AgentMake AI: a software developement kit for developing agentic AI applications that support 14 AI backends and work with 7 agentic components, such as tools and agents. (Developer: Eliran Wong)
Supported backends: anthropic, azure, cohere, custom, deepseek, genai, github, googleai, groq, llamacpp, mistral, ollama, openai, vertexai, xai
9-min introduction 24-min introduction
This SDK incorporates the best aspects of our favorite projects, LetMeDoIt AI, Toolmate AI and TeamGen AI, to create a library aimed at further advancing the development of agentic AI applications.
Windows, macOS, Linux, ChromeOS, Android via Termux
anthropic - Anthropic API [docs]
azure - Azure OpenAI API [docs]
cohere - Cohere API [docs]
custom - any openai-compatible backends that support function calling
deepseek - DeepSeek API [docs]
genai - Vertex AI or Google AI [docs]
github - Github API [docs]
groq - Groq Cloud API [docs]
llamacpp - Llama.cpp Server [docs] - local setup required
mistral - Mistral API [docs]
ollama - Ollama [docs] - local setup required
openai - OpenAI API [docs]
For simplicity, agentmake uses ollama as the default backend, if parameter backend is not specified. Ollama models are automatically downloaded if they have not already been downloaded. Users can change the default backend by modifying environment variable DEFAULT_AI_BACKEND.
agentmake is designed to work with seven kinds of components for building agentic applications:
-
system- System messages are crucial for defining the roles of the AI agents and guiding how AI agents interact with users. Check out our examples.agentmakesupports the use offabricpatterns assystemcomponents for runningagentmakefunction or CLI options READ HERE. -
instruction- Predefined instructions that are added to users' prompts as prefixes, before they are passed to the AI models. Check out our examples.agentmakesupports the use offabricpatterns asinstructioncomponents for runningagentmakefunction or CLI options READ HERE. -
input_content_plugin- Input content plugins process or transform user inputs before they are passed to the AI models. Check out our examples. -
output_content_plugin- Output content plugins process or transform assistant responses after they are generated by AI models. Check out our examples. -
tool- Tools take simple structured actions in response to users' requests, with the use ofschemaandfunction calling. Check out our examples. -
agent- Agents are agentic applications automate multiple-step actions or decisions, to fulfill complicated requests. They can be executed on their own or integrated into an agentic workflow, supported byagentmake, to work collaboratively with other agents or components. Check out our examples. -
follow_up_prompt- Predefined prompts that are helpful for automating a series of follow-up responses after the first assistant response is generated. Check out our examples.
agentmake supports both built-in agentic components, created by our developers or contributors, and cutoms agentic components, created by users to meet their own needs.
Built-in agents components are placed into the following six folders inside the agentmake folders:
agents, instructions, plugins, prompts, systems, tools
To use the built-in components, you only need to specify the component filenames, without parent paths or file extensions, when you run the agentmake signature function or CLI options.
agentmake offers two options for users to use their custom components.
Option 1: Specify the full file path of inidividual components
Given the fact that each component can be organised as a single file, to use their own custom components, users only need to specify the file paths of the components they want to use, when they run the agentmake signature function or CLI options.
Option 2: Place custom components into agentmake user directory
The default agentmake user directory is ~/agentmake, i.e. a folder named agentmake, created under user's home directory. Uses may define their own path by modifying the environment variable AGENTMAKE_USER_DIR.
After creating a folder named agentmake under user directory, create six sub-folders in it, according to the following names and place your custom components in relevant folders, as we do with our built-in components.
If you organize the custom agentic components in this way, you only need to specify the component filenames, without parent paths or file extensions, when you run the agentmake signature function or CLI options.
In cases where a built-in tool and a custom tool have the same name, the custom tool takes priority over the built-in one. This allows for flexibility, enabling users to copy a built-in tool, modify its content, and retain the same name, thereby effectively overriding the built-in tool.
Basic:
pip install --upgrade agentmake
Basic installation supports all AI backends mentioned above, except for vertexai.
Extras:
We support Vertex AI via Google GenAI SDK. As this package supports most platforms, except for Android Termux, we separate this package google-genai as an extra. To support Vertex AI with agentmake, install with running:
pip install --upgrade "agentmake[genai]"
It is recommended not to install agentmake inside the directory ~/agentmake, as ~/agentmake is used by default for placing user custom content.
