Introduction
Veska is a general-purpose Python framework for building AI agent systems from scratch. It supports multi-model providers, streaming, structured output, delegation, memory, and more, without depending on any external AI framework.
Why Veska?
Most AI frameworks are black boxes, layers of abstraction that hide what's happening. Veska gives you full control. Every component is transparent, modular, and built from first principles. You own every line of code.
Key features
- Multi-model — Use Claude, OpenAI, or both, per-agent configurable
- Streaming — Real-time token streaming via async generators
- Structured Output — Pydantic validation with auto-retry
- Delegation — Orchestrator plans and delegates to specialized agents
- Memory — Per-agent and shared memory with pluggable backends
- Tools — Pre-built file, code, and project tools + custom tools
- Security — 3-zone sandbox, command guard, code scanner
- Error Recovery — 3-level system: retry, agent fix, discussion room
- Extended Thinking — Budget-controlled chain-of-thought reasoning
- Human-in-the-Loop — Agents can ask users for input
- Events — Subscribe to framework events for UI updates
- MCP — Connect to external MCP servers for additional tools