UniversalBench: run code, web, databases, and any LLM from your AI, safely

MCP开发工具

提供方 nikhilgogulwar

开发工具 未标注 OAuth 授权 社区 / 第三方 ClaudeCursorGemini CLICodex

An MCP execution platform that gives any AI real infrastructure to act. Through three tools it can run code and shell commands, install packages, run tasks in parallel, search the web, make HTTP calls, read write and query databases, call any LLM, commit to GitHub, take screensho

一句话选型结论

按资料字段整理

UniversalBench: run code, web, databases, and any LLM from your AI, safely 属于开发工具类 MCP 服务器,传输方式未标注,OAuth 授权,可直接接入 Claude、Cursor、Gemini CLI 等,没有现成安装命令,需按官方文档接入。接入前先确认权限范围;涉及写入、删除或付款的操作,建议先在测试环境验证。

分类开发工具 传输方式未标注 鉴权OAuth 授权 部署方式未标注
GitHub 星标 未标注 仓库热度,无公开仓库时显示未标注
工具数 未标注 暴露给模型的工具数量
资源 / 提示词 0 / 0 资源与提示词模板数量
维护方 社区 / 第三方 核验等级 仅资料核对

工具与能力

开发工具
数量来自来源目录的标注,具体工具清单以官方文档为准
工具数量未标注暴露给模型可调用的工具个数
资源数量0可读取的资源条目
提示词数量0内置提示词模板数量

能力标签

支持客户端

适用场景

mcpcode executionai agents

安装与配置

未标注 OAuth 授权
上游未给出传输方式
  1. 1
    准备运行环境

    按官方文档确认运行环境(常见为 Node.js 或 Python),装好后再执行安装命令。

  2. 2
    安装服务

    复制上面的安装命令执行;命令里带占位符的位置,换成自己的密钥或路径。

  3. 3
    写入客户端配置

    在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,让新配置生效。

  4. 4
    验证是否可用

    连接成功后随便调一个只读工具,确认能返回结果,再接入正式流程。

典型使用场景

3 类
按分类与标签整理的常见用法,具体以官方文档为准
代码与仓库操作

在对话里直接查代码、提 issue、走 PR,省掉来回切工具。

本地命令与脚本

把构建、测试、格式化这类重复命令交给模型执行。

接口联调与测试

让模型调你的接口、跑一遍用例,快速定位问题。

选择建议

选之前先看这几条
按条目字段给出的判断,不替你决定用不用

值得选的地方

  • 已适配 Claude、Cursor、Gemini CLI,换客户端不用重新研究配置。

先想清楚的地方

  • 没有现成安装命令,需要按官方文档自己填启动命令和参数。
  • 使用前要准备好OAuth 授权对应的凭据,并按最小权限配置。
  • 资料没标注工具数量,实际能力要装好后确认。
  • 版本更新与行为细节以官方仓库为准。

更适合

  • 需要在开发工具场景里让模型直接调用外部能力的人。
  • 已经在用 Claude、Cursor 的用户,接入路径最短。
  • 想先快速试一下、能接受在本地或自有环境跑服务的人。

不太适合

  • 不方便管理密钥或第三方授权的人。
  • 需要厂商级 SLA 和技术支持的生产环境。

同类对比

6 个同类
同分类里挑最接近的几条,字段以各自详情页为准
MCP 服务器工具数传输方式部署方式安装命令最后核验
UniversalBench: run code, web, databases, and any LLM from your AI, safely当前 未标注 未标注 未标注 无 未标注
Git 838e2ec1 12 stdio · 本地进程 本地部署 有 2026-05-17
Markitdown MCP 2c653a5f 1 stdio · 本地进程 本地部署 有 2026-05-17
JavaGuide 未标注 stdio · 本地进程 本地部署 有 2026-04-11
awesome-claude-skills 未标注 stdio · 本地进程 本地部署 有 2026-04-26
Context7 831acf40 未标注 stdio · 本地进程 本地部署 有 2026-05-17

服务器档案

开发工具
字段来自项目仓库与公开索引,中文说明由本站标注
提供方nikhilgogulwar上游标注的维护方 nikhilgogulwar
分类开发工具本站按上游分类展示
连接方式未标注上游未给出传输方式
鉴权方式OAuth 授权远程服务建议用最小权限凭据
维护方社区 / 第三方核验等级 仅资料核对
包名 / 运行时未标注 · Cross-platform本地安装时用到的包标识与运行时
部署类型未标注上游给出的部署形态
最近核验未标注上游最后核验时间

详细介绍

7.8 KB
来自项目仓库与上游目录的原文,未做删改

UniversalBench

Your AI finally has hands.

