Gemini Scribe

Appearance

OpenAI ​

Gemini Scribe can route chat, summary, completions, rewrite, agent tool-calling, and image generation through OpenAI instead of Google Gemini. Text and agent requests use the OpenAI Chat Completions API; image generation uses the dedicated Images API. Use this with OpenAI Platform billing, or point the plugin at an OpenAI-compatible server — LM Studio, an MLX-served endpoint, Ollama's own OpenAI-compatible endpoint, or similar — running locally or on your network.

This is API-key billing only: there is no "Sign in with ChatGPT" / ChatGPT-subscription (Codex-style) authentication. You need an OpenAI API key, or a placeholder key for a compatible server that doesn't check one.

Setup ​

  1. Get an API key — Visit platform.openai.com/api-keys, create a key, and copy it. If you're only targeting a local compatible server that doesn't validate keys, you can skip this and use any placeholder value instead — the provider still requires a key to be set.
  2. Open the OpenAI card — Open Settings → Gemini Scribe → Providers, then select OpenAI.
  3. Enter your API key — On the API key row, click "Link..." and paste your key (or your placeholder value, for a compatible server). It's stored securely using Obsidian's SecretStorage, the same as the Gemini key.
  4. Route features to it — Open the Features page and set Chat and agent (and anything else you want) to OpenAI. Set Default provider on the Providers page to OpenAI too if you want it to catch everything you haven't routed elsewhere.
  5. Pick models — Chat, summary, completions, rewrite, and image generation each get their own model dropdown on the Features page, populated from GET <OpenAI base URL>/models. Unlike Ollama, OpenAI has no single-resident-model constraint, so picking a different model per feature costs nothing extra. Click Refresh on the OpenAI provider card if a model doesn't show up.

If you're using the real OpenAI API, that's it — the base URL defaults to https://api.openai.com/v1 and you're done.

Using an OpenAI-compatible server ​

Point the OpenAI card's Base URL field at your server instead of the default, and the plugin talks to it the same way it talks to OpenAI.

Example with LM Studio:

  1. Start LM Studio's local server (Developer tab → Start Server). By default it listens on http://localhost:1234/v1.
  2. In Gemini Scribe, set the OpenAI card's Base URL to http://localhost:1234/v1.
  3. LM Studio doesn't check API keys, but the plugin still requires the API key field to be non-empty — enter any placeholder value (e.g. lm-studio).
  4. Click Refresh on the OpenAI provider card to pull in whatever model you have loaded in LM Studio.

The same pattern works for MLX-served endpoints, Ollama's OpenAI-compatible endpoint (http://localhost:11434/v1), or any other server that implements the /v1/chat/completions and /v1/models endpoints. Image generation additionally requires a compatible /v1/images/generations endpoint that returns b64_json; most chat-only local servers do not provide one.

A compatible server's /models catalog is taken at face value — the supported-model allowlist described below applies only to api.openai.com.

Models ​

  • api.openai.com — the endpoint advertises around ninety model ids, most of which this integration can't drive usefully (Responses-API-only, audio, embeddings, older families). The text dropdowns are therefore limited to the three GPT-5.6 models that have been validated against the plugin: gpt-5.6-sol (chat default), gpt-5.6-terra (summary default), and gpt-5.6-luna (completions default). All three accept 922,000 input tokens, which is what the token counter above the message box reflects. The image-generation dropdown offers gpt-image-2.5-flare (default; fast everyday generation) and gpt-image-2.5-sunburst (higher-precision editing and fidelity).
  • Where those numbers come from — OpenAI's /v1/models returns no capability metadata (just an id and a creation date), so context windows are curated in the plugin rather than discovered. The 922,000 figure was measured directly against the live API, not taken from documentation.
  • Compatible servers — an unrecognized model id gets a conservative default: 128k context window, no vision, tool calling assumed. Because /models does not report whether a model supports image generation, these models appear in both the text and image-generation dropdowns; the image picker marks their capability as unreported. Selecting one for image generation is an explicit opt-in and succeeds only if the server implements /v1/images/generations for that model.

