Brave-BYOM-Proxy

A small proxy script that enables the use of OpenAI's flagship models in Brave Browser's Assistant (Leo) via Bring Your Own Model (BYOM).

Why

The BYOM feature in Brave is not compatible with OpenAI's flagship models (e.g., GPT-5) because it always adds a temperature body parameter in API calls, which the newer models do not support, thus causing a 400 response from OpenAI's API and making it unusable.

The proxy script strips the unsupported parameters, extends the timeout to allow models to reason longer, and adds a few adjustable parameters.

Usage

Run it locally, then point Brave to this proxy instead of OpenAI.

# Run locally.
# Requires uv (astral-sh/uv). It will automatically download dependencies into the .venv directory.
uv run fastapi run main.py

# -- Or --

# Build a docker image
docker build --pull -t brave-byom-proxy .

# Run the docker image
docker run --rm --network host brave-byom-proxy
# Or run on specific host and port
docker run --rm --network host brave-byom-proxy --host 192.168.1.123 --port 8001

When exposing it publicly, make sure to set an access token and serve it over HTTPS.

Configuration is read from environment variables or a .env file in the working directory. The following settings are configurable, with their default values:

BYOMPROXY_UPSTREAM_ENDPOINT=https://api.openai.com/v1/chat/completions

# Timeout (seconds) when requesting the upstream endpoint. Reasoning models can
# take longer to respond, so don't set it too short.
BYOMPROXY_REQUEST_TIMEOUT=300

# Empty by default.
BYOMPROXY_ACCESS_TOKEN=

# Log requests for debugging (or curiosity). Set to 1 to enable.
BYOMPROXY_LOG_REQUEST=

# Set to 1 to disable the generation of conversation titles. All new conversations
# will be titled "New conversation."
BYOMPROXY_DISABLE_TITLE_GEN=

# If set, this model will be used for generating conversation titles.
# A faster model is recommended.
BYOMPROXY_TITLE_GEN_MODEL=gpt-4o-mini

When BYOMPROXY_ACCESS_TOKEN is set, prepend it to the upstream API key with a colon in Brave's settings.

For example, with BYOMPROXY_ACCESS_TOKEN=hello and your OpenAI API key sk-proj-xxxxx, the API key set in Brave has to be hello:sk-proj-xxxxx.

In addition, these API parameters are configurable via query parameters:

• reasoning_effort
• service_tier
• verbosity

For example, when the script is running on localhost:8000, you can configure the server endpoint in Brave like this:

http://localhost:8000/v1/chat/completions?reasoning_effort=high&service_tier=priority&verbosity=low

When provided, these values are injected into the request body.

Please refer to the official API document for details.

Tips

Reasoning Summary

Although OpenAI doesn't return the model's reasoning in the chat completions API, you can add an instruction in the system prompt asking the model to prepend a <think></think> block to its response.

For example:

- **Summarize your thinking:** Always start your response with a <think></think> block. Use it to store a concise summary of your reasoning (e.g., path to conclusion) for your own reference later in the conversation. It will be visible only to you.

Alternatives

Try OpenRouter if an additional third party is acceptable.

With OpenRouter, adding the :online suffix to any model name enables research capabilities (access to tools like web search, calculator, and more), and it's compatible with the v1/chat/completions API.