Connect once. Your AI can now run code, push to GitHub, query databases,

deploy, call APIs, manage credentials, and invoke any AI model.

Not describe how. Actually do it.

universalbench.dev ·
Docs ·
Sign up free


The problem

Your AI is genuinely intelligent. It can reason, plan, write code, and solve complex problems.

But it cannot do anything in the real world on its own.

It can write a Python script. It cannot run it. It can suggest a Git commit. It cannot push it. It can write a database query. It cannot execute it. It can describe a deployment. It cannot ship it.

It knows how to do the work. It has no hands to do it with.


What UniversalBench does

UniversalBench gives your AI a cloud computer it can use as a tool.

Connect one URL to any MCP-compatible AI. From that moment your AI can:

  • Run the code it writes, execute Python and bash in an isolated sandbox, see real output, fix errors, iterate
  • Push to GitHub, read repos, commit files, validate code before it lands, roll back automatically if a smoke test fails
  • Deploy and verify, push a change, hit a health check, confirm the new version is serving, roll back if not
  • Query your databases, structured queries, raw SQL, keyword search, inserts, upserts, against any database you connect
  • Call any API, HTTP requests to any public endpoint with network safety built in
  • Invoke any AI model, route to Claude, GPT-4o, Gemini, or any other model, with a cost ceiling enforced before the call runs
  • Store credentials once, save a GitHub token or database URL to the vault, and every relevant tool reads it automatically
  • Run tasks in parallel, up to 8 concurrent execution threads for workflows that need speed
  • Snapshot any web page, capture a full-page screenshot of any URL and get back a link that opens it, to show a result or confirm a page rendered correctly
  • Take files in and hand them back, receive a file from someone through an upload link, work on it, and return a finished file as a clean branded download, with the file never passing through the chat

None of this is described. None of this is suggested. It is executed, confirmed, and returned.


Connect in 30 seconds

1. Sign up free at universalbench.dev. No credit card. 1,000 free calls per month from day one.

2. Copy your API key from the dashboard.

3. Paste this URL into any AI client:```
https://universalbench-mcp.penantiaglobal.workers.dev/u/YOUR_API_KEY


### Claude.ai
Settings → Integrations → Add custom integration → paste the URL.

### Cursor / Windsurf / VS Code Copilot```json
{
  "mcpServers": {
    "universalbench": {
      "url": "https://universalbench-mcp.penantiaglobal.workers.dev/u/YOUR_API_KEY"
    }
  }
}

