Features
Agentic is built around a set of composable features. Each feature is independent and can be used on its own or combined with others.
| Feature | Description |
|---|---|
ILLMBackend | Unified abstraction over any inference source; swap backends without touching agent code |
OpenAIBackend | OpenAI-compatible REST client with streaming, embeddings, vision and health-check |
NativeBackend | Local llama.cpp inference with auto-install from GitHub releases |
BackendRouter | Compose multiple backends; route chat by model name, embeddings to a dedicated model |
LlamaRuntimeInstaller | On-demand runtime installer for CPU, CUDA and Vulkan on Windows and Linux |
| Agent | Multi-turn streaming agent with automatic MCP tool orchestration |
| Image Input | Send images alongside text as URL, local file, or base64 |
| Workflows | Ordered multi-step execution with per-step async guardrails and retry |
| Tool System | Define tools with [Tool] / [ToolParam] attributes; zero boilerplate |
| Tool Context | HTTP headers forwarded to tool methods via ToolContext |
| MCP Server | Expose any IAgentToolSet over HTTP as a Model Context Protocol server |
| Context Compaction | Auto-summarise older history into a structured checkpoint |
| Vector Storage | IStore / ICollection<T> with SQLite or PostgreSQL + pgvector |
| Reasoning Control | Control chain-of-thought effort at global, agent, or request level |
| Inference Config | Sampling and penalty parameters with three-level override |