算力交易中心解读 Agent Plugins 1.0 跨平台兼容:一个 Skill 跑遍 Claude/ChatGPT/Codex 的时代来了

以 Token 聚合与分发为核心的算力交易中心开始进入产业视野。


当技能可以打包带走,算力可以按需取用

2026 年 8 月 6 日,AI 行业六家重要公司——OpenAI、AWS、Microsoft、GitHub、Vercel,以及 Cursor 母公司 Anysphere——罕见地站在了同一张桌子前,共同发布 Agent Plugins 1.0.0 开放规范。

几乎在同一时间,另一条路线也正在成形:以 Token 聚合与分发为核心的算力交易中心开始进入产业视野。

表面上看,两件事毫无关系。

一个解决的是**“AI 技能如何跨平台复用”,另一个解决的是“AI 算力如何跨供应商获取”**。

但如果把视角拉高,就会发现它们正在推动同一场变化:

AI 能力正在从“自己建设”,转向“标准化调用”;从“独占资源”,转向“可复用、可交易的服务”。


一、Agent Plugin 1.0:终结“每个平台重写一遍”的开发噩梦

在 Agent Plugin 出现之前,AI 开发工具生态长期处于碎片化状态。

Cursor 有自己的插件体系,GitHub Copilot 有自己的扩展机制,不同 AI 客户端对 Skill、MCP Server 和配置文件的组织方式也各不相同。

最麻烦的地方在于:

同一个能力,往往需要为不同客户端重新包装。

开发者真正想复用的是 Skill 和 MCP Server 本身,但现实却经常变成:

  • 为 Cursor 做一份;

  • 为 Copilot 再做一份;

  • 换到 ChatGPT,又重新适配;

  • 换一个客户端,又是一套配置。

能力没有变化,重复工作却不断增加。

Agent Plugin 1.0 的思路恰恰相反:先定义一个足够轻量的通用包装标准。

一个插件可以被组织成标准化目录:

plugin/
├── plugin.json
├── skills/
└── mcp.json

其中 plugin.json 负责描述插件基本信息,Skill 放入 skills/,MCP 配置则通过 mcp.json 管理。

开发者按照统一结构进行打包,客户端按照统一结构进行解析。

最直接的价值就是:

一次开发,多端复用。

发布后,ChatGPT/Codex、Cursor、GitHub Copilot、Kiro、VS Code 等主流开发环境陆续支持相关规范,Google 也参与了标准维护生态。

这意味着一个 Skill 不再天然绑定某一个客户端。

开发者真正需要维护的是能力本身,而不是为每个平台重复制作包装。


二、但 Agent Plugin 1.0 只解决了“能不能带走”

值得注意的是,Agent Plugin 1.0 并没有试图解决所有问题。

它重点定义的是:

插件如何组织、如何打包、如何被发现。

至于安装、权限、认证、沙箱、信任验证和安全策略,则更多交给具体客户端处理。

这是一种非常典型的开放标准思路:

先把接口统一,再让生态竞争具体实现。

好处是标准足够轻,开发者可以快速采用。

但它也留下了一个现实问题:

当一个 Agent Plugin 可以跨平台运行时,谁来保证这个插件值得信任?

因此,Agent Plugin 只是 AI 能力标准化的第一步,而不是终点。


三、Unexhub:当算力变成 Token,算力也开始“标准化”

如果说 Agent Plugin 解决的是工具层的可携带性,那么 Token 算力交易中心解决的就是基础设施层的可获取性。

传统算力市场存在一个非常现实的矛盾:

有人有 GPU,却没有足够需求;有人需要算力,却没有 GPU。

对于中小型 AI 供应商而言,自建算力还意味着:

  • 采购 GPU;

  • 搭建机房或云环境;

  • 部署模型;

  • 寻找客户;

  • 等待回款。

大量资金因此被锁在硬件和基础设施中。

Token 聚合与分发模式试图改变这一过程。

它不要求每个企业都自己建设完整算力基础设施,而是将不同供应商、不同模型、不同 API 能力聚合起来,再通过统一的 Token 体系完成计量和结算。

简单来说:

开发者购买的不是一台 GPU,而是完成任务所需要的 AI 能力。

这就像用电。

用户不需要知道电来自哪一座发电厂,只需要按照实际用量支付费用。


四、即时结算:让闲置算力变成现金流

算力交易真正有吸引力的地方,还不只是“能买”。

对于供应商而言,什么时候收到钱同样重要。

传统 B2B 算力交易往往存在较长结算周期。

供应商先提供服务,再等待客户确认、月底对账,最终才能完成回款。

如果采用 Token 化计量和实时结算机制,交易可以进一步拆解为更细的调用单元:

算力被调用 → Token 被消耗 → 服务完成 → 对应收益结算。

这让供应商可以更加灵活地把闲置算力转化为收入。

对于中小供应商而言,这意味着更低的资金占用压力,也意味着 GPU 不必长期躺在机房里等待订单。


五、API 竞价:让算力价格回归供需

Unexhub 进一步设想的,是 API 动态竞价模式。

传统 API 往往采用固定价格:

供应商定价 → 用户购买 → 无论供需如何变化,价格基本不变。

但算力其实是一种高度动态的资源。

凌晨 3 点和下午 3 点,同一块 GPU 的硬件没有变化,但市场需求可能完全不同。

因此,动态竞价可以让价格更加接近实时供需:

空闲时段

算力需求下降 → 供应商降低价格 → 开发者以更低成本运行测试和批量任务。

高峰时段

需求增加 → 稀缺算力价格上升 → 供应商获得更高收益。

这样一来,算力市场就不再只是“统一报价”,而开始具备类似商品市场的价格发现能力。

对于开发者来说,最大的价值不是永远低价,而是:

能够根据任务紧急程度、质量要求和预算,选择更合适的算力价格。


六、两条线最终会在哪里汇合?

现在再回头看 Agent Plugin 和 Token 算力交易中心,就会发现它们其实正在解决 AI 产业链上两个相互对应的问题。

层级过去正在发生的变化
工具层每个平台重复开发Plugin 标准化
模型层单一模型绑定多模型调用
算力层自建 GPU / 固定供应商Token 化、市场化
交易层固定报价动态竞价
应用层自己开发和部署Agent 上架、调用和交易

最终形成的是一条新的 AI 能力链:

标准化技能 → 模型调用 → Token 计量 → 算力交易 → Agent 应用。

过去做一个 AI 应用,更像是“造汽车”。

你需要自己采购发动机、搭底盘、建工厂,再寻找销售渠道。

未来则更像“组装汽车”。

需要什么能力,就调用什么能力;需要什么模型,就接入什么模型;需要多少算力,就支付多少 Token。


七、Unexhub 的角色:把“找模型”变成“找能力”

这也是 Unexhub 想切入的核心位置。

Unexhub 定位于 Token 聚合与分发的算力交易中心,希望把不同规模的 AI 供应商汇聚到统一市场。

目前规划覆盖包括 Gemini、ChatGPT、Claude、DeepSeek、GLM 等主流模型,同时也面向个人开发者和中小型 AI 供应商开放。

对开发者而言,核心价值并不是“多一个 API”。

而是:

一个入口,对接更多模型;根据价格与性能进行路由;按照实际 Token 用量结算。

在价格方面,Unexhub 计划通过供应商竞争和 API 竞价机制降低使用门槛,开发者按 Token 实际消耗付费,部分场景最低可达到市场价格的 3 折。

对于高频调用的团队来说,这种价格差异会被放大。

如果每天需要处理大量 Token,那么哪怕单次调用只节省一小部分,长期累计也可能形成非常明显的成本优势。

与此同时,供应商也可以通过标准化接口快速接入渠道,并按照实际调用量获得分成,实现供给、调用、结算之间的闭环。


八、早期供应商为什么值得关注?

对于算力和模型供应商来说,平台冷启动阶段往往是最关键的窗口。

Unexhub 针对早期入驻供应商提供最低 0 佣金的合作政策。

这意味着早期供应商可以在平台生态形成阶段,以更低的平台成本获得:

  • 更早进入 Token 算力交易市场;

  • 获取开发者和企业用户;

  • 通过 API 竞价参与市场价格发现;

  • 按实际调用获得收益;

  • 实现更灵活的渠道分发和结算。

从平台角度看,这是一种典型的生态冷启动策略:

先扩大供给 → 再吸引需求 → 形成交易 → 最终建立市场流动性。

如果未来 Token 真正成为 AI 能力的标准计量单位,那么早期进入供应侧生态的价值,就不仅仅是“多一个销售渠道”。

而可能意味着:

提前占据 AI 能力交易市场的供给入口。


九、从“自建”到“调用”,AI 正在进入下一阶段

Agent Plugin 1.0 被很多人视为 AI 插件生态迈向标准化的重要一步。

而 Token 算力交易中心,则试图让算力也从专业基础设施变成一种可以直接调用的市场化资源。

二者叠加之后,AI 应用开发的逻辑可能发生根本变化:

技能可以打包带走。

模型可以按需切换。

算力可以按量购买。

Agent 可以直接上架交易。

这意味着 AI 能力正在从“企业自己建设”逐渐走向“生态共同提供”。

从“卖 GPU”到“卖 Token”。

从“卖模型”到“卖能力”。

从“购买软件席位”到“购买智能体完成的结果”。

真正的 AI 应用市场,也许并不是某一个超级模型最终统治一切,而是越来越多的能力被拆分、标准化、定价,并最终像 App 一样被调用。

当技能可以打包带走,算力可以按需取用,AI 应用的门槛也就真正开始下降。


English Version

When Skills Become Portable and Compute Becomes On-Demand

English readers: the full English version starts here. ↓ Keep scrolling for the complete article.

On August 6, 2026, six major players in the AI industry—OpenAI, AWS, Microsoft, GitHub, Vercel, and Anysphere, the company behind Cursor—came together around an unusual common standard: Agent Plugins 1.0.0.

At almost the same time, another development was taking shape: the emergence of compute trading centers built around Token aggregation and distribution.

At first glance, these two developments appear unrelated.

One addresses how AI capabilities can be reused across platforms.

The other addresses how AI compute can be accessed across different suppliers.

But at a higher level, they point toward the same structural shift:

AI capabilities are moving from something companies build themselves to something they can access through standardized services—from proprietary resources to reusable and tradable capabilities.


1. Agent Plugin 1.0: Ending the “Rebuild It for Every Platform” Problem

Before Agent Plugins, the AI development ecosystem was highly fragmented.

Cursor had its own plugin system. GitHub Copilot had its own extension mechanism. Different AI clients also used different structures for Skills, MCP Servers, configuration files, and discovery mechanisms.

The frustrating part was simple:

The capability itself did not change, but developers often had to package it again for every client.

The same Skill might need one implementation for Cursor, another configuration for Copilot, and yet another adaptation for ChatGPT.

Agent Plugin 1.0 takes the opposite approach: define a lightweight, portable packaging standard.

A plugin can follow a standardized structure:

plugin/
├── plugin.json
├── skills/
└── mcp.json

The plugin.json file describes the basic plugin information, Skills live under skills/, and MCP configuration can be defined through mcp.json.

The result is straightforward:

Build once, reuse across multiple clients.

Major development environments, including ChatGPT/Codex, Cursor, GitHub Copilot, Kiro, and VS Code, have moved toward supporting the specification, while Google has also joined the broader standards ecosystem.

The important change is that a Skill no longer has to be permanently tied to a single client.

Developers can focus on maintaining the capability itself, rather than repeatedly rebuilding its packaging.


2. But Agent Plugin 1.0 Only Solves Portability

Agent Plugin 1.0 deliberately does not attempt to solve every problem.

Its primary focus is:

How plugins are packaged and discovered.

Installation, permissions, authentication, sandboxing, trust verification, and security policies remain largely the responsibility of individual clients.

That is a classic open-standard strategy:

Standardize the interface first, then let the ecosystem compete over implementation.

The advantage is speed and simplicity.

The downside is equally obvious:

If an Agent Plugin can run across multiple platforms, who decides whether that plugin should be trusted?

Agent Plugin 1.0 is therefore an important first step toward standardized AI capabilities—not the final destination.


3. Unexhub: When Compute Becomes Tokenized

If Agent Plugin addresses portability at the tool layer, a Token-based compute trading center addresses accessibility at the infrastructure layer.

The problem in the traditional compute market is straightforward:

Some providers have GPUs but lack demand. Others need compute but do not have enough infrastructure.

For smaller AI suppliers, building infrastructure also means:

  • purchasing GPUs;

  • deploying infrastructure;

  • hosting models;

  • finding customers;

  • waiting for payments.

Capital can remain locked in hardware for months.

A Token aggregation and distribution model attempts to change this process.

Instead of requiring every company to build its own complete AI infrastructure, different suppliers, models, and API capabilities can be aggregated and accessed through a unified Token-based measurement and settlement system.

In simple terms:

Developers are not necessarily buying a GPU. They are buying the AI capability required to complete a task.

It is similar to electricity.

Users do not need to know which power plant generated the electricity. They simply pay according to usage.


4. Instant Settlement: Turning Idle Compute into Cash Flow

The appeal of compute trading is not only about access.

For suppliers, when they get paid matters just as much.

Traditional B2B compute transactions can involve long settlement cycles.

Suppliers deliver the service first, wait for confirmation and reconciliation, and only then receive payment.

