Provider Capabilities
Gemini Scribe can run on the Google Gemini (cloud), Ollama (models pulled to your machine, plus optional Ollama cloud models), OpenAI (cloud), or Anthropic (cloud) provider. Connect one or more providers on the Providers page, pick a default provider for anything you haven't routed elsewhere, and route individual features to a specific provider on the Features page. Some features depend on Gemini-specific cloud APIs and are unavailable on Ollama, OpenAI, and Anthropic. This page is the single source of truth for what works where; docs/guide/ollama-setup.md, docs/guide/openai-setup.md, and docs/guide/anthropic-setup.md link here instead of duplicating the table.
Capability matrix
| Feature | Gemini | Ollama | OpenAI | Anthropic |
|---|---|---|---|---|
| Chat | ✓ | ✓ | ✓ | ✓ |
| Tool calling (agent mode) | ✓ | ✓ (model-dependent) | ✓ (model-dependent on compatible servers) | ✓ |
| Vision (image attachments) | ✓ | ✓ (model-dependent, auto-detected) | ✓ (model-dependent, auto-detected) | ✓ |
| Scheduled tasks | ✓ | ✓ (inherits the model's tool/vision limits) | ✓ (inherits the model's tool/vision limits) | ✓ |
| Summaries | ✓ | ✓ | ✓ | ✓ |
| Completions | ✓ | ✓ | ✓ | ✓ |
| Rewrite | ✓ | ✓ | ✓ | ✓ |
| RAG / Vault Semantic Search | ✓ | ✗ — tracked in #705 | ✗ | ✗ |
| Image generation | ✓ | ✗ — tracked in #706 | ✓ (GPT Image via the Images API) | ✗ |
| Google Search grounding | ✓ | ✗ | ✗ | ✗ |
| Google Maps grounding | ✓ | ✗ | ✗ | ✗ |
| URL Context (web fetch tool) | ✓ | ✗ | ✗ | ✗ |
| Deep Research | ✓ | ✗ | ✗ | ✗ |
| PDF / audio / video attachments | ✓ | ✗ (images only) | ✗ (images only) | ✗ (images and PDF; no audio/video) |
| Custom base URL | ✓ | ✗ (uses its own ollamaBaseUrl setting) | ✓ (also targets OpenAI-compatible local servers) | ✗ |
OpenAI's row covers both the real api.openai.com endpoint (billed with your own OpenAI API key) and any OpenAI-compatible server reachable at a custom base URL — LM Studio, an MLX-served endpoint, Ollama's own OpenAI-compatible endpoint, and similar. See the OpenAI Setup Guide for both paths. This is API-key billing only — there is no "Sign in with ChatGPT" / ChatGPT-subscription auth.
Anthropic's column covers Claude models on api.anthropic.com, billed with your own Anthropic API key. It has no custom base URL. Offered models report their own context window (1M tokens for all but Haiku 4.5's 200k); 200k is the fallback for a model the list can't identify. Token counts are estimated rather than fetched from a counting endpoint. See the Anthropic Setup Guide.
Features page
On the Features page, each of these is routed independently, shown as a <provider> · <model> row:
| Row | Covers |
|---|---|
| Chat and agent | Interactive chat, agent sessions, scheduled tasks, hooks |
| Summaries | The "Summarize active file" command and conversation compaction |
| Completions | IDE-style inline suggestions |
| Rewrite | Rewriting selected text — has its own model field, no longer borrows the chat model |
| Web search | Google Search grounding and URL Context (web fetch) |
| Deep research | The Deep Research managed agent (no model field of its own) |
| Vault search index | Semantic search across your vault (RAG; no model field of its own) |
| Image generation | Generating images from a prompt |
Opening a row gives exactly two controls: a provider dropdown (listing only providers that actually support that feature, plus an explicit Off) and, where the feature has a model, a model dropdown filtered to that provider. Chat and agent, Summaries, Completions, and Rewrite each list Gemini, Ollama, OpenAI, and Anthropic. Image generation lists Gemini and OpenAI; Web search, Deep research, and Vault search index list Gemini only today.
Google Maps grounding is not on this page. It's provider-bound to Gemini and listed under the Gemini provider card's "Includes" line instead — see Provider-bound grounding below.
This is what makes a mixed setup work: run chat locally on Ollama for privacy and speed, but keep Gemini for web search and image generation — or run chat and image generation on OpenAI while keeping Gemini for Google-backed tools. Each feature makes its own provider call rather than riding on the chat request, so its credential and destination follow that feature's route.
Provider-bound grounding
Google Maps grounding and Gemini's URL-context surface are provider-bound, not routed features: google_maps registers whenever the Gemini provider is configured, regardless of which provider serves any Features row. Its model follows the Web search row's model when Web search is on Gemini, otherwise the bundled Gemini chat default.
Privacy semantics
Nothing is ever routed to the cloud on your behalf. If a feature's routed provider can't serve it, or isn't connected, that feature stays off — the plugin does not quietly substitute another provider. Turning on a cloud feature is always an explicit, per-feature choice, made on the Features page.
Consequences worth knowing:
- Choosing Gemini for a feature means that feature's requests — including any note content they send — go to Google. The Providers page's privacy note names exactly which features are on a non-local provider.
- Choosing OpenAI for a feature means that feature's requests go to
api.openai.com(or whatever base URL you've configured) instead. Choosing Anthropic sends them toapi.anthropic.com. - Vault search index is the broadest of these: enabling it uploads note content to a cloud file-search store, not just the text of a single request.
