AIRIClaw

A local-first Agent + digital-human assistant. AIRIClaw fuses the agent / skill / gateway runtime of OpenClaw (2026.6.2) with the character and event protocol of Project AIRI (v0.10.2), powered by DeepSeek V4 Pro.

It is a single, self-contained TypeScript monolith — not a bridge between two apps. The pieces that matter are re-implemented from both projects' latest source and wired into one runtime.

input:text ─▶ recall memory ─▶ build system prompt ─▶ DeepSeek agent loop ─▶ output:gen-ai:chat:*
                (AIRI memory)     (AIRI persona +          (OpenClaw           (AIRI protocol →
                                   OpenClaw skills)          agent-core)          avatar / voice)

📖 完整使用文档见 docs/USAGE.md — 安装、配置、CLI 命令、技能编写、角色卡、记忆、Library API、Gateway 接入、常见问题。

Features

• DeepSeek V4 Pro as the brain — OpenAI-compatible client with non_think / think_high / think_max thinking modes, 1M-token context, streaming with reasoning deltas.
• Agent loop (OpenClaw agent-core, 2026.6.2 event names): streaming, tool-calling, iterates until the model stops requesting tools; sequential or parallel tool execution.
• Skills in OpenClaw's SKILL.md format — YAML frontmatter with metadata.openclaw (emoji / always / requires / install / skillKey), and upstream-faithful invocation policy (user-invocable defaults to true, disable-model-invocation hides a skill from the model). Skills become callable tools; invoking one feeds its SKILL.md body back to the model (progressive disclosure). always skills are inlined into the system prompt.
• Character / persona from an AIRI-style character card (personality, scenario, greetings, system prompt, avatar & speech modules).
• Memory: MEMORY.md, DREAMS.md, and dated daily notes, with keyword recall injected into the system prompt every turn.
• AIRI protocol events (input:text, output:gen-ai:chat:*, output:emotion, output:lipsync, output:speech, context:update, spark:*) over a typed event bus — the seam where a digital-human frontend plugs in.
• Gateway server speaking OpenClaw's frame protocol v4 (req/res/event) on port 18789, with every protocol event fanned out to connected clients.
• Voice / lip-sync interfaces (TTSProvider, STTProvider, LipSyncDriver) with safe headless defaults, so an avatar layer has a contract to render against.

Quick start

git clone https://github.com/JacksonTai2007/AIRIClaw.git
cd AIRIClaw
pnpm install
pnpm build      # tsup → dist/
pnpm test       # vitest

export DEEPSEEK_API_KEY=sk-...
node dist/cli.js chat "你好"

Requires Node ≥ 22.

CLI

airiclaw chat [message]       # one-shot, or interactive REPL with no args
airiclaw skills list [-i]     # list skills (-i: user-invocable only)
airiclaw skills search <q>    # search model-invocable skills
airiclaw serve [--port N]     # gateway server (protocol v4, default :18789)
airiclaw config init          # write .airiclaw/config.json
airiclaw status               # resolved config + provider info

Library usage

import { Assistant, loadConfig } from 'airiclaw'

const assistant = new Assistant({ config: await loadConfig() })
await assistant.loadSkills()

assistant.events.on('output:gen-ai:chat:delta', ({ delta }) => process.stdout.write(delta))
assistant.events.on('output:lipsync', ({ mouthOpen }) => avatar.setMouth(mouthOpen))

const { text } = await assistant.chat('hello!')

Architecture

One package under src/, organized by concern:

Module	Origin	Responsibility
src/llm	unified	Provider contract + DeepSeekProvider (V4 Pro)
src/agent	OpenClaw agent-core	Streaming agent loop + tool dispatch
src/skills	OpenClaw skills	SKILL.md parsing, registry, prompt, tool adapter
src/character	AIRI character card	Persona + system-prompt construction
src/memory	OpenClaw/AIRI memory	MEMORY.md / DREAMS.md / daily notes + recall
src/events	AIRI plugin-protocol	Typed protocol event bus
src/gateway	OpenClaw gateway-protocol	Frame protocol v4 + WebSocket server
src/voice	AIRI audio pipeline	TTS / STT / lip-sync interfaces
src/runtime	the fusion	Assistant — one conversational turn, end to end
src/config	—	Config schema + loader
src/cli.ts	—	chat / skills / serve / config / status

Examples live under examples/ (a sample skill and character card).

Status

Headless core: complete, builds, fully tested. The digital-human frontend (Live2D/VRM rendering, real TTS/STT) is intentionally an interface — connect any renderer to the output:* events.

License

MIT