endpoint-and-model
Every Obsidian plugin that talks to a local model asks you for the same two things: an endpoint and a model name. Once you have a server running (LM Studio or Ollama), this is all that's left — and there's exactly one trap worth knowing about.
The endpoint
The endpoint is just the address of your running server — http://localhost:<port>. The port depends on which server you started:
| Server | Default endpoint |
|---|---|
| LM Studio | http://localhost:1234 |
| Ollama | http://localhost:11434 |
No API key, no account, no internet — everything stays on your machine.
The model name
The model name is the exact name your server knows the model by — e.g. qwen2.5-vl in LM Studio, or llama3.1 / nomic-embed-text in Ollama. The plugin usually fetches the list for you once the server is running; if it shows a dropdown, pick from it.
The one trap: a doubled /v1
These plugins speak the OpenAI protocol, where requests go to …/v1/chat/completions. The plugin adds the /v1 itself. So you enter the base URL without /v1:
- ✅
http://localhost:1234 - ❌
http://localhost:1234/v1→ the plugin would build…/v1/v1/…, which silently returns an empty result with no obvious error.
(Both image-to-markdown and vault-rag normalise a trailing /v1 away to be forgiving — but the safe habit is: base URL only.)
→ Connect a plugin: image-to-markdown · vault-rag