Shell wrapper for OpenAI's ChatGPT, DALL-E, Whisper, and TTS. Features LocalAI, Ollama, Gemini and Mistral integration.
Chat completions with streaming by defaults.
Markdown rendering of chat response (optional).
- Text and chat completions with gpt-4o and gtp-4-vision support
- Text editor interface, Bash readline, and cat input modes
- Markdown rendering support in response
- Preview, and regenerate responses
- Manage sessions, print out previous sessions
- Instruction prompt manager, easily create and set the initial system prompt
- Voice in (Whisper) and voice out (TTS) chat / REPL mode (
options -cczw
) - Integration with awesome-chatgpt-prompts and Chinese awesome-chatgpt-prompts-zh, mudler's LocalAI, Ollama, Google AI, and Mistral AI.
- Tiktoken for accurate tokenization (optional)
- Colour scheme personalisation, and a configuration file
- Stdin and text file input support
- Should™ work on Linux, FreeBSD, MacOS, and Termux.
- Generate images
from text input (
option -i
) - Generate variations of images
- Edit images,
optionally edit with
ImageMagick
(generate alpha mask) - Transcribe audio
from various languages (
option -w
) - Translate audio into English text (
option -W
) - Text-to-speeech functionality (
option -z
)
Bash
cURL
, andJQ
These are required for specific features.
Base64
- Image endpoint, vision modelsImageMagick
- Image edits and variationsPython
- Modules tiktoken, markdown, bs4mpv
/SoX
/Vlc
/FFmpeg
/afplay
- Play TTS outputSoX
/Arecord
/FFmpeg
- Record input (Whisper)xdg-open
/open
/xsel
/xclip
/pbcopy
- Open images, set clipboardW3M
/Lynx
/ELinks
/Links
- Dump URL textbat
/Pygmentize
/Glow
/mdcat
/mdless
- Markdown supporttermux-api
/play-audio
/termux-microphone-record
/termux-clipboard-set
- Termux system
A. Download the stand-alone
chatgpt.sh
script
and make it executable:
wget https://gitlab.com/fenixdragao/shellchatgpt/-/raw/main/chatgpt.sh
chmod +x ./chatgpt.sh
B. Or clone this repo:
git clone https://gitlab.com/fenixdragao/shellchatgpt.git
C. Optionally, download and set the configuration file
~/.chatgpt.conf
:
#save configuration template:
chatgpt.sh -FF >> ~/.chatgpt.conf
#edit:
chatgpt.sh -F
# Or
vim ~/.chatgpt.conf
With command line options -cc
, some properties are set automatically to create a chat bot.
Start a new session in chat mode, and set a different temperature (gpt-3.5 and gpt-4+ models):
chatgpt.sh -cc -t0.7
Create Marv, the sarcastic bot manually:
chatgpt.sh -60 -cc --frequency-penalty=0.5 --temp=0.5 --top_p=0.3 --restart-seq='\nYou: ' --start-seq='\nMarv:' --stop='You:' --stop='Marv:' -S'Marv is a factual chatbot that reluctantly answers questions with sarcastic responses:'
Start new chat session from the "unix.txt" Instruction file.
The commands below are aliases for command line option -S .prompt_name
.
chatgpt.sh -cc -S ..unix
chatgpt.sh -cc -..unix
chatgpt.sh -cc..unix
In this case, the custom prompt will be loaded, and the history will be recorded in the corresponding "unix.tsv" file at the cache directory.
The command below starts a chat session, loads the "unix" instruction, and changes to the defaults "chatgpt.tsv" history.
chatgpt.sh -cc..unix /current
chatgpt.sh -cc -S "..unix" /session current
In the following example, load an older session from the current (defaults) history file.
chatgpt.sh -cc /fork current
chatgpt.sh -cc /fork.
To load an older session from a history file that is different from the defaults, there are some options.
Change to it with command !session [name]
, and then !fork
the older session to the active session.
Or, !copy [orign] [dest]
the session from a history file to the current one
or any other history file.
In these cases, a pickup interface should open to let the user choose the correct session from the history file.
To send an image
/ url
to vision models, start the script and then either
set the image with the !img
chat command with one or more filepaths / URLs
separated by the operator pipe |.
chatgpt.sh -cc -m gpt-4-vision-preview '!img path/to/image.jpg'
Alternatively, set the image paths / URLs at the end of the prompt interactively:
chatgpt.sh -cc -m gpt-4-vision-preview
[...]
Q: In this first user prompt, what can you see? | https://i.imgur.com/wpXKyRo.jpeg
TIP: Run chat command !info
to check model configuration!
DEBUG: Set option -VV
to see the raw JSON request body.
🗣️ Chat completion with Whisper:
chatgpt.sh -ccw
Chat in Portuguese with voice in and out:
chatgpt.sh -cczw pt
When text completions is set for chatting with option -c
,
some properties are configured automatically to instruct the bot.
chatgpt.sh -c "Hello there! What is your name?"
This is the pure text completions endpoint. It is typically used to complete input text, such as for completing part of an essay.
One-shot text completion:
chatgpt.sh "Hello there! What is your name?"
