Skip to content

Models and Providers ​

The AI chat uses a model you connect: a coding agent you already have, an API key from a provider, your OpenPencil Cloud account, or your own server. With an API key or your own server, requests go straight from the app to the provider, and a coding agent uses its own connection. OpenPencil Cloud's AI goes through Cloud's own service, which forwards each request to the model provider it uses.

Add a model from Settings → Models, or let guided setup do it.

Providers ​

ProviderModelsHow to connect
AnthropicClaude Sonnet 5.5, Claude Opus 5.5, Claude Fable 5.1, Claude Haiku 5.5API key
Anthropic-compatibleAny model it servesYour server's base URL and key
ChatGPTThe models your plan offers, such as GPT-6.1 Sol, GPT-6 Astra, GPT-6 LunaSign in, in the desktop app
DeepSeekDeepSeek Flash, DeepSeek V4 ProAPI key
Google AIGemini 3.8 Flash, Gemini 4 Argon, Gemini Flash Latest, Gemini 3.1 ProAPI key
MiniMaxMiniMax-M3.1-Flash, MiniMax-M3API key
OpenAIGPT-6.1 Sol, GPT-6 Astra, GPT-6 LunaAPI key
OpenAI-compatibleAny model it servesYour server's base URL and key
OpenRouterClaude Sonnet 5.5, Claude Opus 5.5, Claude Fable 5.1, Claude Haiku 5.5, GPT-6.1 Sol, Gemini 3.8 Flash, GLM-5.3, DeepSeek V4 Pro, Kimi K3, MiniMax-M3.1-Flash, Gemma 4 31B, or any other model it servesAPI key
Z.aiGLM-5.3, GLM-5.3-Flash, GLM-5.3-FlashX, GLM-5V-Turbo, GLM-5.2, GLM-4.7-FlashAPI key

OpenAI-compatible fits most other services and local servers, with a switch between the Completions and Responses APIs. Anthropic-compatible fits services that speak Anthropic's Messages API. Enter the base URL, the key, and the model ID your server expects.

With OpenRouter you can also sign in instead of pasting a key. ChatGPT uses your ChatGPT plan instead of an API key, and its replies count toward the plan's limits. See AI Chat for how its sign-in works.

In the browser and in the desktop app ​

The desktop app can reach any provider, and keeps keys in your system's credential store. It also offers what the browser can't: coding agents and signing in with ChatGPT.

The web app sends requests from the browser, so the provider has to allow that (CORS). The built-in providers do; for your own server or another service, check before relying on it.

When the web app shows "Could not reach this endpoint from the browser", the request never got an answer the browser would accept. Either the provider doesn't allow browser requests, or it does but leaves the permission off its error responses, so a wrong key or model name looks the same. Try the same settings in the desktop app: if they work there, the provider blocks the browser. Testing a Model Provider shows how to check a provider yourself.

Choosing a model ​

The chat works by calling tools, so a model has to support tool calling, and stream it reliably; models differ here even on the same provider. Add vision to let the chat look at images and check its own work.

Reasoning models can spend their whole output budget thinking. If a model stops without doing anything, raise its Max output tokens, or lower its Thinking level.

Each step of a reply sends the conversation and the tool descriptions again, so long replies use many input tokens. Maximum steps per message in Settings → Models → Chat caps that; see Step limit.

Keys and security ​

In the browser, keys are kept in encrypted browser storage so you don't enter them again; Settings can keep them for the session only instead. The desktop app uses your system's credential store. A key stored in the browser can still be used by scripts running on the page, so prefer scoped keys with spending limits where the provider offers them.