Streaming output · Tool calling · MCP integration · Event-driven — Build agents that think and act
AmritaCore is a lightweight Agent runtime built on AmritaSense.
AmritaCore = AmritaSense (workflow engine + events + streaming) + Agent layer (strategy, sessions, tools, MCP, adapters)AmritaSense provides the execution substrate — a native-instruction workflow VM, a bidirectional SuspendObjectStream, and a matcher-based event system. AmritaCore builds the agent layer on top: conversation objects, tool calling, MCP clients, model adapters, and a built-in step-driven ReAct strategy.
Learning path: The documentation follows the natural journey of building agents — run it first, extend it, tune it, then understand the internals. AmritaSense-specific topics are only recapped inline where needed and linked to sense.amritabot.com — they are not duplicated here.
| Stage | Section | What you will get |
|---|---|---|
| ① Run it | Getting Started | Environment, minimal example, first agent |
| ② Use it | Tutorials | Tools, streaming, hooks, memory — step by step |
| ③ Understand it | Concepts | How ChatObject, strategies, events and data fit together |
| ④ Extend it | Extensions & Integration | Adapters, custom tools, MCP, custom tokenizers |
| ⑤ Tune it | Agent Engineering | Prompt engineering, Jinja2 templates, troubleshooting |
| ⑥ Go deeper | Advanced | The workflow engine, suspend/resume, step-loop internals |
| ⑦ Design philosophy | Introduction + Appendix | Why AmritaCore is designed the way it is |
Shortcuts: Prefer reference over prose? Jump straight to the API Reference, the Built-in Capabilities, or the Security Mechanisms.