
- [Providers and models](/ai/providers-and-models)
- [Private AI Mode](/ai/private-ai-mode)
- [Local models](/ai/local-models)
- [How TalkBar uses AI](/ai/how-talkbar-uses-ai)