BrainOS 发布多语言 AI 智能体运行时，Rust 核心配 Python 与 JS 绑定

English | 中文版本

BrainOS — Multi-language AI agent runtime. One framework — Rust core, Python & JS bindings.

You have an LLM. You want it to use tools, talk to other agents, remember conversations, and connect to MCP servers — in Python, JavaScript, or Rust.

BOS is the runtime that makes this work out of the box. One pip install nbos or npm install @open1s/jsbos gets you agents with tools, a pub/sub event bus for multi-agent coordination, MCP client for external tools, skill loading for domain-specific capabilities, and cross-session memory — all backed by a performant Rust core.

# 30-second win — copy, paste, run
pip install nbos && python -c "
from nbos import BrainOS
import asyncio
print(asyncio.run(BrainOS().agent('assistant').ask('say hi')))
"

Quick Start

Python (nbos)

from nbos import BrainOS, tool
from nbos.content import Content, ContentPart

@tool("Add two numbers")
def add(a: int, b: int) -> int:
    return a + b

async with BrainOS() as brain:
    agent = brain.agent("assistant").with_tools(add)
    result = await agent.ask("What is 2+2?")

    # Multimodal: text + image
    content = Content.parts([
        ContentPart.text("What is in this image?"),
        ContentPart.image("https://example.com/photo.jpg"),
    ])
    result = await agent.ask(content)

    # Multimodal: text + audio
    audio_content = Content.parts([
        ContentPart.text("Transcribe this audio"),
        ContentPart.audio("/path/to/audio.wav", "wav"),
    ])
    result = await agent.ask(audio_content)

JavaScript (@open1s/jsbos)

import { BrainOS, ToolDef, Content, ContentPart } from '@open1s/jsbos';

// Create tool using ToolDef
const addTool = new ToolDef(
  'add',
  'Add two numbers',
  (args) => (args.a || 0) + (args.b || 0),
  { type: 'object', properties: { result: { type: 'number' } }, required: ['result'] },
  { type: 'object', properties: { a: { type: 'number' }, b: { type: 'number' } }, required: ['a', 'b'] }
);

const brain = new BrainOS();
await brain.start();
const agent = await brain.agent('assistant')
  .register(addTool)
  .start();
const result = await agent.runSimple('What is 2+2?');

// Multimodal: text + image
const content = Content.parts([
    ContentPart.text('What is in this image?'),
    ContentPart.image('https://example.com/photo.jpg'),
]);
const result2 = await agent.ask(content);

// Multimodal: text + audio
const audioContent = Content.parts([
    ContentPart.text('Transcribe this audio'),
    ContentPart.audio('/path/to/audio.wav', 'wav'),
]);
const result3 = await agent.ask(audioContent);

Rust (agent crate)

use agent::{Agent, AgentConfig};

let config = AgentConfig::default()
    .name("assistant")
    .model("openai/gpt-4o")            // vendor inferred from the model prefix
    .api_key(std::env::var("OPENAI_API_KEY")?);
let agent = Agent::from_config(config);
let result = agent.run_simple("Hello").await?;

Why BOS?

BOS is not another LangChain wrapper or Python-only framework. It's a multi-language runtime built from the ground up for production AI agents.

If you want an agent that speaks more than one language, talks to other agents, and works in production — not just a notebook — BOS is the runtime.

Skills System

Agents can load capabilities from directory-based skill definitions:

# Create skills directory with skill folders
skills_dir = "/path/to/skills"
agent.register_skills_from_dir(skills_dir)

Skill Format

Each skill is a directory containing a SKILL.md file with YAML frontmatter:

skills/
├── python-coding/
│   └── SKILL.md
├── api-design/
│   └── SKILL.md
└── database-ops/
    └── SKILL.md

SKILL.md format:

---
name: python-coding
description: Python coding conventions for this project
category: coding
version: 1.0.0
---

# Python Coding Conventions

Your skill instructions here...

The agent's LLM receives available skills in the system prompt and can call load_skill to retrieve full instructions.

Project Structure

bos/
├── crates/
│   ├── agent/          # AI agent framework with LLM, skills, tools, MCP
│   ├── bus/            # Pub/sub, queryable, caller/callable
│   ├── config/         # TOML/YAML config loading
│   ├── logging/        # Tracing and instrumentation
│   ├── react/          # ReAct reasoning engine
│   ├── qserde/         # rkyv-backed wire format (+ qserde_derive)
│   ├── resource/       # Resource subsystem + `rex` CLI
│   ├── nbos/           # Python bindings (nbos package)
│   └── jsbos/          # Node.js bindings (@open1s/jsbos)
├── docs/               # User guides
│   ├── python-user-guide.md
│   ├── javascript-user-guide.md
│   └── rust-user-guide.md
└── Cargo.toml          # Workspace

Crates

Commands

# Build all
cargo build --all

# Test all
cargo test --all

# Lint
cargo clippy --all
cargo fmt --all

# Python bindings (nbos)
cd crates/nbos && maturin develop

# Node.js bindings (jsbos)
cd crates/jsbos && npm install && npm run build

User Guides

• Python: docs/python-user-guide.md
• JavaScript: docs/javascript-user-guide.md
• Rust: docs/rust-user-guide.md
• 中文: README-ZH.md

Unified API

The nbos package (Python) and @open1s/jsbos (JavaScript) provide consistent high-level APIs:

Low-level Bindings

For direct access to Rust bindings:

