Browser Use
Browser agents and stealth cloud browsers for AI: run web tasks, fill forms, extract data.
作者提供的说明(英文原文)
Drive the browser of ZennoPoster ProjectMaker: open pages, read and click elements, screenshots.
这三项永远不会合并成一个分数:意见、测量和人气是不同的东西。
我们还没有为它写通俗的介绍。下面是我们从注册中心读取到的事实。
选择你使用的 AI 应用,然后复制设置。
你用的是哪个应用?
在上方选择你的应用。我们会记住它,下次按钮会显示“安装到 …”。
链接指向作者自己的文件;我们不托管也不重新打包它。打开文件后,Claude Desktop 会进行安装。
在终端中运行这条命令
浏览器会先询问是否打开应用,然后应用会显示设置并请你确认。密钥的值会留空,由你自己填写。
浏览器会先询问是否打开应用,然后应用会显示设置并请你确认。密钥的值会留空,由你自己填写。
这个服务器在你的电脑上运行,而 ChatGPT 只能连接互联网上的服务器。请选择其他应用。
你的选择通过 Cookie 保存在这台设备上,无需账户。
这些设置是根据作者在注册中心的条目生成的。如果 README 的写法不同,请以 README 为准。
它在你的电脑上运行;安装后,工具列表会显示在你的 AI 应用中。
我们检查的是任何人都能看到的信息;我们没有阅读代码。不适用的检查项不计入总数。 · 检查于 2026年11月6日
注册名称已通过域名 zennolab.com 验证(注册中心会核实域名所有权)
我们还没有从 GitHub 读取这个仓库
我们还没有从 GitHub 读取这个仓库
它在你的电脑上运行,没有互联网端点
在你的电脑上运行,改用本地设置
它在你的电脑上运行;我们只能读取互联网端点的工具列表
与注册中心里这个分类中所有活跃的服务器比较。只有当发布者拥有该仓库时,星标数才计入。
它在 1 项适用的基础检查中通过了 1 项。我们没有阅读代码。在添加密钥或令牌之前,请先看看这个服务器要求哪些权限,并使用只具备所需权限的密钥。
注册中心没有列出必需的密钥,不过有些服务器会在首次使用时要求你登录。
它在你的电脑上运行。你的 AI 应用会在需要时启动它,数据不会经过作者的服务器(除非它自己调用在线服务)。
GitHub 仓库最后更新于–,最新版本(0.4.0)于16天前进入注册中心。
支持 MCP 的应用,例如 Claude Desktop、Claude Code、Cursor 和 VS Code。你的电脑上还需要有 MCPB。
同一分类中有:Browser Use, Chrome DevTools MCP, MCP Server Browser
同一分类中其他作者的服务器,按星标数排序。
如何看卡片:★ GitHub 星标i三大支柱:评价 · 检查 · 使用情况i更新 = 最新版本发布到注册中心的时间i
Browser agents and stealth cloud browsers for AI: run web tasks, fill forms, extract data.
MCP server for Chrome DevTools
MCP server for browser use access
按原文从作者的仓库复制。PopNx MCP 没有审查或背书其内容。HTML、图片和脚本已被移除。链接会离开本站。
English | Русский
The MCP servers are published as self-contained win-x64 binaries via GitHub Releases in the shared repository https://github.com/ZennoLab/zennoposter-mcp/releases. Each server has its own release line (its own tag prefix):
mcp-projectmaker-v*, archive MCP.ProjectMaker-v*-win-x64.zipmcp-instance-v*, archive MCP.Instance-v*-win-x64.zipmcp-zennoposter-v*, archive MCP.ZennoPoster-v*-win-x64.zipmcp-android-v*, archive MCP.Android-v*-win-x64.zipPublicApi documentation (the OpenAPI contract rendered with Redoc, the integrator guide, error codes, versioning policy) lives at https://zennolab.github.io/zennoposter-mcp/ (the pages are published from the docs/ folder of this repository).
These servers talk to ZennoPoster 7.9.3 and newer and to ZennoDroid 2.6.1 and newer. MCP.Instance applies to ZennoPoster only and MCP.Android to ZennoDroid only; MCP.ProjectMaker and MCP.ZennoPoster apply to both. Which server version goes with which product and contract version: https://zennolab.github.io/zennoposter-mcp/compatibility.html.
Install buttons for Cursor and VS Code and the Claude Code commands for every server: https://zennolab.github.io/zennoposter-mcp/install.html.
