Asq is a command-line wrapper around a OpenAI Chat Completions-compatible API of your choice, that keeps things simple and helps your LLM get to the point quickly.
Selling points:
- Opinionated - Asq instructs LLMs to be brief, keeping signal-to-noise high
- Low complexity - hackable by design, so you can extend and customise it for your purposes
- Not slop-coded - largely built by hand = code is more understandable than the average hands-off vibe-coded AI project
- Self-understanding - Asq shows its own source code to LLMs, so they can understand and help you hack on the codebase easily. You don't even need to know Python - just ask Asq to tell you what to change.
Asq is currently preset to provide knowledge related to DevOps, Python and AWS, but it can be very easily customised with minor changes to anything you like - it can just as easily become your personal Ruby on Windows expert, or a JavaScript on ChromeOS expert.
Asq is currently stateless (i. e. won't remember anything between conversations). Session history recording will be added at some point.
Before using Asq, you must configure the API settings. You can also add some information about your personal environment, so Asq can tailor its advice to your tech stack.
By default, Asq will send model requests to Chutes.AI.
If you wish to override the base URL with OpenRouter, llama.cpp (llama-server), a litellm proxy, or any other Chat Completions-compatible HTTP/HTTPS endpoint, you can do it by setting an environment variable named OPENAI_API_KEY to the API base URL supplied by your API provider.
A base URL will usually look something like this: http://localhost:8080/v1 (llama-server from llama.cpp) or https://api.example-provider.org/v1 (hosted API provider)
Note the /v1 at the end - if you are seeing 404/403 errors, this is probably because you got the URL wrong and are including too much or too little of the URL.
You can set up your Chat Completions API key either by setting an environment variable named OPENAI_API_KEY, or by saving it in a file called openai_key.txt in the current working directory. Environment variable takes precedence.
To make sure that Asq can get to the point and provide details that are relevant to your OS, device type (laptop/desktop) and so on, you must provide a description like this in the third person.
Create a file called user_description.txt by copying the provided user_description.txt.template:
cp user_description.txt.template user_description.txt
And fill in any details about yourself that you'd like Asq to know. I personally chose to include some details about my OS, architecture, number of displays and other details that might help with troubleshooting your setup using Asq's help.
Now that Asq is configured, you can run it from the command line using ./asq or docker run -it asq (depending on how you installed it).
This will launch Asq's interface, through which you can interact with the high-powered/turbo model of your choice. Exit with a Ctrl+C, a Ctrl+D, or by typing /quit
By default, typing some text and pressing Enter will submit the question immediately. If you want to send a multi-line message - for example, the contents of a config file, a piece of source code, or a paragraph of text - you must first enter the multi-line mode.
To do that, begin your first line with a /// (or just type /// on its own and press Enter). You can now type/paste as much text as you need. To finish typing in multi-line mode and submit your message, end your last line with /// (or just type /// on its own and press Enter).
The whole block of text will then be submitted as a single message.
Asq supports several /-commands to make the CLI experience more tolerable:
-
/power: Switch between a high-powered and a turbo model. Asq starts in Turbo mode by default, for a couple of reasons:- Turbo models are usually cheaper and a lot faster, while being mostly as good
- High-powered models (in reasoning mode) will take a while to deliberate, which can break flow
-
/: BUGGY (this is a known issue and will be fixed at some point). Removes last message from the message log. Afterwards, press Enter to regenerate Asq's response, or type additional user input if you want to add anything before resubmitting your question. -
/reset: Clears the entire message log. Do this frequently to save on API costs (this reduces token count), prevent crashing (no controlled "forgetting" will be supported before a full refactor is done), and to keep each discussion focused on one topic only -
/source: Show Asq what its own source code looks like (injected as a system message). This will allow it to self-criticise and to explain its operation. This is useful if you are planning to hack on Asq to extend it for your own purposes. This option is off by default, as the full source code contains 3000+ tokens, and including it can be expensive in high-powered model mode. -
/quit: Exits the program
Make sure you configure Asq first using the instructions above - the Dockerfile build process will bake the config into the image. Follow the official documentation if Docker is not yet installed on your system.
Afterwards, build the Docker image:
docker build -t asq .- Chat history recording
- Refactor the codebase heavily, including splitting into files and moving functionality into separate libraries
- Work on a more ergonomic way to do rollbacks, including rolling back N messages and allowing user to edit what they already wrote.
- Less hacky multi-line support
- A simple web search + web scraping feature like that of Auto-GPT to allow Asq to help with technology released after the GPT cut-off date.
- More powerful message log editing. Ideally, all operations will be supported - insertion of user/assistant/system messages at an arbitrary point in the log, editing, deletion.
- More powerful history that works with the context length. This might include selecting a range of messages to be auto-summarised with GPT and replaced with the summary to keep the conversation length under the token limit.
- Translation of the user_description.txt from arbitrary format/writing style into a third-person style (using GPT) on first start-up
- Some sort of a configuration script to make the process of onboarding easier.
- Switchable personas (for different programming tasks, or even non-programming ones) and customisable inline text templates (langchain-style) to get Asq to produce the kind of solutions you want.