Bring your own model

The default: Typeme Cloud

Typeme Cloud is the zero-configuration option. Requests go from the app to Typeme's Worker, then through Cloudflare AI Gateway to the model. There is no upstream key in the app and nothing for you to sign up for.

• Both the main model and the context-decision model are hosted.
• An anonymous per-install identifier is stored in the local key file. It is used for rate limiting, not authentication, and it is not a secret.
• Requests are not retained for training.

If you want to see what the Cloud path sends, read What is actually sent. The short version: your selected text, the matched store's title and background, recent log excerpts, and your profile.

Switch to your own provider

1. Open Settings › AI (⌘,).
2. Set Provider to Custom.
3. Choose a Service: Anthropic, DeepSeek, OpenAI, magpie, Local (OpenAI-compatible), or Custom (OpenAI-compatible).
4. Paste the API Key for that service.
5. Optionally set Model and API endpoint. Empty fields use the preset defaults.
6. Press Validate. A successful check reports “✓ Works”.

The preset decides the protocol. Anthropic uses the Messages API; everything else uses an OpenAI-compatible Chat Completions API. If a provider speaks OpenAI's format but is not in the list, choose Custom (OpenAI-compatible) and fill in the endpoint and model id.

Where keys are stored

Keys live in ~/.config/Typeme/providers.json, readable only by your user (0600). Typeme does not read keys from environment variables or shell configuration, and it does not put them in the context folder or in logs.

Each service keeps its own key, so you can switch between Anthropic and DeepSeek without re-entering anything. Changing the context folder does not move the key file.

Local models

A local model is the only setup where the prompt does not leave the Mac. Two presets cover it:

• magpie — a local gateway that exposes an OpenAI-compatible API at http://127.0.0.1:3425/v1. The key is the literal string magpie; it is a placeholder, not a secret.
• Local (OpenAI-compatible) — point the API endpoint at your server, for example http://127.0.0.1:11434/v1 for Ollama, then set the model id exactly as the server reports it.

Local models vary widely in instruction-following. Translation and polishing are usually fine; Reply benefits from a larger model. You can set a stronger model for one action without changing the global default.

The two model settings

Typeme uses two models, and they are configured separately:

• Main model — writes the translation, polish, or reply. This is the one you configure under Provider.
• Decision model — chooses which store and which recent entries are relevant. Configure it under Context matching › Decision model.

The decision model has three options:

If you set your main provider to a local model but leave the decision model on Typeme Cloud, the context-selection request still goes to the network. For a fully local setup, set the decision model to Main AI as well, and accept the slower matching.

Per-action overrides

Each action can name its own model. Open Settings › Actions, select an action, and fill in Model. Leave it empty to use the global provider. A common setup:

• Translate — a fast, cheap model.
• Reply — a stronger model, because it has to reason about the discussion.
• Polish — whatever is cheapest that follows style rules.

The override is a model id, not a separate provider. It uses the same Service, key, and endpoint as the global configuration.

Cost and privacy

With your own provider you pay that provider directly for the tokens you use. The local request log records provider, model, latency, token counts, and the request/response pair, so you can see where the spend goes.

Privacy follows the provider you choose. Typeme Cloud and remote providers receive the assembled prompt; a local model receives it on your machine. Screenshots and OCR never leave the Mac either way.

If validation fails

1. 401 or 403 — the key is wrong, expired, or lacks access to the selected model. Re-paste it and validate again.
2. 404 — the endpoint or model id is wrong. For OpenAI-compatible servers the base URL usually ends in /v1.
3. 429 — rate limit or no credit on the provider account. Check the provider's dashboard.
4. Timeout — a local server may not be running, or may be loading the model. Try a request from the provider's own CLI first.
5. Works in validation, fails in the panel — check the log from the menu-bar menu. Some models reject parameters such as temperature; leaving Temperature empty uses the provider default.

Questions