Maki

Maki is a Python framework for multi-agent LLM applications that run wherever your models do: on your own hardware (through Ollama or llama.cpp), on a hosted API (OpenAI, Anthropic, OpenRouter), or a mix of both, with the same agent code.

It is for developers who want to build and run tool-using agents on local models first, keep the option of moving any task to a hosted model by swapping one object, and have guardrails switched on by default while those agents touch files, the web, or a broker account.

Documentation: https://bowlofdata.github.io/maki/docs/

Quickstart

Install Ollama, pull a small model, then install Maki:

ollama pull llama3.2
pip install maki-framework

from maki import MakiLLama
from maki.agents import Agent

agent = Agent(
    name="Reviewer",
    maki_instance=MakiLLama(model="llama3.2"),
    role="code reviewer",
    instructions="Focus on bugs, regressions, and missing validation.",
)

print(agent.execute_task("Review this design: a plugin system with file access."))

The package is maki-framework; the import is maki. To run the same agent on a hosted model, swap the backend object:

from maki import MakiAnthropic   # pip install "maki-framework[anthropic]"

agent = Agent(name="Reviewer", maki_instance=MakiAnthropic(model="claude-sonnet-4-5"), role="code reviewer")

Why Maki

• Local and hosted models are equals. Ollama, llama.cpp, OpenAI, Anthropic and OpenRouter sit behind the same LLMBackend contract, and an in-process HuggingFace backend is included.
• Guardrails on by default. Content-derived URLs are checked against private and reserved address ranges at connect time. Plugins are fail-closed, and destructive methods (file writes, FTP transfers, trades) stay disabled until you pass Agent(allow_dangerous_tools=True).
• Small core. Three dependencies (requests, httpx, python-dotenv); everything else is an opt-in extra.
• Agents as services. maki serve exposes any agent over HTTP; AgentProxy consumes it remotely.
• Batteries included. 16 built-in plugins and a workflow engine with dependency resolution, retries, parallel execution, checkpoint/resume and maker/checker verification.

Maki is young (0.x) and deliberately small. If you need a large catalog of third-party integrations or hosted tracing, a larger agent framework will serve you better.

Documentation

Contributing

Bug fixes, documentation, new plugins and suggestions are welcome. Open an issue or pull request on GitHub. Maki is part of Bowl of Data, an open-source AI research community.

License

MIT