NOTE: For multiturn, set option -d
.
A strong Instruction prompt may be needed for the language model to do what is required.
Set an instruction prompt for better results:
chatgpt.sh -d -S 'The following is a newspaper article.' "It all starts when FBI agents arrived at the governor house and"
chatgpt.sh -d -S'You are an AI assistant.' "The list below contain the 10 biggest cities in the w"
Set option -q
to enable insert mode and add the
string [insert]
where the model should insert text:
chatgpt.sh -q 'It was raining when [insert] tomorrow.'
NOTE: This example works with no instruction prompt! An instruction prompt in this mode may interfere with insert completions.
NOTE: Insert mode
works with model gpt-3.5-turbo-instruct
. This endpoint may deprecate.
To enable markdown rendering of responses, set command line option --markdown
,
or run /md
in chat mode. To render last response in markdown once,
run //md
.
The markdown option uses bat
as it has line buffering on by defaults,
however other software is supported.
Set it such as --markdown=glow
or /md mdless
on chat mode.
Type in any of the following markdown software as argument to the option:
bat
, pygmentize
, glow
, mdcat
, or mdless
.
Unless the chat option -c
or -cc
are set, no instruction is
given to the language model. On chat mode, if no instruction is set,
minimal instruction is given, and some options set, such as increasing
temp and presence penalty, in order to un-lobotomise the bot.
Prompt engineering is an art on itself. Study carefully how to craft the best prompts to get the most out of text, code and chat completions models.
The model steering and capabilities require prompt engineering to even know that it should answer the questions.
Set a one-shot instruction prompt with option -S
:
chatgpt.sh -cc -S 'You are a PhD psycologist student.'
chatgpt.sh -ccS'You are a professional software programmer.'
To create or load a prompt template file, set option -S
with the
operator dot and the name of the prompt as argument:
chatgpt.sh -cc -S.psycologist
chatgpt.sh -cc -S..software_programmer
This will load the custom prompt, or create it if it does not yet exist. In the second example, single-shot editing will be skipped after loading prompt software_programmer.
Set a prompt from awesome-chatgpt-prompts or awesome-chatgpt-prompts-zh, (use with davinci and gpt-3.5+ models):
chatgpt.sh -cc -S /linux_terminal
chatgpt.sh -cc -S /Relationship_Coach
chatgpt.sh -cc -S '%担任雅思写作考官'
-
Run chat commands with either operator
!
or/
. -
Edit live history entries with command
!hist
, for context injection. -
Add operator forward slash
/
to the end of prompt to trigger preview mode. -
One can regenerate a response typing in a new prompt a single slash
/
, or//
to have last prompt edited before new request.
Generate image according to prompt:
chatgpt.sh -i "Dark tower in the middle of a field of red roses."
chatgpt.sh -i "512x512" "A tower."
Generate image variation:
chatgpt.sh -i path/to/image.png
chatgpt.sh -i path/to/image.png path/to/mask.png "A pink flamingo."
In this example, a mask is made from the white colour.
Adding a bat in the night sky.
Generate transcription from audio file. A prompt to guide the model's style is optional. The prompt should match the audio language:
chatgpt.sh -w path/to/audio.mp3
chatgpt.sh -w path/to/audio.mp3 "en" "This is a poem about X."
1. Generate transcription from voice recording, set Portuguese as the language to transcribe to:
chatgpt.sh -w pt
This also works to transcribe from one language to another.
2. Transcribe any language audio input to Japanese (prompt should be in the same language as the input audio language, preferably):
chatgpt.sh -w ja "A job interview is currently being done."
3.1 Translate English audio input to Japanese, and generate audio output from text.
chatgpt.sh -wz ja "Getting directions to famous places in the city."
3.2 Also doing it conversely, this gives an opportunity to (manual) conversation turns of two speakers of different languages. Below, a Japanese speaker can translate its voice and generate audio in the target language.
chatgpt.sh -wz en "Providing directions to famous places in the city."
4. Translate audio file or voice recording from any language to English:
chatgpt.sh -W [audio_file]
chatgpt.sh -W
NOTE: Generate phrasal-level timestamps double setting option -ww
, or option -WW
.
Make sure you have got mudler's LocalAI, server set up and running.
The server can be run as a docker container or a binary can be downloaded. Check LocalAI tutorials Container Images, and Run Models Manually for an idea on how to install, download a model and set it up.
┌───────────────────────────────────────────────────┐
│ Fiber v2.50.0 │
│ http://127.0.0.1:8080 │
│ (bound on host 0.0.0.0 and port 8080) │
│ │
│ Handlers ............. 1 Processes ........... 1 │
│ Prefork ....... Disabled PID ..................1 │
└───────────────────────────────────────────────────┘
1. Download a binary of localai
for your system from Mudler's release GitHub repo.
2. Run localai run --help
to check comamnd line options and environment variables.
3. Set up $GALLERIES
before starting up the server:
export GALLERIES='[{"name":"localai", "url":"github:mudler/localai/gallery/index.yaml"}]' #defaults
export GALLERIES='[{"name":"model-gallery", "url":"github:go-skynet/model-gallery/index.yaml"}]'
export GALLERIES='[{"name":"huggingface", "url": "github:go-skynet/model-gallery/huggingface.yaml"}]'
4. Install the model named phi-2-chat
from a yaml
file manually, while the server is running:
curl -L http://localhost:8080/models/apply -H "Content-Type: application/json" -d '{ "config_url": "https://raw.githubusercontent.com/mudler/LocalAI/master/embedded/models/phi-2-chat.yaml" }'
Finally, when running chatgpt.sh
, set the model name:
chatgpt.sh --localai -cc -m luna-ai-llama2
Setting some stop sequences may be needed to prevent the model from generating text past context:
chatgpt.sh --localai -cc -m luna-ai-llama2 -s'### User:' -s'### Response:'
Optionally set restart and start sequences for text completions
endpoint (option -c
), such as -s'\n### User: ' -s'\n### Response:'
(do mind setting newlines \n and whitespaces correctly).
And that's it!
Model names may be printed with chatgpt.sh -l
. A model may be
supplied as argument, so that only that model details are shown.
NOTE: Model management (downloading and setting up) must follow the LocalAI and Ollama projects guidelines and methods.
For image generation, install Stable Diffusion from the URL
github:go-skynet/model-gallery/stablediffusion.yaml
,
and for audio transcription, download Whisper from the URL
github:go-skynet/model-gallery/whisper-base.yaml
.
If the host address is different from the defaults, we need editing
the script configuration file .chatgpt.conf
.
vim ~/.chatgpt.conf
# Or
chatgpt.sh -F
Set the following variable:
# ~/.chatgpt.conf
OPENAI_API_HOST="http://127.0.0.1:8080"
Alternatively, set $OPENAI_API_HOST
on invocation:
OPENAI_API_HOST="http://127.0.0.1:8080" chatgpt.sh -c -m luna-ai-llama2
Visit Ollama repository, and follow the instructions to install, download models, and set up the server.
After having Ollama server running, set option -O
(--ollama
),
and the name of the model in chatgpt.sh
:
chatgpt.sh -cc -O -m llama2
If Ollama server URL is not the defaults http://localhost:11434
,
edit chatgpt.sh
configuration file, and set the following variable:
# ~/.chatgpt.conf
OLLAMA_API_HOST="http://192.168.0.3:11434"
Get a free API key for Google to be able to use Gemini and vision models. Users have a free bandwidth of 60 requests per minute, and the script offers a basic implementation of the API.
Set the enviroment variable $GOOGLE_API_KEY
and run the script
with option --google
, such as:
chatgpt.sh --google -cc -m gemini-pro-vision
OBS: Google Gemini vision models are not enabled for multiturn at the API side, so we hack it.
To list all available models, run chatgpt.sh --google -l
.
Set up a Mistral AI account,
declare the enviroment variable $MISTRAL_API_KEY
,
and run the script with option --mistral
for complete integration.
This project PKGBUILD is available at the Arch Linux User Repository (AUR) to install the software in Arch Linux and derivative distros.
To install the programme from the AUR, you can use an AUR helper
like yay
or paru
. For example, with yay
:
yay -S chatgpt.sh
Install the termux-api
package.
For recording audio (Whisper, option -w
), we defaults to termux-microphone-record
and for playing audio (TTS, option -z
), optionally install play-audio
.
To set the clipboard, it is required termux-clipboard-set
from the termux-api
package.
Markdown in TTS input may stutter the model audio generation a little.
If python
modules markdown
and bs4
are available, TTS input will
be converted to plain text. As fallback, pandoc
is used if present.
Under Termux, make sure to have your system updated and installed with
python
, rust
, and rustc-dev
packages for building tiktoken
.
pkg update
pkg upgrade
pkg install python rust rustc-dev
pip install tiktoken
-
Implement nice features from OpenAI API version 1.
-
Provide the closest API defaults.
-
Let the user customise defaults (as homework).
-
OpenAI API version 1 is the focus of the present project implementation. Not all features of the API will be covered.
-
See BUGS AND LIMITS section in the man page.
-
Bash shell truncates input on
\000
(null). -
Bash "read command" may not correctly display input buffers larger than the TTY screen size during editing. However, input buffers remain unaffected. Use the text editor interface for big prompt editing.
-
Garbage in, garbage out. An idiot savant.
Please leave bug reports at the GitHub issues page.
Read the online man page here.
Alternatively, a help page snippet can be printed with chatgpt.sh -h
.
Many Thanks to all that contributed to this project.
The following projects are worth remarking. They were studied during development of this script and used as referencial code sources.
- TheR1D's shell_gpt
- xenodium's chatgpt-shell
- llm-workflow-engine
- 0xacx's chatGPT-shell-cli
- mudler's LocalAI
- Ollama
- Google Gemini
- f's awesome-chatgpt-prompts
- PlexPt's awesome-chatgpt-prompts-zh
Everyone is welcome to submit issues, PRs, and new ideas!
https://gitlab.com/fenixdragao/shellchatgpt
Mirror
https://github.com/mountaineerbr/shellChatGPT