# Skill: Connect a local LLM server to an agent, editor, or script using base URL and a dummy API key

> Apply this skill when wiring a locally running model (Ollama, LM Studio, or similar) into any OpenAI-compatible client: an editor extension, a custom script, or an AI agent framework.

Source: sodigi-learn -- local-agents/podklyuchaem-k-agentu -- https://sodigi.io/learn/local-agents/podklyuchaem-k-agentu

## When to use
- User has a local model running and wants to connect it to their code editor, a browser extension, or an agent.
- User wants to replace a cloud OpenAI call in their script with a free local model call.
- User is getting connection errors when pointing a tool at their local server and needs to debug the address.
- User wants to confirm a local server is alive before writing client code.

## Core rules
- A local server exposes an OpenAI-compatible HTTP endpoint on localhost; any program that speaks the OpenAI API can be pointed at it.
- The connection requires exactly two fields: base URL (the server address with the `/v1` path) and api key (any non-empty dummy string).
- `localhost` (or `127.0.0.1`) means the request stays inside the machine and never touches the internet.
- The port number varies by tool: check what the server printed at startup; do not guess from memory.
- The `/v1` path suffix signals OpenAI-compatible mode; it must appear in the base URL exactly once.
- The api key field is required by client libraries but ignored by a local server; fill it with any string (e.g., "local").
- Cloud-based AI editors route requests through their own servers and cannot see your localhost; only locally installed tools and your own scripts connect directly.
- The model name in the request must match exactly the name the server reports; copy it from the model list endpoint, do not invent it.

## Procedure
1. Start the local server (via `ollama serve`, the LM Studio server toggle, or equivalent) and note the port it prints.
2. Verify the server is alive: open `http://localhost:<port>/v1/models` in a browser; a JSON list of model names should appear.
3. Copy the exact model name from that list.
4. In the client (script, editor extension, agent config), locate the base URL / endpoint field and enter: `http://localhost:<port>/v1`
5. In the api key field, enter any non-empty dummy string (e.g., "local").
6. Set the model field to the exact name copied in step 3.
7. Send a minimal test request (one short user message) and confirm a response arrives.
8. If the tool adds `/v1` automatically, remove it from the base URL to avoid doubling (e.g., use `http://localhost:<port>` instead).

## Ready-to-use prompt
```
Help me connect my locally running model to [tool name / script language].

Local server details:
- Tool running the server: [Ollama / LM Studio / other]
- Port (from server startup output): [e.g., 11434 or 1234]
- Model name (from /v1/models): [e.g., llama3 or mistral:7b]

Please provide:
1. The exact base URL string to put in the client's endpoint field.
2. What to type in the api key field.
3. A minimal working code snippet (or config snippet) for [my language / tool].
4. How to verify the connection is working without running the full agent.
5. A checklist of the three most common mistakes that cause a "connection refused" or 404 error here.
```

## Pitfalls
- Doubling the `/v1` suffix: the client appends it automatically and the base URL already contains it, resulting in `/v1/v1` and a 404 error.
- Using the wrong port: two local servers run on different ports; always read the port from server startup output.
- Leaving the api key field empty: client libraries treat an empty key as invalid and throw an auth error before the request is even sent.
- Starting the client before starting the server: a server that has not been started looks identical to a broken connection.
- Expecting a cloud-hosted editor (where requests go through the vendor's servers) to reach localhost: it cannot; use a locally installed extension or your own script instead.
- Typing a model name from memory: the server's internal name may differ from the display name; always copy it from the `/v1/models` response.