Multilogin Browser MCP
Enables AI agents to control Multilogin X anti-detect browser profiles via a streamable HTTP MCP interface. Supports profile management, session handling, and a broad catalog of browser actions such as navigation, clicking, typing, screenshots, and more.
README
Multilogin Browser MCP
A FastAPI service that drives Multilogin X anti-detect browser profiles over CDP and exposes them to AI agents through a streamable HTTP MCP server.
Features
- Profile management — list, search, create, and remove Multilogin profiles (batch create is supported without an artificial cap).
- Browser sessions — start a profile, connect Playwright over CDP, and manage the session lifecycle.
- Browser action catalog — navigate, click, type, evaluate JS, screenshots, tabs, cookies, scroll, hover, and more.
- Dual interface — REST API for integrations and MCP for agent frameworks.
- Session watchdog — 15-minute max TTL and idle timeout cleanup.
- Railway-ready —
PORTand.envdriven.
Requirements
- Python 3.12+
- Multilogin desktop app / agent running locally (exposes the launcher at
https://launcher.mlx.yt:45001, resolving to127.0.0.1) - A Multilogin automation token, or username + password
Installation
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
No playwright install is required — we connect over CDP to the browser Multilogin launches.
Configuration
Copy the example and fill in the values:
cp .env.example .env
# Option A: automation token (recommended for servers)
MULTILOGIN_AUTOMATION_TOKEN=your_token
# Option B: username + password
# MULTILOGIN_USERNAME=you@example.com
# MULTILOGIN_PASSWORD=your_password
# Optional overrides
# MULTILOGIN_API_URL=https://api.multilogin.com
# MULTILOGIN_LAUNCHER_URL=https://launcher.mlx.yt:45001
# PORT=8000
# SESSION_MAX_TTL_SECONDS=900
# SESSION_IDLE_TIMEOUT_SECONDS=300
# MULTILOGIN_DEBUG_HTTP=1 # log every outgoing Multilogin request (auth masked)
When MULTILOGIN_AUTOMATION_TOKEN is set it is always used and
MULTILOGIN_USERNAME / MULTILOGIN_PASSWORD are ignored — the two auth modes
are never combined in one request. Set only one of them to avoid confusion.
Run
python main.py
# or
uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload
- REST docs:
http://localhost:8000/docs - MCP endpoint:
http://localhost:8000/mcp(both/mcpand/mcp/are served directly, no redirect)
REST API
| Method | Path | Purpose |
|---|---|---|
| GET | /api/v1/profiles/ |
List/search profiles |
| POST | /api/v1/profiles/search |
Search profiles |
| POST | /api/v1/profiles/ |
Create a profile |
| POST | /api/v1/profiles/bulk |
Create many profiles |
| POST | /api/v1/profiles/remove |
Remove profiles |
| GET | /api/v1/profiles/folders |
List workspace folders |
| POST | /api/v1/profiles/folders |
Create a workspace folder |
| GET | /api/v1/profiles/statuses |
Running profiles (launcher) |
| POST | /api/v1/sessions/ |
Launch a session (start profile + connect) |
| GET | /api/v1/sessions/ |
List active sessions |
| GET | /api/v1/sessions/{id} |
Session status/expiry |
| DELETE | /api/v1/sessions/{id} |
Close session + stop profile |
| GET | /api/v1/browser/actions |
List available browser actions |
| POST | /api/v1/browser/sessions/{id}/action |
Run an action on a session |
| GET | /api/v1/diagnostics |
Effective auth mode, egress IP, and live probes of the Multilogin API |
Example flow
# 1. list profiles
curl http://localhost:8000/api/v1/profiles/
# 1b. create a profile (folder_id defaults to the first workspace folder,
# browser_type/os_type default to mimic/windows)
curl -X POST http://localhost:8000/api/v1/profiles/ \
-H 'content-type: application/json' \
-d '{"name":"agent-1"}'
# 2. launch a session
curl -X POST http://localhost:8000/api/v1/sessions/ \
-H 'content-type: application/json' \
-d '{"profile_id":"<pid>","folder_id":"<fid>"}'
# 3. drive it
curl -X POST http://localhost:8000/api/v1/browser/sessions/<sid>/action \
-H 'content-type: application/json' \
-d '{"action":"navigate","params":{"url":"https://example.com"}}'
# 4. close (auto-closes after 15 min anyway)
curl -X DELETE http://localhost:8000/api/v1/sessions/<sid>
MCP Tools
Point any streamable-HTTP MCP client at http://localhost:8000/mcp/.