The product itself starts internal MCP servers for its built-in AI chat — they live on the internal port band 6107–6113 (on ZennoDroid everything is shifted by +10), receive a least-privilege service key from the host, and ignore the Authorization header of incoming requests. This is internal infrastructure: connecting to it from outside is not supported (permissions there are defined by the service key, not yours), and its ports must not be occupied — a foreign process on a port from this band prevents the built-in server from starting (the product logs an error, but its AI stack is left without that server).
Those internal servers have no console of their own, so the product copies everything they write to their output into its own log, each line prefixed with the process name (MCP.ProjectMaker, Orchestrator, …). Ordinary start-up chatter is tagged Debug and never reaches the log at all; only what the server itself reports as a warning or a failure, and the exit code of a server that dies unprompted, actually get written. That is the place to look when the built-in chat loses one of its servers.
Those internal servers also expect a per-machine secret that the product hands only to its own AI processes, so being able to reach the port is no longer enough to use one: a call without it is refused with `401`, and the refusal is written to the product's log as ingress refused. The same goes for the built-in chat's own API on 6113. If you were calling those ports from your own client, run your own copy on the public band instead, as described below.
For your own LLM client you run a separate copy of the MCP server from the public package: it listens on a public port (the tables below) and talks to the same product PublicApi, but with your ApiKey — with the scopes/tier you chose when issuing the key. Both copies run side by side without interfering with each other.
On ZennoDroid the two servers that serve both products listen +10 higher — MCP.ProjectMaker on 6217, MCP.ZennoPoster on 6220 — the same shift the product applies to its internal band. That way a machine with both products installed can run both sets at once. MCP.Android exists only on ZennoDroid and keeps 6211: nothing on ZennoPoster occupies it.
Default ports, one table per product. A server takes its listen port and BaseUrl from the appsettings.json next to its exe; the ports marked ¹ have no such defaults and are passed on the command line together with the matching BaseUrl (and Target), as in the examples of step 3.
ZennoPoster
| Server | Port | Talks to (PublicApi) | Key in config |
|---|---|---|---|
MCP.ProjectMaker | 6207 | ProjectMaker :5299 | ProjectMaker:ApiKey |
MCP.Instance, Target=projectmaker (editor browser) | 6208 | ProjectMaker :5299 | Instance:ApiKey |
MCP.Instance, Target=zennoposter (task browsers) | 6209 ¹ | ZennoPoster :5300 | Instance:ApiKey |
MCP.ZennoPoster | 6210 | ZennoPoster :5300 | ZennoPoster:ApiKey |
ZennoDroid
| Server | Port | Talks to (PublicApi) | Key in config |
|---|---|---|---|
MCP.ProjectMaker | 6217 ¹ | ProjectMaker :5309 | ProjectMaker:ApiKey |
MCP.ZennoPoster | 6220 ¹ | ZennoPoster :5310 | ZennoPoster:ApiKey |
MCP.Android, Target=projectmaker (editor device) | 6211 | ProjectMaker :5309 | Android:ApiKey |
MCP.Android, Target=zennoposter (task devices) | 6212 ¹ | ZennoPoster :5310 | Android:ApiKey |
¹ Not a built-in default: set it explicitly (--urls plus the matching BaseUrl; the commands are in step 3).
Everything can be overridden through standard ASP.NET Core configuration: the appsettings.json next to the exe, environment variables (ASPNETCORE_URLS, ProjectMaker__ApiKey, …) or command-line arguments (--urls, --ProjectMaker:ApiKey=…, …) — arguments override environment variables, environment variables override appsettings.json.
On the releases page find the latest release of the server you need (by tag prefix), download its *-win-x64.zip and unpack it into any folder. Each archive is a single self-contained .exe + appsettings.json; no additional .NET runtime is required. For Claude Desktop take the .mcpb bundle of the same release instead, see Claude Desktop: install from an .mcpb bundle.
In ProjectMaker: Settings → Api-Keys → Add new API key — a dialog of the same name opens:
Label (an arbitrary key name, to tell keys apart in the list).Max tier (the minimum sufficient for your tasks — T0 for read-only, higher for mutating operations).Scopes you need (only *:read are ticked by default; add others as needed).More on scopes/tiers — security-model.md.
