Runtime governance & proof
Decide what your AI is allowed to do. Prove it.
Policies, sovereignty, budgets and defensible proof for every AI call — one runtime layer, zero rewrites.
- Designed for the EU AI Act
- Hash-chained journal
- EU residency
- France 2030 · candidate
- Data policy — no forbidden categories
- Sovereignty — EU destination confirmed
- Connector — admin-approved
- Weekly budget — team cap reached
a1f4c09e7b2d8f60Every decision — granted or refused — is traced, sealed, defensible.
The problem
Your agents act. Nothing stops them, nothing proves it.
Observability tells you what your AI did — after the fact. A regulator, a CFO or a customer asks something else: what it was allowed to do, and proof the rule held.
Your agents spend before asking
Budgets get observed at month-end. They don't apply at call time.
The AI Act demands evidence, not logs
A log can be edited. Evidence can be defended. Your current tools produce logs.
Sovereignty is promised, not enforced
Nothing technically stops an EU agent from calling a US endpoint.
Nobody can say what was refused
The history of what was blocked — and why — lives nowhere. It's exactly what the auditor asks for.
四个问题
当智能体做出错误决策时，你能回答这四个问题吗？
不是在一个月的取证之后——而是现在，从一条密封的记录中。
是哪个模型做出的决策？
所选模型被密封写入决策轨迹。
基于哪些数据？
每次调用都会记录所应用的数据策略。
成本是多少？
真实成本与预算检查，在支出之前即被记录。
依据什么授权？
主权、连接器与上限——通过校验，否则调用被拒绝。
How it works
One line to change. Three acts.
OpenAI-compatible. Your SDKs, your frameworks, your code — unchanged.
- client = OpenAI(base_url="https://api.openai.com/v1")
+ client = OpenAI(base_url="https://gateway.azothos.ai/v1")Point your existing client at the gateway. That's all.
policy: "support-bot-guardrails"
rules:
- budget: 250 € / week / team # blocks, doesn't alert
- residency: eu-only # refuses outside EU
- data: deny(pii.health)Declare the rule once. It applies to every request, before spend.
$ curl gateway.azothos.ai/v1/journal/export?period=Q2Export the signed evidence. Verifiable by anyone — no need to take our word.
The architecture
A gateway routes. Azoth filters.
Between your applications and the providers, every call is evaluated. What passes is optimized. What doesn't is refused — with proof.
- Policy engine
- Sovereignty
- Fail-closed budgets
- Sealed journal
- Exportable proof
request refused → signed evidence · nothing sent, nothing spent
定位
Azoth OS 是什么——不是什么。
Azoth OS 是
- 面向你的模型与智能体的编排层
- 治理与策略执行层
- 可观测性、可追溯性与优化工具
- 生产基础设施，而非演示
Azoth OS 不是
- 一个新的 AI 模型
- 一个简单的聊天机器人
- 部署之后附加的仪表盘
- 一个抽象的合规承诺
Five minutes, zero refactor