Claude Desktop```json

{
"mcpServers": {
"universalbench": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/client-sse",
"https://universalbench-mcp.penantiaglobal.workers.dev/u/YOUR_API_KEY"]
}
}
}
```

Works with any MCP-compatible client: Claude, ChatGPT, Gemini, GitHub Copilot, Cursor, Windsurf, and more.


What this looks like in practice

Your AI ships code, not just writes it

Without UniversalBench, your AI writes the function, you copy it, run it yourself, see an error, paste it back. Back and forth until something works.

With UniversalBench, your AI writes the function, runs it, sees the error, fixes it, runs it again, confirms it works, pushes the commit, hits the health check, and tells you it is live. You asked once. It shipped.

Your AI works your data, not just talks about it

You have a log file with 50,000 lines. Without a runtime, your AI reads as much as fits in context and estimates the rest.

With UniversalBench, your AI writes a script, runs it against the full file in the sandbox, and returns the exact answer. Every row processed. One round trip.

Your AI manages your infrastructure

> "Check the last 100 database rows for anomalies, open a GitHub issue if you find any, and send me a summary."

One prompt. Your AI queries the database, finds the anomaly, writes the GitHub issue, opens it, and responds with what it found. No copy-pasting. No manual steps. No back and forth.

Your AI coordinates other AIs

> "Run this summarisation task through three models and tell me which gives the best result per dollar."

Your AI calls each model, compares outputs, calculates cost per call, and reports back. UniversalBench enforces your spending ceiling before each call goes out.

Your AI hands you finished files, and shows you results

Drop a messy spreadsheet into a link your AI gives you. It cleans it in the sandbox and hands back a polished file you can download. Ask it to check a page you just shipped, and it captures a full snapshot and returns a link that opens it. The files never pass through the chat, so there is no size limit and nothing to copy and paste.


Three safety guarantees built in

These are not settings you configure. They run on every call, for every customer, by default.

Your AI cannot push code that breaks your service.
Every push is validated before the commit lands. The previous version is captured automatically. Provide a health check URL and we run a smoke test after deploy. If it fails, the previous version is restored.

Your AI cannot burn your API budget.
Every AI model call is cost-estimated before it runs. Calls over your ceiling are rejected. You set the limit. Default is $0.50 per call. Hard cap is $50.

Your AI cannot reach your internal network.
Every outbound HTTP request is checked against a blocklist before it leaves. Private IP ranges, loopback, link-local, and cloud metadata endpoints are blocked at the infrastructure level.


What your AI can do

One connection. Three tools (ub_read, ub_write, ub_ai). Underneath them, your AI can:

  • Run code and shell, execute Python and shell commands in an isolated sandbox, with state and files that persist across a session
  • Install what it needs, pull in any package before running
  • Work in parallel, run multiple tasks at once for speed
  • Search the live web, real results with sources
  • Call any public API, with internal-network access blocked by default
  • Invoke any AI model, Claude, GPT, Gemini, and more, with a cost ceiling enforced before each call
  • Read and ship code on GitHub, read repos, commit with validation, and roll back automatically if a deploy check fails
  • Query your databases, structured queries, raw SQL, search, inserts and updates, against any database you connect, using your own credentials
  • Capture and preview pages, take a full-page snapshot of any URL, or show a page your AI built, as a shareable link
  • Take files in and hand them back, receive a file through an upload link, work on it, and return a finished, correctly-typed download, with files never passing through the chat
  • Persist memory and state, keep context across sessions
  • Store credentials safely, save a credential once in an encrypted vault and the right tools use it automatically, your AI never sees the value

Every capability runs behind the same safety checks: validated before code lands, cost-estimated before a model runs, and network-checked before a request leaves.


Credentials, saved once

Store a credential once and your AI uses it on every future call, without ever seeing the value. Encrypted at rest. Save once, and the relevant tools start working automatically.


Pricing

Free · 1,000 calls per month, no credit card
Pay as you go · $0.008 per call after the free tier
Minimum top-up · $5

No subscription. No lock-in. Credits do not expire.


Documentation

Full capability reference, quickstart guides, and agent-ready prompts at docs.universalbench.dev.


Sign up free, no credit card →

常见问题

7 问
答案由本站依据条目字段整理,不引用上游原文
UniversalBench: run code, web, databases, and any LLM from your AI, safely 怎么安装?

资料没有给出安装命令,请到官方仓库或文档按说明接入;这类条目多数是生态应用,或尚未发布可安装包。

UniversalBench: run code, web, databases, and any LLM from your AI, safely 怎么在 Claude、Cursor、Gemini CLI 里配置?

它的传输方式未标注,在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,配置才会生效。

UniversalBench: run code, web, databases, and any LLM from your AI, safely 收费吗?

本站不代收费用,也没有标注具体价格。是否收费取决于服务本身和背后的接口额度,正式使用前按官方说明确认。

UniversalBench: run code, web, databases, and any LLM from your AI, safely 需要 API Key 吗?

资料标注需要OAuth 授权认证,使用前准备好对应凭据,并按最小权限配置。

UniversalBench: run code, web, databases, and any LLM from your AI, safely 支持哪些客户端?

资料标注支持 Claude、Cursor、Gemini CLI、Codex、Zed;其他客户端只要能配置 MCP Server,也能用同样的命令或地址接入。

UniversalBench: run code, web, databases, and any LLM from your AI, safely 能提供哪些能力?

资料显示能力标签包括 mcp、code execution、ai agents、llm、database、github。具体工具清单以官方文档为准。

UniversalBench: run code, web, databases, and any LLM from your AI, safely 需要什么运行环境?

运行时为 Cross-platform。

热门专题

按标签挑服务器
全部专题 →

服务器信息以作者仓库与官方文档为准;分类、传输方式与鉴权方式的中文说明为中文整理,最近核验 未标注。