Reasoning and tool calling on GPT-5.6 ​

The GPT-5.6 models are reasoning models, which constrains what Chat Completions accepts:

  • Sampling settings are ignored. These models only accept the default temperature and top-p, so the plugin omits both rather than sending a value — moot in practice, since the plugin no longer exposes temperature/top-p settings for any provider.
  • Tool calling runs with reasoning disabled. /v1/chat/completions rejects function tools for these models unless reasoning effort is set to none, so any request carrying tools (all agent-mode turns) pins it there. Agent mode works normally; you just won't get reasoning output alongside tool use.

What works ​

  • Agent chat with streaming, tool calling, and conversation memory
  • Drag-and-drop / paste of image attachments to vision-capable models (auto-detected for known OpenAI models; off by default for unrecognized/compatible-server models)
  • File summarization, IDE-style completions, selection rewriting
  • Image generation from the command palette or generate_image agent tool, saved as PNG in the same background-task flow used by Gemini
  • Custom prompts, projects, agent skills, scheduled tasks, MCP servers
  • Retry with exponential backoff (fixed policy, not configurable) and streaming responses, identical to the Gemini and Ollama providers

What does not work ​

Google Search, Google Maps, URL Context (web fetch), Deep Research, and the vault search index still depend on Gemini cloud services and are unavailable on OpenAI. See the Provider Capabilities reference for the full matrix.

OpenAI image generation is one-shot text-to-image generation through /v1/images/generations; image editing and multi-turn image workflows are not exposed in Gemini Scribe. OpenAI may require API organization verification before GPT Image models can be used.

OpenAI's Chat Completions API also doesn't return model "thinking" the way Gemini does, so no reasoning is shown when talking to api.openai.com. A compatible server that emits a reasoning_content field on its responses will have that content shown as thoughts.

Mixed setup: OpenAI chat and images, Gemini extras ​

You don't have to choose all-or-nothing. On the Features page, each feature can be pointed at a different provider — so you can keep chat and image generation on OpenAI while still using Gemini's web search, deep research, or vault index.

To do that:

  1. Add the OpenAI card and route Chat and agent to it, as above.
  2. Add the Gemini card too (Settings → Gemini Scribe → Providers), with your Gemini API key. It's needed by any feature routed to Gemini, even though chat stays on OpenAI.
  3. Open the Features page and route each cloud feature where you want it — for example Web search to Gemini and Image generation to OpenAI.

Nothing moves between providers unless you move it — a feature routed to a provider that can't serve it, or isn't connected, stays off; the plugin never silently substitutes another provider. See Provider Capabilities for the full privacy semantics.

Tips ​

  • Vision model detection — Known OpenAI model ids report vision support from a curated metadata table; models from a compatible server default to no vision since the plugin has no way to know the server's capabilities. If your server serves a vision-capable model and attachments aren't working, this is why.
  • Tool calling — All offered OpenAI chat models support function calling; on GPT-5.6 models it runs with reasoning disabled (see above). Dedicated GPT Image models use the Images API instead. Compatible-server text models default to "tool calling assumed" — if a model genuinely doesn't support it, tool calls will fail rather than being pre-filtered.
  • A model you want isn't listed — On api.openai.com only the validated GPT-5.6 and GPT Image models above are offered. To use another model, point the base URL at a proxy that serves it; catalogs from non-api.openai.com URLs aren't filtered. Unrecognized compatible-server models remain available for text and are also offered as explicit image-generation opt-ins because /v1/models does not report image-generation capability.
  • Placeholder API keys — Required even when your server doesn't check them; the provider won't initialize with an empty key field.
  • Refreshing the model list — Click Refresh on the OpenAI provider card after changing the base URL, loading a different model in a local server, or whenever a dropdown looks stale.