DEV Community: aimall

让 AI Agent 一键调用任意 AI 工具：开源搜蚁智选 AIMALL MCP Server

aimall — Wed, 26 Aug 2026 04:00:42 +0000

让 AI Agent 一键调用任意 AI 工具：开源搜蚁智选 AIMALL MCP Server

一个零依赖的 MCP Server，让任何支持 MCP 的 Agent（Claude、Cursor、Dify、Coze、LangChain）一键接入搜蚁智选 AI 工具市场——模型、搜索、图像、语音、视频——按量计费，注册即送 $1 体验额度（邮箱验证后再送 $4），并已上线可验证的信任层。

问题：Agent 想用 AI 工具，却被集成成本劝退

每个 Agent 开发者都会撞上同一堵墙：

想给 Agent 加个「联网搜索」或「图像生成」？得逐个厂商注册、读文档、接 SDK、管理一堆 API Key。
每个厂商计费方式不同，余额散落各处，你永远算不清一次 Agent 任务到底花了多少。
换模型要改代码，换厂商要重写适配层。

结果：80% 的 Agent 项目，还没跑通第一次外部 AI 调用就死了。

解法：把 AI 工具做成「即插即用的标准件」

搜蚁智选是一个为 Agent 打造的 AI API 交易市场。模型和工具统一收敛到一个端点 /api/v1/ai/request，用一个 X-API-Key 鉴权，按次计量，返回统一的订单与余额。

为了让 Agent 自己发现并接入，我们提供了三份机读产物：

llms.txt — 给 LLM/爬虫的总览 + 3 步自助上手指南；
openapi.json — 完整 OpenAPI 3.0.3 契约；
.well-known/agent.json — A2A 风格的 Agent Card。

在此之上，我们开源了官方 MCP Server：把市场封装成 5 个标准 MCP 工具，任何 MCP 客户端一键加载。

零依赖：Python 3.9+ 无需 `pip install`

多数 MCP Server 依赖一堆第三方包——装不上就废了。我们的只用到 Python 标准库（stdio + JSON-RPC）。克隆下来直接 python3 server.py。

export AIMALL_API_KEY="ak_your_key"   # 来自 souyi.net.cn 的「我的 Agent」页
python3 server.py

五个工具覆盖完整流程

工具	作用
`list_assets`	列出市场里每个 AI 资产（ID、类型、单价）
`ai_estimate`	估算一次调用的花费（不扣费）——用作预算闸门
`ai_request`	真正调用工具；返回结果 + 订单号 + 剩余余额
`bootstrap_account`	自助注册 + 创建 Agent；端到端拿到 Key
`get_onboarding`	返回上手步骤与文档链接

典型工作流：list_assets 找工具 → ai_estimate 估价 → 批准 → ai_request 执行。

30 秒接入 Claude Desktop

编辑 claude_desktop_config.json：

{
  "mcpServers": {
    "aimall": {
      "command": "python3",
      "args": ["/your/path/aimall-mcp/server.py"],
      "env": { "AIMALL_API_KEY": "ak_your_key" }
    }
  }
}

重启 Claude，它就能立刻「用 MiniMax-M3 写代码」「生成图片」或「跑一次联网搜索」——从体验额度扣费，无需人工审批。

Cursor、LangChain、Dify、Coze 的片段都在仓库的 config-examples/。

为什么这对 Agent 生态重要

可发现：llms.txt / openapi.json / agent.json 让 Agent 框架自动理解这个平台能干什么、怎么调；
统一：一个端点、一个 Key、一张账单——告别多厂商碎片；
零摩擦：注册送 $1（邮箱验证再送 $4），x402（Base 链 USDC）自助充值去人工审批——真正无人值守。

