GPTbot is a simple bot that uses different APIs to generate responses to messages in a Matrix room.
gpt-3.5-turbo
and gpt-4
, gpt-4o
, whisper
and tts
) and compatible APIs (e.g. ollama
)dall-e-2
/dall-e-3
models!gptbot calculate
command
!gptbot help
and !gptbot coin
To run the bot, you will need Python 3.10 or newer.
The bot has been tested with Python 3.12 on Arch, but should work with any current version, and should not require any special dependencies or operating system features.
The recommended way to install the bot is to use pip to install it from PyPI.
# Recommended: activate a venv first
python -m venv venv
. venv/bin/activate
# Install the bot
pip install matrix-gptbot[all]
This will install the latest release of the bot and all required dependencies for all available features.
You can also use pip install git+https://git.private.coffee/privatecoffee/matrix-gptbot.git
to install the latest version from the Git repository.
A docker-compose.yml
file is provided that you can use to run the bot with
Docker Compose. You will need to create a config.ini
file as described in the
Running
section.
# Clone the repository
git clone https://git.private.coffee/privatecoffee/matrix-gptbot.git
cd matrix-gptbot
# Create a config file
cp config.dist.ini config.ini
# Edit the config file to your needs
# Initialize the database file
sqlite3 database.db "SELECT 1"
# Optionally, create Pantalaimon config
cp contrib/pantalaimon.example.conf pantalaimon.conf
# Edit the Pantalaimon config file to your needs
# Update your homeserver URL in the bot's config.ini to point to Pantalaimon (probably http://pantalaimon:8009 if you used the provided example config)
# You can use `fetch_access_token.py` to get an access token for the bot
# Start the bot
docker-compose up -d
WARNING: Using end-to-end encryption seems to sometimes cause problems with file attachments, especially in rooms that are not encrypted, if the same user also uses the bot in encrypted rooms.
The bot itself does not implement end-to-end encryption. However, it can be used in conjunction with pantalaimon.
You first have to log in to your homeserver using python fetch_access_token.py
,
and can then use the returned access token in your bot's config.ini
file.
Make sure to also point the bot to your pantalaimon instance by setting
homeserver
to your pantalaimon instance instead of directly to your
homeserver in your config.ini
.
Note: If you don't use pantalaimon, the bot will still work, but it will not be able to decrypt or encrypt messages. This means that you cannot use it in rooms with end-to-end encryption enabled.
Clone the repository and install the requirements to a virtual environment.
# Clone the repository
git clone https://git.private.coffee/privatecoffee/matrix-gptbot.git
cd matrix-gptbot
# If desired, activate a venv first
python -m venv venv
. venv/bin/activate
# Install the bot in editable mode
pip install -e .[dev]
# Go to the bot directory and start working
cd src/gptbot
Of course, you can also fork the repository on GitHub and work on your own copy.
Generally, the main
branch is considered unstable and should not be used in
production. Instead, use the latest release tag. The main
branch is used for
development and may contain breaking changes at any time.
For development, a feature branch should be created from main
and merged back
into main
with a pull request. The pull request will be reviewed and tested
before merging.
The bot requires a configuration file to be present in the working directory.
Copy the provided config.dist.ini
to config.ini
and edit it to your needs.
The bot can then be run with python -m gptbot
. If required, activate a venv
first.
You may want to run the bot in a screen or tmux session, or use a process
manager like systemd. The repository contains a sample systemd service file
(gptbot.service
) that you can use as a starting point. You will need to
adjust the paths in the file to match your setup, then copy it to
/etc/systemd/system/gptbot.service
. You can then start the bot with
systemctl start gptbot
and enable it to start automatically on boot with
systemctl enable gptbot
.
Analogously, you can use the provided gptbot-pantalaimon.service
file to run
pantalaimon as a systemd service.
Once it is running, just invite the bot to a room and it will start responding
to messages. If you want to create a new room, you can use the !gptbot newroom
command at any time, which will cause the bot to create a new room and invite
you to it. You may also specify a room name, e.g. !gptbot newroom My new room
.
Note that the bot will respond to all messages in the room by default. If you
don't want this, for example because you want to use the bot in a room with
other people, you can use the !gptbot roomsettings
command to change the
settings for the current room. For example, you can disable response generation
with !gptbot roomsettings always_reply false
.
With this setting, the bot will only be triggered if a message begins with
!gptbot chat
. For example, !gptbot chat Hello, how are you?
will cause the
bot to generate a response to the message Hello, how are you?
. The bot will
still get previous messages in the room as context for generating the response.
The bot has a selection of tools at its disposal that it will automatically use to generate responses. For example, if you send a message like "Draw me a picture of a cat", the bot will automatically use DALL-E to generate an image of a cat.
Note that this only works if the bot is configured to use a model that supports
tools. This currently is only the case for OpenAI's gpt-3.5-turbo
model. If
you wish to use gpt-4
instead, you can set the ForceTools
option in the
[OpenAI]
section of the config file to 1
. This will cause the bot to use
gpt-3.5-turbo
for tool generation and gpt-4
for generating the final text
response.
Similarly, it will attempt to use the gpt-4-vision-preview
model to "read"
the contents of images if a non-vision model is used.
There are a few commands that you can use to explicitly call a certain feature
of the bot. For example, if you want to generate an image from a text prompt,
you can use the !gptbot imagine
command. For example, !gptbot imagine a cat
will cause the bot to generate an image of a cat.
To learn more about the available commands, !gptbot help
will print a list of
available commands.
The bot supports voice input and output, but it is disabled by default. To
enable it, use the !gptbot roomsettings
command to change the settings for
the current room. !gptbot roomsettings stt true
will enable voice input using
OpenAI's whisper
model, and !gptbot roomsettings tts true
will enable voice
output using the tts
model.
Note that this currently only works for audio messages and .mp3 file uploads.
Help, the bot is not responding!
First of all, make sure that the bot is actually running. (Okay, that's not really troubleshooting, but it's a good start.)
If the bot is running, check the logs, these should tell you what is going on. For example, if the bot is showing an error message like "Timed out, retrying", it is unable to reach your homeserver. In this case, check your homeserver URL and make sure that the bot can reach it. If you are using Pantalaimon, make sure that the bot is pointed to Pantalaimon and not directly to your homeserver, and that Pantalaimon is running and reachable.
If you need help figuring out what went wrong, feel free to open an issue.
Help, the bot is flooding the room with responses!
The bot will respond to all messages in the room, with two exceptions:
!gptbot <command>
.There is a good chance that you are seeing the bot responding to its own messages. First, stop the bot, or it will keep responding to its own messages, consuming tokens.
Check that the UserID provided in the config file matches the UserID of the bot. If it doesn't, change the config file and restart the bot. Note that the UserID is optional, so you can also remove it from the config file altogether and the bot will try to figure out its own User ID.
If the User ID is correct or not set, something else is going on. In this case, please check the logs and open an issue if you can't figure out what's going on.
This project is licensed under the terms of the MIT license. See the LICENSE file for details.