Boxline MCPサーバー

Boxline MCP server

Give your AI agents the infrastructure they need: browsers, shells, storage and isolated machines.

This server gives any MCP client (Claude, Cursor, VS Code and others) an isolated cloud machine with a real browser, a shell and a disk, through Boxline. The agent can open pages, click and type, read a page as Markdown, run commands, read and write files, fetch pages and search the web.

Set it up

You need a Boxline API key (console → API keys) and Node 20 or newer.

Claude Desktop, Cursor and other clients that take a JSON config:

{
  "mcpServers": {
    "boxline": {
      "command": "npx",
      "args": ["-y", "@boxline/mcp"],
      "env": { "BOXLINE_API_KEY": "bxl_your_key" }
    }
  }
}

Claude Code:

claude mcp add boxline --env BOXLINE_API_KEY=bxl_your_key -- npx -y @boxline/mcp

BOXLINE_API_URL points it at another API (default https://api.boxline.dev).

Tools

The server keeps nothing between calls: session_create returns a sessionId, and every tool that works in a session takes it as a required sessionId (session_list finds an id again; a tool that acts resumes a stopped session first, as it was; browser_read and browser_screenshot say to call session_resume instead, because a resume starts a machine and billing; files_list and files_read read a stopped session's saved files without resuming it). session_create, session_list, credentials_list and the web_* tools need no session.

A session is a browser, a shell or both: session_create takes browser and shell (both true unless you say otherwise). browser: false with a shell is a shell-only session, with no live view: the browser_* tools answer browser_disabled.

Every tool says what it does to the world in its MCP annotations (a title, and read-only, destructive and open-world hints), so a client can decide what to ask you about. There are 26 tools:

Group	Tools
Sessions	session_create (a browser, a shell or both; optionally from a saved browser profile, and with credentials exported into the shell), session_list (find a session again), session_stop (saves the session as it is and stops billing), session_resume (brings a stopped session back as it was), session_delete (deletes it and what it saved, for good)
Browser	browser_navigate, browser_act (one or more steps in plain English or as exact actions: hover, scroll, drag, right and double clicks…), browser_click (a selector or x/y), browser_type (text, or a saved credential), browser_press (keys: Enter, Control+A, ctrl+a Delete), browser_read (Markdown or text), browser_screenshot, browser_login (sign in with a saved password), browser_computer (screenshot-driven computer use)
Shell	shell_exec (bash in the session)
Files	files_list, files_read, files_write, files_delete (the session's workspace, its disk; files_delete also removes a folder with everything in it)
Web	web_fetch (a page through a real browser), web_search (with the top pages as Markdown), web_screenshot (a quick screenshot), web_extract (JSON from up to 10 pages, from a prompt or a JSON Schema), web_crawl_start and web_crawl_get (a whole site or section, robots.txt respected); none needs a session
Credentials	credentials_list (names, types, sites and a password's code source for the saved passwords and secrets, never their values)

browser_act is the tool for everything the others do not cover. A plain-English step ("click the Sign in link") is carried out by a model, like a step of an agent run: it is billed as one and needs a plan that includes plain-English steps. An exact action ({"action": "hover", "selector": "text=Products"}) uses no model. Each step gets one line saying what happened; an evaluate step also gives the value it returned ({"action": "evaluate", "expression": "document.title"}). There are no agent-run tools: an MCP client is already an agent.

Signing in without showing the agent a password

Save a website password (or a secret) as a credential in the Boxline console, the CLI or an SDK. The agent calls credentials_list to see what exists (names, types and the sites each may be typed on; never a value), then types it with browser_type:

browser_type { "sessionId": "<id>", "credential": "SHOP", "field": "username", "selector": "#email" }
browser_type { "sessionId": "<id>", "credential": "SHOP", "field": "password", "selector": "#password" }
browser_type { "sessionId": "<id>", "credential": "SHOP", "field": "otp", "selector": "#code" }   // the current 2FA code, if the password has a 2FA key

Or sign in in one call with browser_login { "sessionId": "<id>", "credential": "SHOP", "url": "https://shop.example.com/login" }: a short run in the session types the credential on its sites only, and the agent sees neither the password nor any code. A password whose codeSource (in credentials_list) is push or url waits for the code or sign-in link that your system sends, up to its timeout (the agent should tell you a code is needed); you send it with the CLI (boxline credentials push-code SHOP) or an SDK (credentials.pushCode), never through the chat.

The platform types the value into the field itself, only on the sites the credential was saved for, and never shows it to the agent, the chat or the logs: wherever a result would show it (browser_read, the page's elements, a screenshot, an error), the text is replaced by %SHOP.password% and the field is covered. This hides a value; it does not protect it. It is a guard against showing a credential by accident, not a boundary: code that runs in the page or the session (an evaluate action in browser_act, a command in shell_exec) can still read a field and print it, for example reversed or in base64. Use a password only in sessions whose pages and scripts you trust, and save it with the narrowest sites. session_create can also start from a saved browser profile (cookies and logins from an earlier sign-in; persistProfile saves new ones back) and export credentials into the session's shell as environment variables, for those credentials_list marks exportable (scope shell or all, or shell: true); others are refused. With a shell, the password a profile signs in with is exported too when it is exportable. No tool creates or changes a credential, so a password never passes through the chat: add them yourself, outside the conversation.