This SDK is designed to offer a single signature function agentmake for interacting with all AI backends, delivering a unified experience for generating AI responses. The main APIs are provided with the function agentmake located in this file.
Find documentation at https://github.com/eliranwong/agentmake/blob/main/docs/README.md
The following examples assumes Ollama is installed as the default backend.
To import:
from agentmake import agentmake
To run, e.g.:
agentmake("What is AI?")
To work with parameter tool, e.g.:
agentmake("What is AgentMake AI?", tool="search/google")
agentmake("How many 'r's are there in the word 'strawberry'?", tool="magic")
agentmake("What time is it right now?", tool="magic")
agentmake("Open github.com in a web browser.", tool="magic")
agentmake("Convert file 'music.wav' into mp3 format.", tool="magic")
agentmake("Send an email to Eliran Wong at eliran.wong@domain.com to express my gratitude for his work.", tool="email/gmail")
A cross-platform solution to work with a tool that is placed in a sub-folder, e.g.:
agentmake("Extract text from image file 'sample.png'.", tool=os.path.join("perplexica", "github"))
To work with parameters input_content_plugin and output_content_plugin, e.g.:
agentmake("what AI model best", input_content_plugin="improve_writing", output_content_plugin="translate_into_chinese", stream=True)
To work with plugin that is placed in a sub-folder, e.g.:
agentmake("你好吗？", output_content_plugin=os.path.join("chinese", "convert_simplified"))
To automate prompt engineering:
agentmake("what best LLM training method", system="auto", input_content_plugin="improve_prompt")
To work with parameter system, instruction, follow_up_prompt, e.g.:
agentmake("Is it better to drink wine in the morning, afternoon, or evening?", instruction="reflect", stream=True)
agentmake("Is it better to drink wine in the morning, afternoon, or evening?", instruction="think", follow_up_prompt=["review", "refine"], stream=True)
agentmake("Provide a detailed introduction to generative AI.", system=["create_agents", "assign_agents"], follow_up_prompt="Who is the best agent to contribute next?", stream=True, model="llama3.3:70b")
To work with parameter agent, e.g.:
agentmake("Write detailed comments about the works of William Shakespeare, focusing on his literary contributions, dramatic techniques, and the profound impact he has had on the world of literature and theatre.", agent="teamwork", stream=True, model="llama3.3:70b")
To specify an AI backend:
agentmake("What is Microsoft stock price today?", tool=os.path.join("search", "finance"), backend="azure")
To work collaboratively with different backends, e.g.
messages = agentmake("What is the most effective method for training AI models?", backend="openai")
messages = agentmake(messages, backend="googleai", follow_up_prompt="Can you give me some different options?")
messages = agentmake(messages, backend="xai", follow_up_prompt="What are the limitations or potential biases in this information?")
agentmake(messages, backend="mistral", follow_up_prompt="Please provide a summary of the discussion so far.")
As you may see, the agentmake function returns the messages list, which is passed to the next agentmake function in turns.
Therefore, it is very simple to create a chatbot application, you can do it as few as five lines or less, e.g.:
messages = [{"role": "system", "content": "You are an AI assistant."}]
user_input = "Hello!"
while user_input:
messages = agentmake(messages, follow_up_prompt=user_input, stream=True)
user_input = input("Enter your query:\n(enter a blank entry to exit)\n>>> ")
You may take a look at out our built-in components for more ideas:
Command CLI are designed for quick run of AI features.
