Evaggelos Balaskas - System Engineer

LiteLLM AI Gateway (LLM Proxy)

This project shows a simple pattern: run multiple local model servers, place LiteLLM in front of them, and expose one OpenAI-compatible endpoint for apps to use.

In this setup, Lite LLM sits on http://127.0.0.1:4000/v1 and routes requests to:

• LM Studio on port 1234
• vLLM on port 8000
• Osaurus on port 1337

That gives you one clean API for local testing, model switching, and app integration.

Why this setup is useful

If you already have local models running in different tools, LiteLLM gives you one gateway instead of several different endpoints.

That means you can:

• keep one API base URL
• switch models by alias
• expose multiple local backends behind the same interface
• plug the endpoint into apps that expect an OpenAI-style API

What this article shows

This guide walks through the full flow:

1. LiteLLM exposes /v1/models and returns the three configured aliases.
2. A chat app can select one of those LiteLLM model names and answer normally.
3. Khoj can be pointed at LiteLLM by setting the API base to http://127.0.0.1:4000/v1.
4. Khoj chat models can then use a LiteLLM alias such as ministral-lmstudio.

Quick start

Use Python 3.12 or 3.13 for the LiteLLM virtual environment to avoid uvloop path a known incompatibility with Python 3.14’s asyncio internals.

Set up a virtual environment

virtualenv -p python3.12 venv-litellm/
cd venv-litellm/
source ./bin/activate

pip install "litellm[proxy]"

LiteLLM Configuration

Create config.yaml:

Below is an example based on my local setup.

model_list:
  - model_name: ministral-lmstudio
    litellm_params:
      model: openai/mistralai/ministral-3-3b
      api_base: http://127.0.0.1:1234/v1
      api_key: lmstudio

  - model_name: llama3-vllm
    litellm_params:
      model: openai/mlx-community/Llama-3.2-3B-Instruct-4bit
      api_base: http://127.0.0.1:8000/v1
      api_key: vllm

  - model_name: qwen3-osaurus
    litellm_params:
      model: openai/qwen3.5-0.8b-mlx-4bit
      api_base: http://127.0.0.1:1337/v1
      api_key: osaurus

You need the openai/ prefix for each model to use LiteLLM’s OpenAI-compatible provider.

Start LiteLLM

litellm --config config.yaml --port 4000

If LiteLLM starts correctly, you should see the three model aliases loaded.

Test the proxy

List models:

curl -s http://127.0.0.1:4000/v1/models | jq .

Test chat

Send a chat request:

curl -s http://127.0.0.1:4000/v1/chat/completions
  -H "Content-Type: application/json"
  -d '{
    "model": "ministral-lmstudio",
    "messages": [
      {"role": "user", "content": "Capital of Greece?"}
    ]
  }' | jq .

If everything is wired correctly, LiteLLM will forward the request to the matching backend and return a normal OpenAI-style response.

A typical result looks like this:

{
  "id": "chatcmpl-oukopuooxti6xg92g6qx9b",
  "created": 1773142359,
  "model": "ministral-lmstudio",
  "object": "chat.completion",
  "system_fingerprint": "mistralai/ministral-3-3b",
  "choices": [
    {
      "finish_reason": "stop",
      "index": 0,
      "message": {
        "content": "The capital of Greece is **Athens**.",
        "role": "assistant",
        "provider_specific_fields": {
          "refusal": null
        }
      },
      "provider_specific_fields": {}
    }
  ],
  "usage": {
    "completion_tokens": 10,
    "prompt_tokens": 539,
    "total_tokens": 549
  },
  "stats": {}
}

Connect LiteLLM to Khoj

You can use the same LiteLLM endpoint with any app that supports an OpenAI-style API. In this example, I use Khoj.

Once LiteLLM is running, Khoj only needs one API configuration:

• Name: litellm
• API key: litellm
• API base URL: http://127.0.0.1:4000/v1

Then create a chat model in Khoj using one of the LiteLLM aliases, for example:

• ministral-lmstudio

That is the key idea of this project: Khoj does not need to know whether the model is coming from LM Studio, vLLM, or Osaurus. It only talks to LiteLLM.

Khoj AI model API configuration

Khoj chat model configuration

Khoj using a LiteLLM chat model

That’s it!

