AmritaConstant is an open‑source collective building next‑gen AI Agent infrastructure.
We believe frameworks should be sharp tools, not bloated fortresses. We tackle the pain points of heavyweight architectures, concurrency bottlenecks, and costly customization in mainstream Agent ecosystems — delivering a lightweight, modular, production‑ready toolkit.
Our DNA: break the mold, experiment in public. We intentionally choose unconventional paths, prioritise real‑world usefulness over polished demos, and embrace the messiness of genuine exploration.
We don’t aim to “support everything” — we aim to define new ways to build.
| Project | What it does | Status |
|---|---|---|
| AmritaCore | High‑performance Agent runtime core — pluggable, event‑driven, lightweight, and concurrent‑first | ✅ Production‑ready |
| AmritaSense | General‑purpose workflow orchestration engine — the support library and runtime for Core. ISA‑based, with native GOTO/CALL, PointerVector addressing, and async interrupt |
✅ Available |
| AmritaBot | Full‑stack LLM Agent chatbot (built on NoneBot2) — WebUI, multi‑adapter, permission system | 🚧 Usable, but no i18n (by design: it’s our tech experimentation platform, keeping the core lean and fast‑iterating) |
| AtomInline | Universal Agent CLI built on AmritaCore — lightweight, scriptable, and designed for seamless integration into existing dev toolchains | 🚧 Under development |
| AmritaBase | Lightweight LLM infrastructure — a simpler alternative to LangChain (vector, memory, tool abstractions) | 🚧 Under development |
Why no i18n for AmritaBot? It’s intentional. We avoid callback hell and multi‑scenario adaptation overhead, focusing instead on rapid Core iteration and real‑time error logging. It’s a working lab, not a polished product.
We’re not building “yet another framework.” We’re exploring what Agent infrastructure can be when you:
- Reject over‑abstraction — code should be readable, not a black box of layers
- Accept the cost of experimentation — some paths only reveal themselves when you walk them
- Anchor responsibility in humans — AI assists, but final decisions stay with people (see AACLP)
If you’re tired of “enterprise‑grade” bloat and curious about unproven but promising edges — you’ll feel at home here.
- Telegram (Chinese):
- QQ Group:
(ID: 1006893368)
- Discord:
Whether you’re building with Amrita, contributing code, or just curious — if you have your own take on making Agents lighter, faster, and more controllable, we’d love to have you.
Let’s build something different. Together.