Check for CLI options, run:
agentmark -h
Two shortcut commands:
ai == agentmake
aic == agentmake -c with chat features enabled
The available CLI options use the same parameter names as the agentmake function for AI backend configurations, to offer users a unified experience. Below are some CLI examples, that are equivalent to some of the examples mentioned above:
ai What is AI?
ai What is AgentMake AI --tool search/google
ai Convert file music.wav into mp3 format. --tool task
ai Send an email to Eliran Wong at eliran.wong@domain.com to express my gratitude for his work --tool email/gmail
ai Extract text from image file sample.png. --tool=ocr/openai
ai What is Microsoft stock price today? -t search/finance -b azure
ai what AI model best --input_content_plugin improve_writing --output_content_plugin translate_into_chinese
ai what best LLM training method --system auto --input_content_plugin improve_prompt
ai 你好吗？ --output_content_plugin=chinese/convert_simplified
ai Is it better to drink wine in the morning, afternoon, or evening? --instruction think --follow_up_prompt review --follow_up_prompt refine
ai Write detailed comments about the works of William Shakespeare, focusing on his literary contributions, dramatic techniques, and the profound impact he has had on the world of literature and theatre --agent teamwork --model "llama3.3:70b"
CLI options are handy for testing, e.g. simply use a newly developed tool file with -t option and run:
ai What is AgentMake AI? -t ~/my_folder/perplexica.py
For quick start, run:
agentmake -ec
For more options:
To use ollama as the default backend, you need to download and install Ollama. To use backends other than Ollama, you need to use your own API keys. There are a few options you may configure the AI backends to work with agentmake.
Specify AI backend configurations as parameters when you run the agentmake signature function agentmake.
Setting configurations via option 1 overrides the default configurations set by option 2 and option 3, but the overriding is effective only when you run the function, with the specified configurations. Default configurations described below in option 2 and 3 still apply next time when you run the agentmake function, without specifying the AI backend parameters. This gives you flexibility to specify different settings in addition to the default ones.
You may manually export individual environment variables listed in https://github.com/eliranwong/agentmake/blob/main/agentmake.env
- Make a copy of the file
agentmake.env, located in the package directory, as:
either ~/agentmake/agentmake.env
cd agentmake # where you installed agentmake
cp agentmake.env ~/agentmake/agentmake.env
or <package_directory>/.env:
cd agentmake # where you installed agentmake
cp agentmake.env .env
- Edit the file manually with a text editor, e.g.
etextedit ~/agentmake/agentmake.env
- Save the changes
The changes apply next time when you run agentmake function or cli.
Use built-in agentmake cli option to edit the variables:
agentmake -ec
What does this command do?
- It automatically makes a copy of file
agentmake.envand save it as<package_directory>/.envif both<package_directory>/.envand~/agentmake/agentmake.envdo not exist. - It uses the text editor, specified in
DEFAULT_TEXT_EDITOR, to open the configuration file~/agentmake/agentmake.envif it exists or<package_directory>/.envif~/agentmake/agentmake.envdoes not exist.
Remember to save your changes to make them effective.
Make sure the extra package genai is installed with the command mentioned above:
pip install --upgrade "agentmake[genai]"
To configure, run:
ai -ec
Enter the path of your Google application credentials JSON file as the value of VERTEXAI_API_KEY. You need to specify your project ID and service location, in the configurations, as well. e.g.:
VERTEXAI_API_KEY=~/agentmake/google_application_credentials.json
VERTEXAI_API_PROJECT_ID=my_project_id
VERTEXAI_API_SERVICE_LOCATION=us-central1
To test Gemini 2.0 with Vertex AI, e.g.:
ai -b vertexai -m gemini-2.0-flash Hi!
- Please do not edit the file
agentmake.env, that is located in the package directory, directly, as it is restored to its default values upon each upgrade. It is recommended to make a copy of it and edit the copied file. - Multiple API keys are supported for running backends
cohere,github,groqandmistral. You may configure API keys for these backend in the.envfile by using commas,as separators, e.g.COHERE_API_KEY=cohere_api_key_1,cohere_api_key_2,cohere_api_key_3
fabric is a fantastic third-party project that offers a great collection of patterns.
agentmake supports the use of fabric patterns as entries for the system or instruction parameters when running the agentmake signature function or CLI options.
To use a fabric pattern in agentmake:
- Install fabric
- Specify a fabric pattern in
agentmakeparametersystemorinstruction, by prefixing the selected pattern withfabric.
agentmake("The United Kingdom is a Christian country.", tool="search/searxng", system="fabric.analyze_claims")
Both local backends ollama and llamacpp support GPU accelerations.
- add more examples
- convert availble ToolMate AI tools into tools that runable with this SDK (... in progess ...)
- add documentation about tool creation