The key is set in the MCP server's own configuration (not in the MCP client's headers — the server accepts connections from loopback only and does not read the Authorization header of incoming requests). A server whose port in the tables above has no ¹ can take the key from ApiKey in the appsettings.json next to its exe and start with no arguments; the ports marked ¹ are always set on the command line, together with the matching BaseUrl (and Target). The key can also be passed as an environment variable or an argument, as below. Run only the servers you need, from the block of your product.
ZennoPoster
# ProjectMaker (editor): 6207 -> :5299
.\ZennoLab.AI.MCP.ProjectMaker.exe --ProjectMaker:ApiKey=zp_xxx
.\ZennoLab.AI.MCP.Instance.exe --Instance:ApiKey=zp_xxx
.\ZennoLab.AI.MCP.Instance.exe --urls http://localhost:6209 --Instance:Target=zennoposter --Instance:BaseUrl=http://localhost:5300/api/v1 --Instance:ApiKey=zp_xxx
.\ZennoLab.AI.MCP.ZennoPoster.exe --ZennoPoster:ApiKey=zp_xxx
ZennoDroid
# ProjectMaker (editor): 6217 -> :5309
.\ZennoLab.AI.MCP.ProjectMaker.exe --urls http://localhost:6217 ` --ProjectMaker:BaseUrl=http://localhost:5309/api/v1 --ProjectMaker:ApiKey=zp_xxx
.\ZennoLab.AI.MCP.ZennoPoster.exe --urls http://localhost:6220 ` --ZennoPoster:BaseUrl=http://localhost:5310/api/v1 --ZennoPoster:ApiKey=zp_xxx
.\ZennoLab.AI.MCP.Android.exe --Android:ApiKey=zp_xxx
.\ZennoLab.AI.MCP.Android.exe --urls http://localhost:6212 --Android:Target=zennoposter --Android:BaseUrl=http://localhost:5310/api/v1 --Android:ApiKey=zp_xxx
MCP.Android mounts twice for the same reason MCP.Instance does on ZennoPoster: the Android domain is served both by ProjectMaker's PublicApi (the device you see in the editor) and by the runner's (the devices its tasks are driving). Like MCP.Instance it takes a Target — projectmaker for the editor's device, zennoposter for the runner's — which selects the instructions the server hands to the AI at initialize (one device whose id may be left at 0, versus one device per worker thread with the id taken from list_devices); the tool set is the same either way. Target and BaseUrl are set together as a pair, exactly as for MCP.Instance (see the note below). Target appeared in MCP.Android 0.4.0; earlier versions differ only by --urls and BaseUrl.
Important note on the two copies of `MCP.Instance` and of `MCP.Android`: for these servers, Target (which instructions they serve to the AI — about the ProjectMaker editor or about the runner) and BaseUrl (where HTTP requests actually go) are configured only together, as a pair, and are not linked in code (Target=projectmaker → BaseUrl at the PM PublicApi :5299, Target=zennoposter → BaseUrl at the ZP PublicApi :5300; on ZennoDroid :5309 and :5310). On startup the server makes a best-effort check via the target host's /capabilities and logs a warning on mismatch, but if the target host is unreachable at startup the check is silently skipped — a mismatch is then not detected, and the AI gets instructions about one host while requests go to another.
Besides listening on a port, every server can speak MCP over its own stdin and stdout. Start the exe with --stdio and the client that launched it owns the connection: there is no port to pick, nothing else on the machine can reach the server, and several clients can each run their own copy. This is the mode MCPB bundles and Claude Desktop use; Claude Code and MCP Inspector support it as well. The mode is available from server versions 0.5.0 (ProjectMaker, ZennoPoster, Android) and 0.3.0 (Instance).
Only JSON-RPC goes to stdout; every log line goes to stderr. The key is read from the configuration as usual, which for a subprocess means the environment variable __ApiKey (ProjectMaker__ApiKey, Instance__ApiKey, ZennoPoster__ApiKey, Android__ApiKey); the -- :ApiKey= argument and appsettings.json next to the exe work too. The Instance and Android copies take their Target and BaseUrl pair the same way (Instance__Target, Instance__BaseUrl, and so on).
# Claude Code: -e stores the value in the server's own config entry, not in the launch command
$env:ProjectMakerApiKey = " " claude mcp add projectmaker --transport stdio -e "ProjectMakerApiKey=$env:ProjectMaker__ApiKey" -- "C:\mcp\ZennoLab.AI.MCP.ProjectMaker.exe" --stdio
The same entries for a client that reads a mcpServers configuration (Claude Desktop, Cursor
只显示了开头部分;其余内容请在仓库中阅读