MCP Client

Connect to MCP servers via stdio or HTTP transport:

Python

from nbos import McpClient

# Process-based server
client = await McpClient.spawn("npx", ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"])
await client.initialize()

# HTTP server
client = McpClient.connect_http("http://127.0.0.1:8000/mcp")
await client.initialize()

# Use tools
tools = await client.list_tools()
result = await client.call_tool("echo", '{"text": "hello"}')

JavaScript

import { McpClient } from '@open1s/jsbos';

// Process-based server
const client = await McpClient.spawn("npx", ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]);
await client.initialize();

// HTTP server
const client = McpClient.connectHttp("http://127.0.0.1:8000/mcp");
await client.initialize();

// Use tools
const tools = await client.listTools();
const result = await client.callTool("echo", '{"text": "hello"}');

HTTP Server Example

# Start an MCP HTTP server
python3 crates/examples/mcp_http_server.py
# Server runs on http://127.0.0.1:8000/mcp

Configuration

Create ~/.bos/conf/config.toml:

# OpenAI-compatible endpoint
[global_model]
api_key = "your-api-key"
base_url = "https://api.openai.com/v1"
model = "gpt-4"

# NVIDIA NIM
[llm.nvidia]
api_key = "nv-..."
base_url = "https://integrate.api.nvidia.com/v1"
model = "nvidia/llama-3.1-nemotron-70b-instruct"

# Google AI (Gemini, etc.)
[llm.google]
api_key = "google-api-key"
base_url = "https://generativelanguage.googleapis.com/v1"
model = "gemini-pro"

# OpenRouter
[llm.openrouter]
api_key = "or-..."
base_url = "https://openrouter.ai/api/v1"
model = "anthropic/claude-3-haiku"

[bus]
mode = "peer"
listen = ["127.0.0.1:7890"]

Or use environment variables: OPENAI_API_KEY, LLM_BASE_URL, LLM_MODEL

Examples

See the examples directories:

• Python: crates/nbos/examples/
• JavaScript: crates/jsbos/examples/
• Rust: crates/examples/ (includes agent_skill_demo.rs)

Multimodal Demos

# Python multimodal (text, image, audio)
python crates/nbos/examples/demo_multimodal.py

# JavaScript multimodal (text, image, audio)
node crates/jsbos/examples/demo_multimodal.js

# JavaScript audio demo
node crates/jsbos/examples/demo_audio.js

MCP Demos

# JavaScript MCP HTTP demo
node crates/jsbos/examples/mcp_http_agent_demo.js

# Python MCP HTTP demo (run server first, then use)
python3 crates/examples/mcp_http_server.py

License

MIT OR Apache-2.0

Changelog

v3.0.2 (2026-10-07)

• Fixed: rate-limit/retry waits and per-run token, tool-call and tool-time metrics now recorded on the live LLM path
• Fixed: npm publish pipeline hardened — serialized publishes, dist-tag repair, stray-tag guard, idempotent re-runs
• Docs: READMEs refreshed for the 3.0.2 baseline release

v3.0.1 (2026-10-07)

• Added: Memory store (in-memory/file-backed) with metadata filters, {{memory}} prompt placeholder, and a Memory class in Python/JS
• Added: react::template prompt templates; AsyncFunctionTool::from_fn
• Added: Retry classification by error type; rate limiter honours retry_backoff/auto_wait
• Added: Per-call min/max LLM wall time in metrics, surfaced by both bindings

v2.3.6 (2026-07-01)

• Added: skill_dir to load_skill tool result — LLM receives the parent path of SKILL.md for accessing bundled resources

v2.3.5

• Added: Skills system: directory-based skill loading with SKILL.md files
• Added: load_skill tool — LLM can load skill instructions on demand
• Added: Skill caching with TTL in ReAct engine
• Fixed: Skill load performance — avoid redundant filesystem reads

v2.3.0 (2026-06-09)

• Added: Audio support — Content.audio(), ContentPart.audio() for multimodal LLM calls
• Added: LLM vendor configuration — [llm.google], [llm.nvidia], [llm.openrouter] in config
• Added: demo_audio.js and demo_multimodal.js examples
• Added: demo_multimodal.py Python example
• Improved: Content API — unified Content/ContentPart/Binary classes for text, images, audio
• Fixed: Content serialization for multimodal messages

v2.2.0 (2026-05-18)

• Added: agent.stop() — stop a running agent/react/stream via ESC key
• Added: agent.is_running() — check if agent is currently running
• Added: Stop flag and running state in NAPI Agent layer
• Added: StreamToken::Stopped variant for stream cancellation
• Added: Concurrent call prevention with error "Agent is already running"
• Added: Stream returns {status: "stopped"|"completed"} JSON string
• Breaking: stream() now returns String (JSON status object) instead of ()

v2.1.4 (2026-05-15)

• Fixed: getPerfMetrics() now correctly updates after stream() calls
• Added: Token usage tracking — totalInputTokens and totalOutputTokens are now recorded from LLM responses for react(), runSimple(), and stream()
• Added: Tool invocation tracking — toolInvocationCount records actual tool calls made by the agent
• Renamed: Metrics fields for clarity: callCount → llmCallCount, toolCallCount → toolInvocationCount
• Added: agent_metrics_demo.js example to verify metrics recording

v2.1.3 (2026-05-13)

• Previous release

Version: 3.0.2 | Last Updated: 2026-10-07