- Each provider card only needs its own credential: the Gemini card's API key is required only once some feature is actually routed to Gemini (not only when Gemini is your default provider); the OpenAI and Anthropic cards' API key fields work the same way for their providers.
- A hosted provider that needs an API key gets no traffic at all until you add one — not even a model-list request. The OpenAI and Anthropic cards only query
/modelsonce an API key is configured (or, for OpenAI, once the base URL points at a custom server). - Ollama needs no key, so its card always checks the configured Ollama URL (
localhostby default) for pulled models, even if no feature is routed to Ollama. - With every feature on Ollama and every selected model pulled locally, nothing leaves your machine — the Providers page shows Ollama as the only connected provider and the privacy note is empty. Pointing the OpenAI provider's base URL at a local OpenAI-compatible server (LM Studio, MLX, ...) achieves the same thing for OpenAI-routed features — the request never leaves your machine even though the provider is "OpenAI".
- Ollama can also serve cloud-hosted models (
gpt-oss:120b-cloud,deepseek-v4-pro:cloud, …), which reach the daemon like local ones but run on ollama.com. Selecting one for any Ollama-served feature means requests for that feature leave your machine even though the provider is "Ollama" — provider alone no longer implies on-device execution. Detection uses theremote_hostfield Ollama reports for these entries, not the model name. See Cloud models.
Models
Each model dropdown on the Features page lists the models of the provider serving that feature, so a chat-on-Ollama / summaries-on-Gemini setup offers the right models in each row.
Ollama keeps one model resident at a time, so its summary and completions rows default to Same as chat model. Choosing a distinct model there is supported but means Ollama reloads a model on every switch — worth it for a small, fast completions model, rarely worth it otherwise. OpenAI has no such constraint: each feature resolves its own default model when left on "Default for this provider" (chat, summaries, and completions each have a different OpenAI default), and switching between them costs nothing extra. Anthropic works the same way (Opus 5 for chat and rewrite, Sonnet 5 for summaries, Haiku 4.5 for completions).
Notes
- Tool calling — Whether an Ollama model can call tools depends on the model itself; most modern instruct models (Llama 3.2, Qwen 2.5, Mistral 0.3, …) support it, smaller or older models may not. OpenAI's offered chat models support tool calling; its dedicated image models do not use that path. All offered Claude models support tools.
- Vision — Ollama vision support is auto-detected per model from its
/api/showcapabilities (with a template/name-hint fallback for older Ollama versions) — no manual configuration is needed when you pull a new multimodal model. OpenAI vision support is auto-detected from a curated metadata map for known OpenAI model ids; an unrecognized model id (typically one served by an OpenAI-compatible server) defaults to no vision. - Model discovery — The Ollama picker is built from the daemon's
/api/tags, which lists locally pulled models only. A cloud model must be pulled before it appears, even if you already use it from the Ollama app or CLI. The cloud tag takes one of two forms depending on whether the model carries a size tag —ollama pull gpt-oss:120b-cloudbutollama pull glm-5.2:cloud; see Cloud models for the exact syntax per model. The OpenAI picker is built fromGET <openaiBaseUrl>/models, enriched with curated metadata (context window, vision support, and image-generation support) for current OpenAI models; unrecognized ids — including everything from a compatible server — get a conservative text-model default (128k context, no vision, tool calling assumed). Click Refresh on the OpenAI provider card after a compatible server's catalog changes. - RAG, Google Search, Google Maps, URL Context, and Deep Research call Google's cloud APIs directly and require a Gemini API key. Image generation calls either Google or OpenAI according to its feature route. Their agent tools are only registered when the corresponding feature is routed to a provider that supports them and that provider has a key configured on this device — a tool the agent can see but cannot run is worse than one that isn't offered; RAG's indexing service isn't initialized otherwise. What stays visible either way is the command palette and settings UI: the Generate image command, the RAG Pause/Resume/Show status commands, and the Vault search index page remain in place, but invoking one while its feature has no provider shows a notice pointing at the Features page rather than failing the call.
- Scheduled tasks run through the same chat/tool-calling path as interactive agent sessions, so a task that needs vision or tool calling on Ollama or OpenAI is still bound by that model's capabilities.
- Context limits and token counting follow the model in hand, not a global setting — a Gemini model is counted through Google's
countTokensendpoint against a 1M-token window, while a local, OpenAI, or Anthropic model uses a calibrated chars-per-token estimate against that model's own window. In a mixed setup each applies to its own model. - Ollama context window — Resolved per model from the daemon rather than assumed: the runtime allocation from
/api/pswhen the model is loaded, the model's advertised maximum from/api/showbefore that, and a conservative 32k fallback only when the daemon is unreachable. This matters because Ollama's own default window is VRAM-derived (4k under 24 GiB) and often a small fraction of what the model supports — see Set the context window first. - OpenAI reasoning — the Chat Completions API doesn't return model "thinking" the way Gemini and Ollama can, so no reasoning is shown for
api.openai.com. A compatible server that emits areasoning_contentfield on its responses will have that content shown as thoughts, same as the other providers. - Anthropic reasoning — Claude models other than Haiku 4.5 run with adaptive thinking, and a summary of that reasoning is shown as thoughts. Opus 5 and Fable 5.1 also opt into Anthropic's server-side refusal fallback — see the Anthropic Setup Guide.
Changing the default provider or any feature's provider takes effect immediately — no data is lost, and each provider's model choices persist across changes (providerModelMemory restores the model you had the last time a feature was on that provider).
See the Ollama Setup Guide, the OpenAI Setup Guide, and the Anthropic Setup Guide for installation/configuration steps and provider-specific tips.