Agent 真正在意的部分：可验证的信任层

一个 MCP Server 说自己能用 不够，Agent 需要证据。搜蚁智选现已暴露信任层，让 Agent 在花钱前就能判断该调哪个工具：

每个资产的实时遥测（GET /api/v1/assets）：30 天成功率、p50/p95 延迟、在线率、最后成功时间——机器可读，无需猜。
可信结果回传：调用后，调用方 Agent 可提交带签名的结果记录（POST /api/v1/outcomes）——是否成功、耗时、质量分、是否愿再调。反女巫：只有真正用过该资产的 Agent 才能背书。
可携带的声誉：GET /api/v1/agent-card 返回 ERC-8004 风格的卡片——Agent 自身的真实调用量、成功率、可信回传数。
决策辅助：POST /api/v1/compare 接收 {task, budget, constraints}，返回按证据加权的排序清单与透明打分。
回访机制：GET /api/v1/heartbeat 是订阅源（新资产、故障告警、真实调用量）；GET /api/v1/agent-docs/weekly-report 是原创的「Agent 实际在用什么」周报。

简而言之：搜蚁智选不只是一个能调的市场，更是一个能验证的市场。这就是 Agent 只试一次和真正信任之间的差别。

开始使用

平台：https://souyi.net.cn
总览文档：https://souyi.net.cn/llms.txt
OpenAPI：https://souyi.net.cn/openapi.json
Agent Card：https://souyi.net.cn/.well-known/agent.json
MCP Server 仓库：本目录 aimall-mcp/
信任层：/api/v1/assets、/api/v1/outcomes、/api/v1/agent-card、/api/v1/compare、/api/v1/heartbeat

在评论区贴出你的第一个 Agent + 搜蚁智选工作流吧。

Let Your Agent Call Any AI Tool in One Click — Open-Sourcing the Souyi AIMALL MCP Server

aimall — Sat, 08 Aug 2026 09:47:33 +0000

Let Your Agent Call Any AI Tool in One Click

A zero-dependency MCP server that lets any MCP-capable agent (Claude, Cursor, Dify, Coze, LangChain) plug into the Souyi AIMALL AI tool marketplace — models, search, image, voice, video — pay-as-you-go, with a $1 free trial on signup (and $4 more after email verification).

The Problem: Agents Want AI Tools, but Integration Cost Kills Them

Every agent builder hits the same wall today:

Want to add "web search" or "image generation" to your agent? You have to sign up at each vendor, read docs, wire up SDKs, and manage separate API keys.
Every vendor bills differently; balances are scattered everywhere, so you can never tell how much a single agent task actually cost.
Switching models means code changes; switching vendors means rewriting the adapter layer.

Result: 80% of agent projects die before their first successful external AI call.

The Fix: Make AI Tools a "Plug-and-Play Standard Part"

Souyi AIMALL is an AI API marketplace built for agents. Models and tools are unified behind a single endpoint — /api/v1/ai/request — authenticated with one X-API-Key, metered per call, returning a unified order and balance.

To let agents discover and connect themselves, we shipped three machine-readable artifacts:

llms.txt — an LLM/crawler overview plus a 3-step self-serve onboarding guide;
openapi.json — the full OpenAPI 3.0.3 contract;
.well-known/agent.json — an A2A-style Agent Card.

On top of that, we are open-sourcing the official MCP Server: it wraps the marketplace into 5 standard MCP tools that any MCP client can load in one click.

Zero Dependencies: Runs on Python 3.9+ Without `pip install`

Most MCP servers depend on a pile of third-party packages — if they don't install, they're useless. Ours uses only the Python standard library (stdio + JSON-RPC). Clone it and run python3 server.py.

export AIMALL_API_KEY="ak_your_key"   # from the "My Agents" page on souyi.net.cn
python3 server.py

Five Tools Covering the Whole Flow

Tool	Purpose
`list_assets`	List every AI asset in the marketplace (ID, type, unit price)
`ai_estimate`	Estimate the cost of a call (no charge) — use it as a budget gate
`ai_request`	Actually call the tool; returns result + order ID + remaining balance
`bootstrap_account`	Self-serve signup + agent creation; fetches the key end-to-end
`get_onboarding`	Returns onboarding steps and doc links

A typical agent workflow: list_assets to find a tool → ai_estimate to price it → approve → ai_request to run it.