Sessions are billed while they run: stop them with session_stop when the work is done, or let them stop at their time limit. A stopped session is free and is kept for your plan's retention days (7 on Free, 30 on Hobby and Startup, 90 on Scale): session_resume brings it back with its tabs and files, and session_delete removes it at once, with its recording and logs. A tool that acts, given the id of a stopped session, resumes it first; browser_read and browser_screenshot say to call session_resume; files_list and files_read read the saved files as they are.

Self-hosting over HTTP

boxline-mcp --http serves the same tools over HTTP at POST /mcp, for clients that connect to a URL instead of starting a process. It speaks the MCP revision 2026-07-28 and, without sessions, the 2025-era protocol older clients still use.

MCP_ALLOWED_HOSTS=mcp.example.com BOXLINE_API_URL=https://api.boxline.dev npx -y @boxline/mcp --http

• Sign-in: every request carries its own Boxline API key as Authorization: Bearer <API key>, or the OAuth access token an MCP client got when a person clicked "Connect" (below). The server holds no key of its own, never logs a token, and calls the API with the token of the request it is serving. Before a request is served the token is checked against the API (one GET /v1/auth/me, remembered for a minute by its hash): a token the API refuses (expired, revoked, unknown) is 401, so a client signs in again or refreshes it instead of getting a tool error. Without a token the answer is 401 with WWW-Authenticate: Bearer.
• OAuth sign-in (for ChatGPT, Claude and other clients that offer "Connect" instead of a key field) is for Boxline's hosted server, https://mcp.boxline.dev/mcp. Boxline's sign-in issues tokens for that address only. A server you host yourself takes API keys (Authorization: Bearer bxl_…). The hosted server sets MCP_RESOURCE (its public URL) and MCP_AUTH_SERVER (the issuer, https://api.boxline.dev), then publishes protected resource metadata (/.well-known/oauth-protected-resource, RFC 9728) and answers a missing or refused token with 401 and a WWW-Authenticate: Bearer resource_metadata="…" challenge, which is what those clients follow to sign the person in. An OAuth token (bxo_…) always gets the directory tools (no saved credentials); a bxl_ API key gets MCP_TOOLS.
• Stateless: each request is served on its own, so any number of copies behind a load balancer can answer any request, and a restart loses nothing. Nothing is remembered between calls: the sessionId that session_create returns is the only handle (as above). GET /mcp and DELETE /mcp answer 405.
• Put it behind HTTPS (a reverse proxy or load balancer). GET /healthz answers 200 ok for health checks, on any host.
• Environment:

Variable	Default
MCP_PORT	8081	Port to listen on (0 picks a free one).
MCP_HOST	0.0.0.0	Address to bind.
BOXLINE_API_URL	https://api.boxline.dev	The Boxline API the server calls.
MCP_ALLOWED_HOSTS	none	Comma-separated host names the server answers to (mcp.example.com), against DNS rebinding. localhost, 127.0.0.1 and [::1] always pass; * turns the host and origin checks off. A request whose Host is not listed gets 403, and so does a browser Origin that is not one of them.
MCP_RESOURCE, MCP_AUTH_SERVER	none	Both or neither: turn on OAuth sign-in discovery (above; Boxline's hosted server only).
MCP_OPENAI_CHALLENGE	none	The token ChatGPT's domain verification expects at /.well-known/openai-apps-challenge.
MCP_TOKEN_CACHE_SECONDS	60	How long a token's check against the API is remembered (0: every request).
MCP_TOOLS	full	directory leaves out browser_login, credentials_list and every credential parameter (credential and field on browser_type, credentials on session_create), and browser_act refuses an action that names a credential or logs in: for a server listed in a public connector directory.

Request bodies are limited to 4 MiB (413), at most 32 requests run at once (MCP_MAX_IN_FLIGHT; more get 503), and at most 8 per bearer token (MCP_MAX_IN_FLIGHT_PER_TOKEN; more get 503, so one caller cannot hold every slot). Live view URLs are left out of tool results. Long tool calls keep their connection alive with SSE comment frames every 15 seconds.

Links

• Docs: docs.boxline.dev
• Node SDK: Boxline-dev/sdk-node · Python SDK: Boxline-dev/sdk-python

MIT licence.

---

This repository holds the Boxline MCP server (@boxline/mcp). It is copied from Boxline's main repository on every change. Issues and pull requests are welcome here; accepted changes are made there and arrive with the next copy.