search

smol-ai / developer

the first library to let you embed a developer agent in your own app!
https://twitter.com/SmolModels
MIT License
11.76k stars 1.03k forks source link

readme

🐣 smol developer

Deploy agent on e2b button

Morph Morph

Human-centric & Coherent Whole Program Synthesis aka your own personal junior developer

Build the thing that builds the thing! a smol dev for every dev in every situation

This is a "junior developer" agent (aka smol dev) that either:

  1. scaffolds an entire codebase out for you once you give it a product spec
  2. gives you basic building blocks to have a smol developer inside of your own app.

Instead of making and maintaining specific, rigid, one-shot starters, like create-react-app, or create-nextjs-app, this is basically is or helps you make create-anything-app where you develop your scaffolding prompt in a tight loop with your smol dev.

After the successful initial v0 launch, smol developer was rewritten to be even smol-ler, and importable from a library!

Basic Usage

In Git Repo mode

# install
git clone https://github.com/smol-ai/developer.git
cd developer
poetry install # install dependencies. pip install poetry if you need

# run
python main.py "a HTML/JS/CSS Tic Tac Toe Game" # defaults to gpt-4-0613
# python main.py "a HTML/JS/CSS Tic Tac Toe Game" --model=gpt-3.5-turbo-0613

# other cli flags
python main.py --prompt prompt.md # for longer prompts, move them into a markdown file
python main.py --prompt prompt.md --debug True # for debugging
This lets you develop apps as a human in the loop, as per the original version of smol developer.

*engineering with prompts, rather than prompt engineering* The demo example in `prompt.md` shows the potential of AI-enabled, but still firmly human developer centric, workflow: - Human writes a basic prompt for the app they want to build - `main.py` generates code - Human runs/reads the code - Human can: - simply add to the prompt as they discover underspecified parts of the prompt - manually runs the code and identifies errors - *paste the error into the prompt* just like they would file a GitHub issue - for extra help, they can use `debugger.py` which reads the whole codebase to make specific code change suggestions Loop until happiness is attained. Notice that AI is only used as long as it is adding value - once it gets in your way, just take over the codebase from your smol junior developer with no fuss and no hurt feelings. (*we could also have smol-dev take over an existing codebase and bootstrap its own prompt... but that's a Future Direction*)

In this way you can use your clone of this repo itself to prototype/develop your app.

In Library mode

This is the new thing in smol developer v1! Add smol developer to your own projects!

pip install smol_dev

Here you can basically look at the contents of main.py as our "documentation" of how you can use these functions and prompts in your own app:

from smol_dev.prompts import plan, specify_file_paths, generate_code_sync

prompt = "a HTML/JS/CSS Tic Tac Toe Game"

shared_deps = plan(prompt) # returns a long string representing the coding plan

# do something with the shared_deps plan if you wish, for example ask for user confirmation/edits and iterate in a loop

file_paths = specify_file_paths(prompt, shared_deps) # returns an array of strings representing the filenames it needs to write based on your prompt and shared_deps. Relies on OpenAI's new Function Calling API to guarantee JSON.

# do something with the filepaths if you wish, for example display a plan

# loop through file_paths array and generate code for each file
for file_path in file_paths:
    code = generate_code_sync(prompt, shared_deps, file_path) # generates the source code of each file

    # do something with the source code of the file, eg. write to disk or display in UI
    # there is also an async `generate_code()` version of this

In API mode (via Agent Protocol)

To start the server run:

poetry run api

or

python smol_dev/api.py

and then you can call the API using either the following commands:

To create a task run:

curl --request POST \
  --url http://localhost:8000/agent/tasks \
  --header 'Content-Type: application/json' \
  --data '{
    "input": "Write simple script in Python. It should write '\''Hello world!'\'' to hi.txt"
}'

You will get a response like this:

{"input":"Write simple script in Python. It should write 'Hello world!' to hi.txt","task_id":"d2c4e543-ae08-4a97-9ac5-5f9a4459cb19","artifacts":[]}

Then to execute one step of the task copy the task_id you got from the previous request and run:

curl --request POST \
  --url http://localhost:8000/agent/tasks/<task-id>/steps

or you can use Python client library:

from agent_protocol_client import AgentApi, ApiClient, TaskRequestBody

...

prompt = "Write simple script in Python. It should write 'Hello world!' to hi.txt"

async with ApiClient() as api_client:
    # Create an instance of the API class
    api_instance = AgentApi(api_client)
    task_request_body = TaskRequestBody(input=prompt)

    task = await api_instance.create_agent_task(
        task_request_body=task_request_body
    )
    task_id = task.task_id
    response = await api_instance.execute_agent_task_step(task_id=task_id)

...

examples/prompt gallery

I'm actively seeking more examples, please PR yours!

sorry for the lack of examples, I know that is frustrating but I wasnt ready for so many of you lol

major forks/alternatives

please send in alternative implementations, and deploy strategies on alternative stacks!

innovations and insights

Please subscribe to https://latent.space/ for a fuller writeup and insights and reflections

Please subscribe to https://latent.space/ for a fuller writeup and insights and reflections

caveats

We were working on a Chrome Extension, which requires images to be generated, so we added some usecase specific code in there to skip destroying/regenerating them, that we haven't decided how to generalize.

We dont have access to GPT4-32k, but if we did, we'd explore dumping entire API/SDK documentation into context.

The feedback loop is very slow right now (time says about 2-4 mins to generate a program with GPT4, even with parallelization due to Modal (occasionally spiking higher)), but it's a safe bet that it will go down over time (see also "future directions" below).

future directions

things to try/would accept open issue discussions and PRs: