[](https://github.com/arian-gogani/nobulex/actions/workflows/ci.yml)
[](https://www.bestpractices.dev/projects/10338)
[](https://pypi.org/project/nobulex/)
[](https://www.npmjs.com/package/@nobulex/core)
[](https://opensource.org/licenses/MIT)
[](./drafts/draft-gogani-nobulex-proof-of-behavior-00.txt)
**每个人都有信用评分。每家企业都有。**
**而 AI agents 却一无所有。**
Nobulex 是面向自主 AI agents 的信用与信任协议。
Agents 通过经过验证的行为赚取 Trust Capital。信任度越高,访问权限越多。
自主权是赚取的,而非被授予的。
[官网](https://nobulex.com) · [在线体验](https://nobulex.com/demo) · [快速开始](./GETTING-STARTED.md) · [规范](./drafts/draft-gogani-nobulex-proof-of-behavior-00.txt) · [PyPI](https://pypi.org/project/nobulex/) · [npm](https://www.npmjs.com/package/@nobulex/core)
## 安装
```
pip install nobulex
```
```
npm install @nobulex/core
```
```
from nobulex.agent import Agent
agent = Agent("my-agent")
receipt = agent.act("send_email", scope="user@example.com")
assert receipt.verify() # tamper-proof
```
## 工作原理
每一次 agent 操作都会产生一个加密凭证 —— 在执行前后均进行 Ed25519 签名,并通过哈希链相连以提供防篡改证据。第三方可以在不信任 agent 或运营者的情况下验证完整历史记录。
以下是一次运行中的完整理念(`python -m nobulex demo`):
```
generated 3 receipts
allow: 141ca2947a7e819b8bdebbf8... verified=True
allow: f3377758ac94d812535cbb99... verified=True
deny: 85b2dfd6b87f2678795726e4... verified=True
trust score: 23.26
tamper test:
modified receipt verified=False (tamper detected)
```
修改凭证中的任何一个字节,验证都会失败。这就是全部的保证。
**性能:** 在 p50 下约为 13,683 个签名凭证/秒(Python SDK,单核)。完整的签名与链接凭证端到端耗时约 73 μs。请参阅 [BENCHMARKS.md](./docs/BENCHMARKS.md) 了解详细分解;可通过 `python3 scripts/benchmark.py` 进行复现。
凭证会累积成 **Trust Capital** —— agent 的信用评分。
| 等级 | Trust Capital | 访问级别 |
|------|--------------|--------------|
| Restricted | 0 -- 30 | 只读,沙盒化执行 |
| Standard | 30 -- 60 | 最高 $500 的财务操作,API 访问权限 |
| Trusted | 60 -- 85 | 跨组织操作,受监管市场 |
| Sovereign | 85+ | 完全自主,自我主导 |
创造更多价值的 agents 会获得更多访问权限。发生偏离的 agents 会被自动切断。这不是惩罚——而是数学。
## 快速开始
### Python(推荐用于 AI agents)
```
# 或从源码安装
pip install git+https://github.com/arian-gogani/nobulex.git#subdirectory=packages/python
```
```
# 或立即尝试 CLI demo
git clone https://github.com/arian-gogani/nobulex.git
cd nobulex/packages/python
pip install -e .
python -m nobulex demo
```
```
from nobulex import Agent
agent = Agent("my-agent")
receipt = agent.act("send_email", scope="user@example.com")
assert receipt.verify() # Cryptographic proof
print(agent.trust_score) # Trust Capital: 13.86
```
#### LangChain 集成
```
from nobulex.integrations.langchain import NobulexAuditHandler
handler = NobulexAuditHandler(agent_id="my-agent")
agent.invoke({"input": "..."}, config={"callbacks": [handler]})
handler.export("audit.json") # signed, hash-chained audit trail
```
#### CrewAI 集成
```
from nobulex.integrations.crewai import NobulexCrewAudit
audit = NobulexCrewAudit(agent_id="my-crew")
audit.record_task("credit_check", "loan-app-4821")
audit.export("audit.json")
```
#### Google ADK 集成
```
from nobulex.integrations.google_adk import NobulexADKCallback
cb = NobulexADKCallback(agent_id="my-agent")
@cb.wrap_tool("web_search")
def search(query):
return do_search(query)
cb.export("audit.json")
```
#### PydanticAI 集成
```
from nobulex.integrations.pydantic_ai import NobulexPydanticAIAudit
audit = NobulexPydanticAIAudit(agent_id="typed-agent")
audit.record_tool("get_weather", {"city": "Berlin"}, {"temp": 22})
audit.export("audit.json")
```
#### Haystack 集成
```
from nobulex.integrations.haystack import NobulexHaystackAudit
audit = NobulexHaystackAudit(agent_id="my-pipeline")
audit.record_component("Retriever", {"query": "test"}, {"docs": 5})
audit.export("audit.json")
```
#### LlamaIndex 集成
```
from nobulex.integrations.llama_index import NobulexLlamaIndexAudit
audit = NobulexLlamaIndexAudit(agent_id="llama-agent")
audit.record_tool("web_search", {"query": "test"}, {"results": 5})
audit.export("audit.json")
```
#### 验证任何导出的记录(无需信任运营者)
```
from nobulex.chain import verify_audit_trail
report = verify_audit_trail("audit.json", authorized_keys=AGENT_PUBLIC_KEY)
