- Introduction
- Installation
- Installation through Docker
- Usage
- Setting up Open-WebUI
- Setting up continue.dev
- Contributing
This project simulates the Ollama API using Venice.ai's text generation capabilities, allowing users to access a lifetime Pro account without token fees, usage costs, or subscriptions. The project uses Selenium automation framework to interact with Venice in a clever way, making it suitable for personal inference and integration with IDEs and LLM front-ends.
Beta official API: Venice.ai has launched Beta API access for Pro users, so you don't need this tool anymore. I still can help you get Lifetime Pro account. Learn more about the official API here, including how to set it up with the tools I use.
If you would become a lifetime Pro user on Venice, head over to the offer page, the pro accounts are still on sale.
Please note that currently they are seed-based (not e-mail/password), you need to check the video on how it works.
Since Venice.ai launched their official API, this project is no longer needed. It still works as of November 2024, but I do not plan any other updates and will instead use the official API.
To install the project, follow these steps:
Make sure pip and python commands are Python 3.11 or higher. If you are not sure, run this command:
pip --version
python --version
Especially note if pip and python are different python versions, that might cause problems.
If they are lower version, you might try calling a command with version (python3.11, pip3.11 or python3 and pip3)
Navigate to the project directory using cd ollama-like-venice
Start by first creating and activating a Python virtual environment (optional but recommended):
python -m venv .venv
source .venv/bin/activate
Install the required dependencies
pip install -r requirements.txt
If you would rather install this through Docker, follow these steps:
docker build -t ollama-like-api-for-venice .
docker run -p 9999:9999 -e VENICE_USERNAME="USERNAME" -e VENICE_PASSWORD="PASSWORD" ollama-like-api-for-venice
Note that docker support is currently beta.
First make sure that the virtual environment is activated by running source .venv/bin/activate
in the project's directory.
Then set the environment variables, run the following commands in your terminal:
export VENICE_USERNAME="your-username"
export VENICE_PASSWORD="your-password"
Replace "your-username" and "your-password" with your actual Venice credentials.
If you have a life time pro account based on MOR token in your wallet, you can login using seed instead of username and password:
export VENICE_SEED="abandon abandon abandon abandon abandon abandon abandon abandon abandon abandon abandon about"
Run the server using python ollama_like_server.py
If you have a pro account, I recommend adding --ensure-pro flag, sometimes the pro account is not recognized and thus you don't have access to better models.
If you need to set additional options, see the options:
# python ollama_like_server.py --help
usage: ollama_like_server.py [-h] [--username USERNAME] [--password PASSWORD]
[--host HOST] [--port PORT] [--timeout TIMEOUT]
[--selenium-timeout SELENIUM_TIMEOUT]
[--headless] [--no-headless] [--debug-browser]
[--docker] [--seed SEED] [--ensure-pro]
Ollama-like API for venice.ai
options:
-h, --help show this help message and exit
--username USERNAME Venice username
--password PASSWORD Venice password
--host HOST Local host address
--port PORT Server port
--timeout TIMEOUT Timeout for generating tokens from Venice (seconds)
--selenium-timeout SELENIUM_TIMEOUT
Selenium timeout (seconds)
--headless Run Selenium in headless mode
--no-headless Disable headless mode and run with a visible browser
window
--debug-browser Enable browser debugging logs
--docker Do not run Chrome sandbox (required for docker)
--seed SEED Seed to log in with WalletConnect
--ensure-pro Ensure that Venice recognized the user has a pro
account
If you get WebDriver errors, first make sure that you can run the Chrome binary from the command line.
Replace /path/to/your/google-chrome
with your path to Chrome.
/path/to/your/google-chrome --headless --disable-gpu --dump-dom https://bitcoin.org/ | head -n 20 | grep "Bitcoin"
The output should be something like:
<title>Bitcoin - Open source P2P money</title>
<meta name="description" content="Bitcoin is an innovative payment network and a new kind of money. Find all you need to know and get started with Bitcoin on bitcoin.org.">
If it does not work, use your operating system's package manager to install Chrome and all the required libraries.