With Token-based metering and real-time settlement, the transaction can be broken into much smaller units:

Compute is consumed → Tokens are used → Service is delivered → Revenue is settled.

This gives smaller suppliers a more flexible way to monetize idle infrastructure.

Instead of leaving GPUs underutilized while waiting for large contracts, suppliers can potentially turn unused capacity into incremental revenue.


5. API Bidding: Letting Supply and Demand Set the Price

Unexhub is also exploring a dynamic API bidding model.

Traditional API pricing is usually fixed:

Supplier sets the price → customer pays the price → supply and demand have limited influence on the final rate.

But compute is a highly dynamic resource.

The same GPU can have completely different economic value at 3 a.m. and 3 p.m.

A dynamic bidding mechanism can therefore make pricing more responsive to market conditions.

During low-demand periods

Lower demand → lower prices → developers can run experiments and batch workloads at lower cost.

During peak periods

Higher demand → higher prices → suppliers receive greater returns for scarce capacity.

The value for developers is not necessarily that prices are always low.

It is that developers can potentially choose compute according to urgency, performance requirements, and budget.


6. Where Do These Two Trends Meet?

Agent Plugins and Token-based compute trading may look like two separate developments.

But they address two complementary problems across the AI stack.

LayerTraditional ModelEmerging Model
Tool layerRebuild for each platformStandardized Plugins
Model layerDepend on one modelMulti-model access
Compute layerBuild GPUs / rely on one supplierTokenized, market-based compute
Pricing layerFixed pricingDynamic bidding
Application layerBuild and deploy yourselfPublish, access, and trade Agents

Together, they form a new AI capability chain:

Standardized skills → model access → Token metering → compute trading → Agent applications.

Building an AI application used to feel like manufacturing a car.

You had to source the engine, build the chassis, construct the factory, and establish distribution.

The emerging model looks more like assembling a car.

Need a specific capability? Call it.

Need a different model? Switch it.

Need more compute? Purchase it.

Need to distribute an Agent? Publish it.


7. Unexhub: Turning “Find a Model” into “Find a Capability”

This is where Unexhub aims to position itself.

Unexhub is designed as a Token aggregation and distribution center for AI compute, bringing together AI suppliers of different sizes within a unified marketplace.

The planned ecosystem includes major models such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, while also opening the market to individual developers and smaller AI providers.

For developers, the value is not simply having another API endpoint.

It is having:

one access point, multiple models, intelligent routing, and usage-based Token settlement.

Through supplier competition and API bidding, Unexhub aims to lower the cost of AI consumption. Developers can pay based on actual Token usage, with selected scenarios potentially reaching as low as 30% of prevailing market prices.

For high-volume AI applications, even a relatively small per-call saving can become significant when multiplied across millions or billions of Tokens.

Suppliers, meanwhile, can connect their APIs to distribution channels and receive revenue based on actual usage, creating a more complete loop between supply, consumption, and settlement.


8. Why Early Suppliers May Have an Advantage

The early stage of a marketplace is often the most important period for suppliers.

Unexhub plans to offer early participating suppliers as low as 0% platform commission.

That can provide early suppliers with several practical benefits:

  • earlier access to the emerging Token compute marketplace;

  • exposure to developers and enterprise customers;

  • participation in API-based market pricing;

  • revenue based on actual usage;

  • more flexible distribution and settlement.

From the platform's perspective, this is a classic ecosystem bootstrapping strategy:

Build supply first → attract demand → generate transactions → establish liquidity.

If Token eventually becomes a standard unit for measuring AI capabilities, early participation in the supply ecosystem could mean more than simply gaining another sales channel.

It could mean:

Getting an early position in the supply side of an emerging AI capability marketplace.


9. From “Build It Yourself” to “Call It When You Need It”

Agent Plugin 1.0 represents an important step toward standardizing the AI plugin ecosystem.

Token-based compute marketplaces attempt to make compute itself more accessible, measurable, and tradable.

Together, they point toward a fundamental change in how AI applications are built:

Skills can be packaged and carried across platforms.

Models can be switched on demand.

Compute can be purchased according to usage.

Agents can be published and traded.

AI capabilities are therefore moving from something every company must build independently toward something increasingly supplied by an ecosystem.

From selling GPUs to selling Tokens.

From selling models to selling capabilities.

From buying software seats to buying outcomes produced by intelligent agents.

The future of the AI application market may not be about one super-model dominating everything.

It may instead be about thousands of specialized capabilities being standardized, priced, distributed, and accessed like applications.

When skills become portable and compute becomes on-demand, the barrier to building AI applications begins to fall for real.