If you want to use Claude Code together with Osaurus, there are two different pieces to understand:

1. Model backend — the LLM that answers your prompts
2. MCP tools — the tools Claude Code can call

This is the most important idea:

• Osaurus MCP gives Claude Code access to tools
• Osaurus API can also be used as the model backend, if your setup supports it

These are separate.

Install Claude Code and Osaurus

Let’s start by installing both tools via homebrew on a macbook.

Disclaimer: I like asaurus because it’s small and amazing, I find Ollama big and ugly in macbook.

claude code installation

brew install --cask claude-code

osaurus

brew install --cask osaurus

Open osaurus ui to setup osaurus, in this blog post we will not cover this.

language models

At some point you will download a couple LLMs or SLMs to start with osaurus and you should already have install some tools.

curl -s http://localhost:1337/v1/models | jq .

{
  "data": [
    {
      "id": "llama-3.2-3b-instruct-4bit",
      "created": 1772877371,
      "object": "model",
      "owned_by": "osaurus",
      "root": "llama-3.2-3b-instruct-4bit"
    },
    {
      "id": "qwen3-vl-4b-instruct-8bit",
      "created": 1772877371,
      "object": "model",
      "owned_by": "osaurus",
      "root": "qwen3-vl-4b-instruct-8bit"
    },
    {
      "id": "qwen3.5-0.8b-mlx-4bit",
      "created": 1772877371,
      "object": "model",
      "owned_by": "osaurus",
      "root": "qwen3.5-0.8b-mlx-4bit"
    }
  ],
  "object": "list"
}

status

❯ osaurus status
running (port 1337)

tools

❯ osaurus tools list
osaurus.browser  version=1.2.0
osaurus.fetch  version=1.0.2
osaurus.filesystem  version=1.0.3
osaurus.git  version=1.0.3
osaurus.images  version=1.0.3
osaurus.macos-use  version=1.2.1
osaurus.search  version=1.0.4
osaurus.time  version=1.0.3
osaurus.vision  version=1.0.1

Connect Claude Code to Osaurus via a MCP server

So by default claude code with autostart an interactive configuration setup to connect with your anthropic subscription or with any major ai subscription. We want to override this behaviour to enable claude to connect with osaurus. best way to do that is via an mcp server.

Create ~/.claude.json:

cat > ~/.claude.json <<EOF
{
  "theme": "dark-daltonized",
  "hasCompletedOnboarding": true,
  "mcpServers": {
    "osaurus": {
      "command": "osaurus",
      "args": [
        "mcp"
      ]
    }
  }
}
EOF

This tells Claude Code to start Osaurus as an MCP server.

Note on hasCompletedOnboarding: Setting this to true prevents a startup error where Claude Code tries to connect to Anthropic’s servers before your local endpoint is configured. It is not required for the MCP setup itself, but it avoids a confusing first-run failure.

Note on MCP config location: MCP servers must be defined in ~/.claude.json (or a project-local .mcp.json). Placing them in ~/.claude/settings.json will not work — that file is for environment variables and permissions, not MCP server definitions.

Configure Claude Code to use Osaurus as the model endpoint

Create ~/.claude/settings.json:

mkdir -p ~/.claude/

cat > ~/.claude/settings.json <<EOF
{
  "env": {
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:1337",
    "ANTHROPIC_AUTH_TOKEN": "osaurus",
    "ANTHROPIC_MODEL": "qwen3-vl-4b-instruct-8bit"
  }
}
EOF

This does three things:

• points Claude Code to your local Osaurus server
• authenticates with the local Osaurus endpoint using a static token
• selects the model to use

Note on ANTHROPIC_MODEL vs ANTHROPIC_DEFAULT_SONNET_MODEL: ANTHROPIC_MODEL sets the model directly and is the simpler choice when Osaurus exposes a single model. ANTHROPIC_DEFAULT_SONNET_MODEL overrides only the model Claude Code uses when it internally requests a “sonnet”-class model — useful if you want different models for different internal roles, but unnecessary for a basic local setup.

and

Claude Code requires custom auth token values to be explicitly approved. ANTHROPIC_AUTH_TOKEN is for that

Without this, Claude Code may still prompt for authentication even though your token is set.

Start Claude Code

Run:

claude

Inside Claude Code, you can check your setup with:

/status

Simple mental model

Think of it like this:

• Model = the brain
• MCP = the toolbox

Changing the model does not remove the tools.

That is enough to get started.

Brave’s built-in privacy-first AI assistant, Leo, supports connecting to a local OpenAI-compatible server. This means your conversations never leave your machine — no cloud, no telemetry, just your browser talking to your own model.