Make sure your password does not contain special characters. If you do, change it to password containing letters and numbers only - it can be long. I don't know why, but sometimes Selenium does not type these characters correctly.
Also you can try running ollama-live-venice with --no-headless
flag to see what is happening in the Chrome window.
Now you can use the provided API with Open-WebUI. You can install it according to the instructions, but I think this is simpler and works better:
mkdir open-webui && cd open-webui
python -m venv venv
source venv/bin/activate
pip install open-webui
cd open-webui
source venv/bin/activate
ENABLE_RAG_WEB_SEARCH=true RAG_WEB_SEARCH_ENGINE=duckduckgo OLLAMA_BASE_URL="http://127.0.0.1:9999" open-webui serve --port 8082
Now open-webui should listen on http://127.0.0.1:8082/, visit it through the browser.
If you run local ollama, you can set the URL like this:
cd open-webui
source venv/bin/activate
ENABLE_RAG_WEB_SEARCH=true RAG_WEB_SEARCH_ENGINE=duckduckgo OLLAMA_BASE_URLS="http://127.0.0.1:9999;http://127.0.0.1:11434" open-webui serve --port 8082
I recommend setting the ollama URLs usng the method above, but you can also try using the web interface - it does not always persist restarts though.
After login to the Open-WebUI interface, click on your account (top right icon), then click on Admin Panel, choose the Settings tab, go to Connections. Under the Ollama API either change the URL (if you don't run local Ollama), or click "+" to add new entry and type in http://127.0.0.1:9999 (adjust the port to match your local server configuration, 9999 is just default). If you set it through OLLAMA_BASE_URL above, the it might be already set correctly.
Note: If it does not work, make sure you put http://127.0.0.1:9999 and not localhost, the library used by Open-WebUI sometimes has trouble with resolving, especially if you have both ipv4 and ipv6 host entries for localhost. Also, make sure you are using http, not https.
Then click on the arrows icon, which will verify the connection. If everything is ok, you should see the new models under new chat - the models are "llama-3.1-405b-akash-api", "dolphin-2.9.2-qwen2-72b", "llama-3.2-3b-akash", "nous-theta-web", "nous-hermes3a-web"
The *-web models have access to web search, llama-3.1-405b is the currently best open model. Dolphin is the most uncensored model.
Note that you can also access web through "#" command in open-webui prompt, in this case web search is faciliated by open-webui, not venice.
To use continue.dev to help with your web development needs (or writing README-s like this one😅), install the extension into your Visual Studio Code or JetBrains-based IDE (I haven't tested JetBrains).
Then press the Continue button in the status bar and click on "Configure autocomplete options". An editor will appear. Add the models to models array, for example this is a complete configuration:
{
"models": [
{
"title": "venice llama3.1-405b",
"provider": "ollama",
"apiBase": "http://127.0.0.1:9999",
"model": "llama-3.1-405b-akash-api"
},
{
"title": "venice dolphin-2.9.2-qwen2-72b",
"provider": "ollama",
"apiBase": "http://127.0.0.1:9999",
"model": "dolphin-2.9.2-qwen2-72b"
},
{
"title": "venice llama-3.2-3b-akash",
"provider": "ollama",
"apiBase": "http://localhost:9999",
"model": "llama-3.2-3b-akash"
},
{
"title": "venice nous-theta-web",
"provider": "ollama",
"apiBase": "http://localhost:9999",
"model": "nous-theta-web"
},
{
"title": "venice nous-hermes3a-web",
"provider": "ollama",
"apiBase": "http://localhost:9999",
"model": "nous-hermes3a-web"
}
],
"tabAutocompleteModel": {
"title": "Starcoder",
"provider": "ollama",
"model": "starcoder2:3b"
},
"embeddingsProvider": {
"provider": "ollama",
"model": "nomic-embed-text"
},
"allowAnonymousTelemetry": false,
"docs": []
}
I recommend installing also ollama and running the starcoder2 model locally. For search you will need the embeddings provider and that should also run locally. The embeddings model is used for indexing your codebase and then providing relevant snippets to the model and it needs to run locally through local ollama (not this project, Venice does not provide embeddings API).
Embeddings are very lightweight, so they can run locally on almost any hardware.