Browsa

Trình duyệt đám mây cho tác nhân AI: thực hiện mọi tác vụ web bằng tiếng Anh thông thường trên Phantom, trình duyệt riêng của Browsa, với liên kết xem trực tiếp, lịch chạy, webhook có chữ ký và thông tin đăng nhập đã lưu.

Tài liệu

Browsa

@browsa/mcp

Drive real browsers from Claude Desktop, Cursor, Cline, and any MCP-compatible client.

MCP (Model Context Protocol) server for Browsa: turn any LLM into a real-browser agent with one config line. The agent gets a fresh cloud Chromium on Browsa's own build, and a live view link so you can watch it work.

npm version

What this is

Most "AI agent" setups stop at the LLM. We give the LLM a real browser. Through MCP, Claude can now:

  • Navigate, fill forms, click, scroll, scrape
  • Return results as structured JSON
  • Use your own HTTP or SOCKS5 proxy per browser, or a tier of Browsa's built-in network where it is on (see proxy_network)
  • Stay logged in across calls via persistent identities

No code from you: just configure Claude Desktop (or Cursor / Cline / Zed) once, and ask in natural language.

Install

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "browsa": {
      "command": "npx",
      "args": ["-y", "@browsa/mcp"],
      "env": {
        "BROWSA_API_KEY": "agt_live_..."
      }
    }
  }
}

No LLM secret goes in the config: save your Anthropic, OpenAI or Google key once at app.browsa.io/dashboard/llm-keys and tasks use the workspace's default saved key. (Prefer to keep it local? Add "BROWSA_LLM_API_KEY": "sk-ant-..." instead.)

Restart Claude Desktop. Ask: "Go to news.ycombinator.com and tell me the top story title."

Cursor

~/.cursor/mcp.json:

{
  "mcpServers": {
    "browsa": {
      "command": "npx",
      "args": ["-y", "@browsa/mcp"],
      "env": { "BROWSA_API_KEY": "agt_live_..." }
    }
  }
}

Cline / Continue / Zed

Any MCP-spec-compliant client. Point at npx -y @browsa/mcp and set BROWSA_API_KEY.

Env

VarRequiredWhat
BROWSA_API_KEYyesYour agt_live_* key from browsa.io → Keys
BROWSA_LLM_KEY_IDnoId of a saved LLM key to use by default (see list_llm_keys). Without it, the workspace's default saved key for the model's provider is used
BROWSA_LLM_API_KEYnoA per-task LLM provider key sent with every task. Not needed once a key is saved at dashboard/llm-keys. Wins over BROWSA_LLM_KEY_ID
BROWSA_LLMnoDefault LLM, e.g. anthropic:claude-sonnet-5-5, or a hosted tier such as browsa-fast. Unset: the saved key's provider default model (claude-opus-5-5 with BROWSA_LLM_API_KEY)
BROWSA_COUNTRYnoDefault egress country (e.g. US)
BROWSA_BASE_URLnoOverride prod URL (staging / self-hosted)

Which LLM key a task uses

Per call, first match wins: a hosted tier (llm: "browsa", "browsa-fast" or "browsa-smart": no key, see Models), then llm_api_key, then llm_key_id, then BROWSA_LLM_API_KEY, then BROWSA_LLM_KEY_ID. With none of these, no key is sent and the server uses the workspace's default saved key for the provider of llm (with llm unset: the first default in the order anthropic, openai, google).

If no key is saved at all:

  • No model named (no llm argument, no BROWSA_LLM): run_task and create_schedule still work with zero config. They retry once on Browsa's hosted model (which adds credits per task) and the result has a note saying so: "No saved LLM key found, so this ran on Browsa's hosted model (adds credits per task; see list_models). Save a key at https://app.browsa.io/dashboard/llm-keys to use your own." The same happens if the server has saved keys turned off (LLM_KEYS_UNAVAILABLE). A schedule created this way stays on the hosted model; create it again after saving a key.
  • A model named: there is no fallback. The tool returns an error (code LLM_KEY_REQUIRED) that says to save a key at app.browsa.io/dashboard/llm-keys, set BROWSA_LLM_API_KEY, or pass llm: "browsa".

Models

run_task and create_schedule take llm and reasoning:

llmWhat it is
"browsa"The default hosted model
"browsa-fast"Fastest and lightest
"browsa-smart"For harder tasks
anything else, e.g. "anthropic:claude-sonnet-5-5"A provider model, run with your own key (saved, or llm_api_key)

Hosted tiers need no LLM key and add a flat number of credits per task. Which model each tier runs on, and what it costs, are server settings that can change: the list_models tool returns them (free, read-only), along with the providers you can bring a key for.

reasoning is "low", "medium" or "high"; leave it out for the tier's or the provider's default. Ask in plain language, for example: "Use browsa-smart with high reasoning to compare the pricing pages of these three sites."

run_task and get_job results carry status and outcome. status: completed only means the run ended; outcome: failed with it means the agent finished but could not do the task, and outcome_message and final_result say why. signals lists what was seen on the way (a bot check, a proxy failure) when there is any.

When the job has them, run_task and get_job results include usage (input_tokens, output_tokens, cache_read_tokens, cache_write_tokens, requests), llm_tier, reasoning (the level that was applied) and llm_cost_usd. llm_cost_usd is an estimate from public list prices for runs with your own key; it is null when unknown and always null for hosted runs, which are paid in credits.

Tools exposed

ToolWhat
run_taskRun a browser-driven task. Returns the final result, outcome and steps. Optional llm, reasoning, llm_key_id.
get_jobPoll a long-running task. Optional server-side wait_seconds.
cancel_jobCancel a running task.
respond_to_jobAnswer an input_required prompt (a decision, or a CAPTCHA the task could not solve itself: see https://browsa.io/captcha).
start_browserStart a real cloud browser and get its CDP URL, to drive yourself with Playwright or Puppeteer.
stop_browserStop a browser from start_browser and release it.
proxy_networkWhether Browsa's built-in proxy network is on, its tiers, countries and price. Read-only and free.
observeRead the page of a browser from start_browser as compact text with element ids. No Playwright needed.
actOne action in that browser (click, type, press, select, scroll, navigate, wait and more); returns what changed.
list_browsersList this workspace's browsers and their status.
creditsShow credit balance + top-up packs.
list_webhooks / create_webhook / delete_webhookWebhook CRUD.
list_identitiesList persistent browser identities (use one via profile_id to reuse warmed cookies).
create_schedule / list_schedulesRecurring tasks. create_schedule takes llm, reasoning and llm_key_id; the schedule stores only the key's id.
list_job_filesOutput files a task wrote.
list_llm_keysSaved LLM keys: id, provider, masked key, default flag, status. Read-only; never returns a secret.
list_modelsHosted tiers (model, default reasoning, credits per task) and own-key providers. Read-only and free.

Why MCP-first matters

Most agent platforms gate distribution behind "write code → deploy → call our API." We gate it behind "paste config → ask Claude." For an AI agent platform, the prospect's first 30 seconds is the make-or-break moment. With this MCP server, the first 30 seconds is a working browser doing what they asked.

Source + Issues

License

MIT.