Architecture

Four layers on top of a shared infrastructure layer:

• Public API: maki/__init__.py lazy-loads all exports on first access.
• LLM backends: MakiLLama, MakiOpenAI, MakiAnthropic, MakiOpenRouter, MakiLlamaCpp and HFBackend all implement the abstract LLMBackend contract.
• Agent system: Agent composes PluginHandler and ReasoningEngine mixins; AgentManager orchestrates agents via WorkflowTask and WorkflowState.
• Distributed layer: AgentServer (FastAPI) exposes agents over HTTP; AgentProxy is the remote-agent client with circuit breaking; DistributedAgentManager mixes local and remote agents.
• Infrastructure: Connector (hardened HTTP: URL validation, connect-time IP pinning), shared data classes, typed exceptions, runtime config and logging.

Plugins sit alongside the agent layer. They are loaded on demand and invoked when the model emits a TOOL: directive or through native tool-calling APIs.

Public API

Top-level imports from maki:

• MakiLLama, MakiOpenAI, MakiAnthropic, MakiOpenRouter, MakiLlamaCpp, HFBackend
• LLMBackend, BackendType
• Agent, AgentManager
• GenerationConfig, LLMResponse, Message, ToolCall
• ConversationMemory, RateLimiter
• Connector, Utils
• config

Desktop app

The repository includes a PySide6/QML desktop shell. Install maki-framework[gui] and run maki-gui.

Contributing

Bug fixes, documentation, new plugins and suggestions are welcome. Open an issue or pull request on GitHub.

pip install -e ".[dev]"
pytest          # tests marked @pytest.mark.network are excluded by default

Maki is part of Bowl of Data, an open-source AI research community. Licensed under MIT.