This guide uses Osaurus on a MacBook M4 Pro, running the qwen3.5-0.8b-mlx-4bit model as a local example. Any OpenAI-compatible local server (LM Studio, Ollama, llama.cpp, etc.) will work the same way.

About the Model

Qwen3.5-0.8B is Alibaba’s latest small language model, released in March 2026. Despite its compact size, it is a native multimodal model — meaning it supports both text and vision (image understanding) out of the box. It runs efficiently on Apple Silicon via MLX quantization, making it an excellent fit for local inference on a MacBook M4 Pro with minimal RAM usage.

The mlx-4bit suffix means the model weights are 4-bit quantized for Apple Silicon using the MLX framework — fast, low-memory, and runs entirely on-device.

Prerequisites

• Brave Browser installed (check latest version)
• A local LLM server running and reachable at http://localhost:<port>
• Your server responds to POST /v1/chat/completions (OpenAI-compatible API)

Verify your server is working before continuing:

curl -s -X POST http://localhost:1337/v1/chat/completions
  -H "Content-Type: application/json"
  -d '{
    "model": "qwen3.5-0.8b-mlx-4bit",
    "messages": [{"role": "user", "content": "Say hello"}]
  }' | jq .

You should get a JSON response with a choices[0].message.content field. If that works, you’re ready.

example output

{
  "id": "chatcmpl-88053214C2DC",
  "object": "chat.completion",
  "created": 1772783955,
  "model": "qwen3.5-0.8b-mlx-4bit",
  "choices": [
    {
      "finish_reason": "stop",
      "message": {
        "content": "Hello! How can I help you today? 😊",
        "role": "assistant"
      },
      "index": 0
    }
  ],
  "usage": {
    "prompt_tokens": 2,
    "completion_tokens": 8,
    "total_tokens": 10
  }
}

Step 1 — Enable Required Brave Flags

Before Leo can connect to a local server, you need to enable two feature flags in Brave.

Open a new tab and go to:

brave://flags

Search for and enable each of the following:

After enabling both flags, click Relaunch to restart Brave.

Step 2 — Open Leo Settings

Once Brave restarts, open Leo settings by navigating to:

brave://settings/leo-ai

Or open the Leo sidebar (chat bubble icon) → click the Settings gear icon.

Step 3 — Add a Custom Model

In the Leo settings page, scroll down to Bring your own model and click Add new model.

Fill in the fields as follows:

Click Save model.

Note: The server endpoint must be the full path including /v1/chat/completions, not just the base URL.

Step 4 — Select Your Local Model

Back in the Leo chat panel:

1. Click the model selector dropdown (shows the currently active model name).
2. Select the model you just added — e.g. Osaurus.

Leo will now route all requests to your local server.

Step 5 — Start Chatting

Type a message in the Leo input box and press Enter.

How It Works

Your request goes to http://localhost:1337/v1/chat/completions — entirely on your machine. Nothing is sent to Brave’s servers or any external service.

You type in Leo
      │
      ▼
Brave sends POST /v1/chat/completions
      │
      ▼
localhost:1337  (your local server — Osaurus)
      │
      ▼
Model inference on Apple Silicon (MLX / 4-bit quantized)
      │
      ▼
Response streams back to Leo in your browser

No internet required after setup. No data leaves your device.

Tips

• Model name must match exactly what your server reports — check it with:

curl http://localhost:1337/v1/models | jq .

eg.

{
  "data": [
    {
      "object": "model",
      "id": "llama-3.2-3b-instruct-4bit",
      "created": 1772791159,
      "root": "llama-3.2-3b-instruct-4bit",
      "owned_by": "osaurus"
    },
    {
      "object": "model",
      "id": "qwen3.5-0.8b-mlx-4bit",
      "created": 1772791159,
      "root": "qwen3.5-0.8b-mlx-4bit",
      "owned_by": "osaurus"
    }
  ],
  "object": "list"
}

• Leo context features (summarize page, ask about selected text) also work with local models — Leo includes the page content as part of the prompt automatically.
• Since Qwen3.5-0.8B supports vision, with Vision Support enabled you can paste or drag images into Leo and the model will analyze them — all locally.
• Start your local server before opening Brave, or you’ll get a connection error when Leo tries to reach it.

That’s it. You now have a fully local, private AI assistant inside your browser — no accounts, no subscriptions, no data leaving your machine.