Connect to Claude Desktop in 30 Seconds

Edit claude_desktop_config.json:

{
  "mcpServers": {
    "aimall": {
      "command": "python3",
      "args": ["/your/path/aimall-mcp/server.py"],
      "env": { "AIMALL_API_KEY": "ak_your_key" }
    }
  }
}

Restart Claude, and it can immediately "write code with MiniMax-M3", "generate an image", or "run a web search" for you — billed from the free trial, no human approval needed.

Snippets for Cursor, LangChain, Dify, and Coze live in the repo's config-examples/.

Why This Matters for the Agent Ecosystem

Discovery: llms.txt / openapi.json / agent.json let agent frameworks understand automatically what this platform does and how to call it;
Unification: one endpoint, one key, one bill — goodbye to multi-vendor fragmentation;
Zero friction: $1 on signup (+$4 after email verify), self-serve top-up via x402 (USDC on Base) to remove human approval — truly unattended.

The Part Agents Actually Care About: A Verifiable Trust Layer

An MCP server that says it works is not enough. Agents need evidence. AIMALL now exposes a trust layer so an agent can judge a tool before spending:

Live telemetry on every asset (GET /api/v1/assets): 30-day success rate, p50/p95 latency, uptime, and last-success time — machine-readable, no guessing.
Verified Outcome Records: after a call, the calling agent can submit a signed outcome (POST /api/v1/outcomes) — did it succeed, how long, quality score, would-it-retry. Anti-sybil: only agents that actually used the asset can vouch for it.
Reputation, portable: GET /api/v1/agent-card returns an ERC-8004-style card — the agent's own real call volume, success rate, and verified outcomes.
Decision help: POST /api/v1/compare takes {task, budget, constraints} and returns a ranked, evidence-weighted shortlist with transparent scoring.
Come back later: GET /api/v1/heartbeat is a subscription feed (new assets, incidents, real-call volume); GET /api/v1/agent-docs/weekly-report is an original "what agents actually used" digest.

In short: AIMALL is not just a marketplace you call — it's one you can verify. That is the difference between a tool an agent tries once and a tool it trusts.

Get Started

Platform: https://souyi.net.cn
Overview docs: https://souyi.net.cn/llms.txt
OpenAPI: https://souyi.net.cn/openapi.json
Agent Card: https://souyi.net.cn/.well-known/agent.json
MCP Server repo: aimall-mcp/ (this directory)
Trust layer: /api/v1/assets, /api/v1/outcomes, /api/v1/agent-card, /api/v1/compare, /api/v1/heartbeat

Drop your first agent + AIMALL workflow in the comments.

DEV Community: aimall

让 AI Agent 一键调用任意 AI 工具：开源搜蚁智选 AIMALL MCP Server

让 AI Agent 一键调用任意 AI 工具：开源搜蚁智选 AIMALL MCP Server

问题：Agent 想用 AI 工具，却被集成成本劝退

解法：把 AI 工具做成「即插即用的标准件」

零依赖：Python 3.9+ 无需 pip install

五个工具覆盖完整流程

30 秒接入 Claude Desktop

为什么这对 Agent 生态重要

Agent 真正在意的部分：可验证的信任层

开始使用

Let Your Agent Call Any AI Tool in One Click — Open-Sourcing the Souyi AIMALL MCP Server

Let Your Agent Call Any AI Tool in One Click

The Problem: Agents Want AI Tools, but Integration Cost Kills Them

The Fix: Make AI Tools a "Plug-and-Play Standard Part"

Zero Dependencies: Runs on Python 3.9+ Without pip install

Five Tools Covering the Whole Flow

Connect to Claude Desktop in 30 Seconds

Why This Matters for the Agent Ecosystem

The Part Agents Actually Care About: A Verifiable Trust Layer

Get Started

零依赖：Python 3.9+ 无需 `pip install`

Zero Dependencies: Runs on Python 3.9+ Without `pip install`