How josh-xt/agixt works
Project type and purpose
• AGiXT is a modular “AI Automation Platform” implemented in Python. It provides a FastAPI‐based RESTful API for orchestrating AI instruction management, task execution, conversation chains, memory handling, and a plugin/extension system.
• It integrates multiple AI back ends (OpenAI, Hugging Face, local ONNX models, etc.), manages agents, prompts, chains, and supports third-party authentication via SSO.
File structure
• Top level
• start.py, Dockerfile, docker-compose.yml: initialization and deployment scripts
• pyproject.toml, requirements.txt, setup.py: packaging and dependencies
• agixt/ (core package)
• app.py: FastAPI app entrypoint
• endpoints/: HTTP route handlers (Agent, Chain, Completions, Conversation, Memory, Prompt, Provider, Tasks, Health, GQL, etc.)
• Agent.py, Chain.py, Task.py, TaskMonitor.py, Conversations.py, Models.py, Prompts.py, Providers.py, Memories.py, DB.py, Interactions.py, MagicalAuth.py, SessionTracker.py, etc.: core orchestration modules
• extensions/: plugin modules (web browsing, cameras, databases, file system, scheduled tasks, etc.)
• providers/: adapters to various AI providers (OpenAI, Azure, HuggingFace, GPT4Free, local)
• sso/: OAuth2 handlers for GitHub, Google, Microsoft, Alexa, Tesla, etc.
• onnx/: local model artifacts and tokenizer configs
• prompts/, chains/, conversations/: JSON/YAML definitions of workflows
• Configuration & supporting directories
• docker-compose*.yml, .github/workflows, docs/, examples/, tests/
README insights
a. Main components
1. Client layer
– SDKs (Python, TypeScript, Dart, Rust, Zig)
– UIs (AGiXT Management UI, AGiXT Interactive UI)
2. API Gateway & Request ingress
– HTTP and WebSocket endpoints on Uvicorn/FastAPI (start.py → app.py)
3. Core backend modules
– Request router (endpoints/)
– Agent Manager (Agent.py, Workspaces.py, TaskMonitor.py)
– Chain Orchestrator (Chain.py, Chains definitions)
– Task Executor (Task.py)
– Memory Manager (Memories.py, vector/SQ store via DB.py)
– Prompt Manager (Prompts.py, prompt templates)
– Extension Executor (extensions/)
– Provider Adapter Layer (providers/) + ONNX runtime
– Authentication & SSO (MagicalAuth.py, sso/)
– Session Tracker & Logging (session_tracker.py, DB.py)
4. Data stores
– Relational DB (SQLite/Postgres) for metadata, logs
– Vector DB or embedding store (in-process or external)
5. External systems
– AI providers (OpenAI, Azure, Hugging Face, GPT4Free, local ONNX)
– Third-party services via extensions (web browsing, databases, cameras, scheduled tasks, etc.)
b. Relationships & data flows
• Clients → API Gateway → FastAPI endpoints
• Endpoints invoke core modules (Agent Manager → Chain Orchestrator or Prompt Inference or Task Executor)
• Core orchestrators read/write metadata to Relational DB, query/update vectors in Vector DB
• For inference: Provider Adapter → call external AI API or ONNX runtime
• Extension Executor → external services via plugin interfaces
• Results → format & return to client, with logging/tracing
• SSO flows: Client ←→ MagicalAuth / sso handlers
c. Architectural patterns & principles
• Modular monolith with plugin/extension architecture
• Adapter pattern for AI providers
• Workflow/chain orchestration pattern
• Clear separation of concerns: routing, orchestration, persistence, external integrations
• Dependency injection via FastAPI
d. Key technologies & frameworks
• Python 3.10+, FastAPI, Uvicorn
• SQLAlchemy or similar via DB.py for relational DB
• ONNX Runtime for local inference
• Docker & Docker Compose for deployment
• External AI APIs (OpenAI, Azure, Hugging Face, etc.)
• OAuth2 for SSO, REST for provider communication
Essential details
• Don’t over-modularize; group logically related modules under higher-level components
• SSO/authentication flow
• If using Docker Compose, show each service container (API service, ONNX service if separate, DB, vector store) and network links