assert report["chain_intact"] and report["authenticated"]
```
### JavaScript / TypeScript
```
npm install @nobulex/core
npx tsx examples/trust-capital-demo.ts
```
```
Agent starts at RESTRICTED tier (Trust Capital: 0)
Action 1: read_data — ALLOWED (Trust Capital: 12)
Action 2: read_data — ALLOWED (Trust Capital: 24)
Action 3: process_payment — BLOCKED (insufficient trust)
Action 4: read_data — ALLOWED (Trust Capital: 36)
Action 5: read_data — ALLOWED (Trust Capital: 48)
Agent promoted to STANDARD tier
Action 6: process_payment — ALLOWED (Trust Capital: 65)
Agent promoted to TRUSTED tier (Trust Capital: 89)
Action 8: approve_contract — ALLOWED
```
## 协议
```
DECLARE ──► ENFORCE ──► PROVE ──► ACCUMULATE
Covenant Pre-execution Receipt chain Trust Capital
defines receipt blocks verified by earned over
the rules violations third parties time
before they
happen ──► more access
──► more receipts
──► higher trust
```
飞轮效应:更多的 Trust Capital 带来更有价值的工作,从而产生更多凭证,进而建立更高的 Trust Capital。问责制由此成为最具盈利能力的策略。
## 代码
```
import { createDID, parseSource, EnforcementMiddleware, verify } from '@nobulex/core';
const agent = await createDID();
const spec = parseSource(`
covenant SafeTrader {
permit read;
permit transfer (amount <= 500);
forbid transfer (amount > 500);
forbid delete;
}
`);
const mw = new EnforcementMiddleware({ agentDid: agent.did, spec });
await mw.execute(
{ action: 'transfer', params: { amount: 300 } }, // allowed
async () => ({ success: true }),
);
await mw.execute(
{ action: 'transfer', params: { amount: 600 } }, // BLOCKED before execution
async () => ({ success: true }), // never runs
);
const result = verify(spec, mw.getLog());
console.log(result.compliant); // true
```
## 发展情况
独立、可验证的信号(每一项均附有证据链接):
| | 内容 | 证据 |
|---|---|---|
| | **OWASP CheatSheetSeries** | 第 8-11 节(JCS 规范化、跨 agent 问责制、制裁名单时效性、监管映射)由 Jim Manico 于 2026 年 6 月合并至 master 分支([PR #2210](https://github.com/OWASP/CheatSheetSeries/pull/2210)) |
| | **vaara v0.50** | 独立第三方采用 —— Henri Sirkkavaara 发布了欧盟《AI 法案》第 12 条的审计记录,并引用了 nobulex 的签名凭证设计([GitHub](https://github.com/vaaraio/vaara)) |
| | **Dify Marketplace** | 在 60,000+ star 的平台上开放了插件 PR([PR #2500](https://github.com/langgenius/dify-plugins/pull/2500))。LangGenius 社区运营确认其架构可靠,且 Trust Capital 具有真正的差异化优势。 |
| | **Microsoft AI Agents for Beginners** | 已提交 PR,拟在第 18 课《使用加密凭证保护 AI Agents》中将 nobulex 添加为 Python 生产级凭证库([PR #571](https://github.com/microsoft/ai-agents-for-beginners/pull/571)) |
| | **AgentAudit AI** | 五点整合合作伙伴关系生效。签名样本凭证可在 10 行 Python 代码中实现端到端验证([fixture](./fixtures/agentaudit-specimen-v1.json)) |
| | **Microsoft AGT** | 列入 [ADOPTERS](https://github.com/microsoft/agent-governance-toolkit/pull/1703)(PR 由 Microsoft 维护者合并) |
| | **builderz-labs / mission-control** | 跨会话 Trust Capital RFC 作为公开 issue 被接受;已交付 TypeScript 参考实现 |
欧盟《AI 法案》第 12 条执行日期:2026 年 8 月 2 日。
## 为什么是现在
AI agents 正在部署到生产环境中,但却缺乏问责基础设施。
- **86%** 的 AI agents 在部署时未经安全审批 (CSA, 2026)
- **UUMit** 推出了首个无任何身份验证的 A2A 市场
- **$138B+** 投入到物理 AI 中,且零问责层
- 顶级模型在处理实际问题时得分仅为 **10-15%** (LemmaBench),且在失败时毫无可追溯性
Agents 已经部署。资金正在涌入。而问责基础设施却尚未存在。我们正在构建它。
## 标准
| 标准 | 状态 |
|---|---|
| Proof-of-Behavior 规范 | [`draft-gogani-nobulex-proof-of-behavior-00`](./drafts/draft-gogani-nobulex-proof-of-behavior-00.txt) |
| Microsoft AGT | 列入 [ADOPTERS](https://github.com/microsoft/agent-governance-toolkit/pull/1703)(PR 已合并) |
| CTEF v0.3.2 | 14/14 字节匹配一致性 |
| A2A Protocol | 凭证行已提议;URN 方案 `urn:nobulex:receipt:
` |
| NIST RFI | 已提交正式意见 |
## 开发
```
git clone https://github.com/arian-gogani/nobulex.git
cd nobulex && npm install
npx vitest run # tests
npx tsx examples/demo.ts # end-to-end
npx tsx benchmarks/bench.ts # benchmarks
```
[官网](https://nobulex.com) · [体验](https://nobulex.com/demo) · [npm](https://www.npmjs.com/package/@nobulex/core) · [规范](./drafts/draft-gogani-nobulex-proof-of-behavior-00.txt) · [X @nobulexlabs](https://x.com/nobulexlabs)
**[为本项目加星](https://github.com/arian-gogani/nobulex/stargazers)** 以持续关注
MIT License