Discovery
list_profilessearch_profileslist_folderscreate_folder
Profile management
create_profile— create one profilecreate_profiles— create any number of profiles in one call (no artificial limit)remove_profiles
Sessions
launch_sessionsession_statuslist_sessionsclose_session
Browser
browser_action— run any action by namelist_browser_actions— see the full catalog
The full action catalog includes: navigate, click, type_text, press_key, get_text, get_html, get_url, screenshot, wait_for, evaluate_js, list_tabs, switch_tab, new_tab, close_tab, go_back, go_forward, reload, scroll, hover, select_option, set_checked, upload_file, cookies.
Example client config
{
"mcpServers": {
"multilogin-browser": {
"url": "http://localhost:8000/mcp/"
}
}
}
Troubleshooting
405 Method Not Allowed when creating a profile — you are calling an
endpoint that does not accept that method. Profile creation is
POST /api/v1/profiles/ (REST) or the create_profile / create_profiles
MCP tools. Note that the MCP endpoint itself only speaks POST (plus GET/DELETE
for SSE and session teardown) on /mcp.
503 with "Cannot reach Multilogin" — the local launcher
(https://launcher.mlx.yt:45001) or api.multilogin.com is unreachable.
Start the Multilogin X desktop app / agent, or fix
MULTILOGIN_LAUNCHER_URL / MULTILOGIN_API_URL. Only launcher calls
(/statuses, session launch/stop) need the desktop app; profile CRUD goes to
the cloud API and works headlessly with a token.
502 with "answered 501 with an HTML page" — the POST never reached the
Multilogin API. api.multilogin.com sits behind an edge that answers POSTs
from blocked networks (cloud/datacenter IPs, e.g. Railway) with a canned
501 Unsupported method ('POST') HTML page, while GETs from the same host and
token reach the real API (they return proper JSON such as
401 Wrong JWT token). It is therefore not a credential or request-format
problem. Confirm with GET /api/v1/diagnostics, which probes both methods and
reports the egress IP. Fixes: run this service from a network Multilogin
accepts (the machine running Multilogin X, which the launcher/CDP part needs
anyway), or ask Multilogin support to allowlist your egress IP.
MCP requests rejected with a security error — set MCP_ALLOWED_HOSTS /
MCP_ALLOWED_ORIGINS to your public host when running behind a proxy. When
both are empty, DNS-rebinding protection is disabled and all hosts are
accepted.
Deployment
Colocated with Multilogin X (recommended)
Run the server on the machine that has the Multilogin X desktop app / agent installed, and expose it with a tunnel. This is the only layout where both halves work:
- The launcher binds to
127.0.0.1(launcher.mlx.ytresolves to127.0.0.1), soGET /api/v2/profile/f/{folder}/p/{profile}/startreturns a port whose CDP endpoint isws://127.0.0.1:<port>/devtools/.... There is no cloud CDP URL — only a process on the same machine can attach. api.multilogin.comblocks profile/workspace POSTs from cloud/datacenter networks (see Troubleshooting), so cloud hosts can't do profile CRUD either.
cp .env.example .env # set MULTILOGIN_AUTOMATION_TOKEN
./scripts/serve-with-tunnel.sh
That starts the server on 127.0.0.1:8000 and a Cloudflare quick tunnel,
printing a public https://<random>.trycloudflare.com URL. For a URL that
survives restarts, create a named tunnel once and pass it in:
cloudflared tunnel login
cloudflared tunnel create mlx-mcp
cloudflared tunnel route dns mlx-mcp mlx-mcp.yourdomain.com
TUNNEL_NAME=mlx-mcp ./scripts/serve-with-tunnel.sh
Point the agent at https://<your-tunnel-host>/mcp. With a stable hostname,
set MCP_ALLOWED_HOSTS=<your-tunnel-host> in .env so DNS-rebinding
protection is on and scoped to it.
Railway
- Push the repo.
- Set the environment variables from
.envin Railway. - Set the start command to
python main.py(oruvicorn app.main:app --host 0.0.0.0 --port $PORT).
Note on the launcher
The Multilogin desktop launcher currently resolves to 127.0.0.1. For remote hosts (e.g. Railway), the launcher must either run on the same host or be reachable through a tunnel / VPN.
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。