Endpoints

Endpoints connect margin to an AI model. You can use local models (completely private, offline) or cloud APIs.

Active Endpoint

Select which endpoint margin should use for AI requests:

• .env Default: Uses settings from your .env file. This is the traditional method and useful if you prefer environment-based configuration.
• Custom endpoints: Any endpoint you've added via the UI.

Adding an Endpoint

1. Fill in the Add New Endpoint form:

   Field	Description
   Name	A label for this endpoint (e.g., "Ollama", "OpenAI", "My Local Model").
   Base URL	The full URL of your API (e.g., http://localhost:11434/v1, https://api.openai.com/v1).
   API Key	Your API key (required for cloud providers, optional for local).
   Model	The model name (e.g., gpt-4o, qwen2.5-coder). Optional -- some endpoints auto-detect.
   Context Window	The maximum context size in tokens (default: 8192).
   Thinking Model	When enabled (default), inline reasoning tags (e.g. <think>, <|channel|>) are stripped from the response. Disable to see the raw output including any tags the model emits.

   When Thinking Model is checked, a Custom Thinking Tags section appears where you can add open/close tag pairs for models with non-standard reasoning tags.

2. Click Test Connection to verify the endpoint works.

3. Click Save Endpoint.

Common Local Endpoints

Provider	Default URL	Default Port
Ollama	http://localhost:11434	11434
LM Studio	http://localhost:1234/v1	1234

Cloud Providers

Provider	Base URL
OpenAI	https://api.openai.com/v1
Anthropic	https://api.anthropic.com/v1
Google Gemini	https://generativelanguage.googleapis.com/v1beta
Grok (xAI)	https://api.x.ai/v1
OpenRouter	https://openrouter.ai/api/v1
Groq	https://api.groq.com/openai/v1

All cloud providers require an API key. You can get one from their respective console or dashboard.

API keys are stored locally in settings.json and never sent anywhere except to the endpoint you configure. No data leaves your machine unless you explicitly connect to a cloud provider.

Reasoning Settings

Many models produce internal reasoning before generating their final output. margin can detect and separate this reasoning text from the visible response.

Thinking Model Toggle

Each endpoint has a Thinking Model checkbox. It controls how inline reasoning tags in the content field are handled:

• On (default): Inline tags like <think>, <\|channel\|>, and custom tag pairs are stripped from the prose output. The reasoning text is captured and shown in the "Thought Process" dropdown instead.
• Off: Tags pass through as raw text — you'll see <think>...</think> or <|channel|>...</channel|> directly in the editor output.

Some models (including LM Studio, DeepSeek, and some OpenAI-compatible providers) use a separate reasoning_content field in the API response rather than inline tags. This reasoning text is always captured in the "Thought Process" dropdown regardless of the toggle — the toggle only controls tag stripping from the main content stream.

Custom Thinking Tags

Different models use different tags to delimit their reasoning. margin supports these out of the box:

Opening	Closing
<|channel|>	<channel|>, <|channel|>, </channel|>, |channel|>
<think>	</think>

If your model uses non-standard tags (e.g., [REASONING]...[/REASONING]), add them in the endpoint form:

1. Open the endpoint for editing.
2. Check Thinking Model.
3. Under Custom Thinking Tags, enter the opening tag (e.g., [REASONING]) and closing tag (e.g., [/REASONING]).
4. Click Add Tag — it appears as a removable chip.
5. Save the endpoint.

Multiple custom tag pairs are supported. Added tags are stripped alongside the built-in defaults.

Prepend Thinking Preamble

When enabled, this injects a system prompt that instructs the model to use reasoning tags before every response. Useful for models that can reason internally but don't do it by default.

Finding Your Model's Tags

If you see raw tags like <think>... or <|channel|>... in your output, the model is using reasoning tags that aren't being filtered. Check your endpoint's Thinking Model toggle and add any missing custom tags. If the toggle is on and you still see tags, add them as custom tags.

See AI Assist > Reasoning & Thinking for how reasoning appears in the editor.

Editing and Deleting Endpoints

• Edit: Click the pencil icon on an endpoint card to modify its settings.
• Delete: Click the trash icon to remove an endpoint.