agentic.js

Give AI
a body.

Today's AI can only type. It can't see you, can't hear you, can't decide how to reach you, can't remember yesterday. Eight modules that change that.

The problem
AI is a brain in a jar
Every AI product ships the same interface: a text box. ChatGPT, Claude, Gemini — the most powerful technology in human history, and the only way to interact with it is typing.

Humans don't work like that. We see each other. We choose how to communicate — a tap on the shoulder, a whisper, a notification. We remember what happened. We act on intent, not instruction.

AI needs the same capabilities. Not as features bolted on, but as fundamental building blocks.
👁
It can't see youNo awareness of who's there, whether they're paying attention, or how they feel.
🎯
It can't choose how to reach youAlways the same channel. Can't decide between voice, notification, light, or silence.
🧠
It forgets everythingEvery conversation starts from zero. No continuity, no growth, no relationship.
The anatomy
Eight organs for a complete being
Each module is one capability. One file, zero dependencies, works in any browser or Node.js. Compose them into whatever shape of AI you need.
👁
sense
The eyes. Face detection, attention tracking, expression reading, head pose — via MediaPipe. AI sees who's there and whether they're engaged.
new AgenticSense(video)
🧠
core
The brain. LLM calls with tools, vision, streaming, and structured output. One function — ask() — replaces an entire framework.
ask(prompt, config)
⚡
act
The will. Multimodal input → structured action decision. Doesn't ask "what can I do" — asks "what should I do." Intent, not instruction.
act.decide(input)
🎨
render
The expression. Streaming markdown renderer — syntax highlighting, zero flicker, dark/light themes. AI's thoughts made visible in real-time.
create(el, opts)
🗣️
voice
The voice. TTS + STT in one library. OpenAI-compatible. Push-to-talk. AI speaks and listens.
voice.speak(text)
💭
memory
The memory. Conversation context + knowledge retrieval. Auto-trimming, persistence, fork. Conversations that build on each other.
createMemory(opts)
📦
store
The long-term memory. SQLite-first persistence — browser WASM + Node.js. Key-value simple, raw SQL when you need it. 8 backends, auto-fallback.
createStore(name)
🦀
claw
The body. Wires core + memory + tools into a living agent. Multi-session, heartbeat, lifecycle. The runtime that holds it all together.
createClaw(opts)
Architecture
Sense → Think → Act
A complete perception-cognition-action loop. Like a nervous system.
📝
Input
text / image / audio
→
👁
sense
perceive
→
⚡
core
think
→
🎯
act
decide
→
✨
render
express
🎙
voice
🧠
memory
💾
store
🦀
claw
Use one organ or all eight. They compose by convention, not coupling.
Playground
Talk to it
Core + memory + render running together in your browser. API key stays local.
Send a message to start.
Memory auto-trims. Markdown streams in real-time.
Memory: 0 tokens · 0 turns Ready
Code
A living agent in 12 lines
Not a demo. A complete agent with memory, knowledge recall, and streaming.
createClaw — the whole body
import { createClaw } from 'agentic-claw' const agent = createClaw({ apiKey: 'sk-ant-...', knowledge: true, }) // Teach it something await agent.learn('docs', longDocument) // It remembers, it recalls, it reasons await agent.chat('What does the doc say?') await agent.chat('Tell me more') // it remembers
~30KB
Total size
0
Dependencies
8
Modules
0
Build steps
Philosophy
Why organs, not features

Tools extend input. Actions extend output.

Tools let AI pull information in — search, compute, query. Actions let AI push intent out — speak, notify, illuminate. Two symmetric halves. Most frameworks only have the first.

One file you can read

Every module is a single file. No hidden complexity. When something breaks, you read the source, find the line, fix it. That's the entire debugging story.

Frameworks shape thinking

Call it a "tool" and developers think "what can AI access?" Call it "act" and they think "what should AI do?" The abstraction changes the question, and the question changes the product.

AI is not a backend

It's a participant. It should perceive the room, remember the conversation, choose how to respond, and express itself. That's what these eight modules enable.