unbrowser by Unchained

unbrowser: лёгкий браузер для LLM-агентов без Chrome

От сообществаРаботаетБез входаГлобальныйБесплатноТолько чтение

Что умеет

  • Переходить по URL и разбирать HTML в DOM
  • Искать элементы по CSS-селекторам и диагностировать запросы
  • Извлекать текст, таблицы и карточки со страницы
  • Заполнять формы, кликать и отправлять данные
  • Строить карту маршрутов и модель страницы

Какие данные видит

Ассистент видит содержимое запрошенных веб-страниц через хостед-эндпоинт, предназначенный для публичного тестирования, а не для боевых сценариев.

Нужен ли аккаунт

Не нужен: сервер работает без входа

unbrowser — MCP-браузер для агентов, работающий без Chrome: один нативный бинарник, компактные BlockMap-представления страниц, устойчивые сессии и ограниченное выполнение JavaScript. Ассистент может переходить по ссылкам, извлекать текст и таблицы, заполнять формы и кликать по элементам.

Как подключить

Как подключить unbrowser by Unchained к Claude

Адрес сервера, который нужно вставить в настройках Claude:

https://unbrowser--protostatis-dev.run.tools

  1. Откройте Claude (сайт claude.ai или приложение для компьютера).
  2. Перейдите в Настройки → Коннекторы (в английском интерфейсе: Settings → Connectors).
  3. Нажмите Добавить пользовательский коннектор (Add custom connector).
  4. Вставьте адрес сервера, скопированный ниже, в поле URL удалённого MCP-сервера и нажмите Добавить.
  5. В чате нажмите + → Коннекторы и включите новый коннектор.

Пользовательские коннекторы доступны на тарифах Free, Pro, Max, Team и Enterprise. На Free можно добавить только один. В Team и Enterprise коннектор сначала добавляет владелец организации в Настройки организации → Коннекторы.

Авторизация

После нажатия Добавить откроется окно входа в сервис. Войдите под своим аккаунтом и подтвердите доступ. Пароль Claude не видит.

Как подключить unbrowser by Unchained к ChatGPT

Адрес сервера, который нужно вставить в настройках ChatGPT:

https://unbrowser--protostatis-dev.run.tools

  1. Откройте ChatGPT в браузере (chatgpt.com). Нужен тариф Plus, Pro, Business, Enterprise или Edu.
  2. Один раз включите режим разработчика: Настройки → Приложения → Дополнительные настройки → Режим разработчика (Settings → Apps → Advanced settings → Developer mode).
  3. Откройте Настройки → Коннекторы и нажмите Создать (Create).
  4. Заполните форму: Название (любое), Описание (одна фраза, что делает сервис), URL MCP-сервера (скопируйте ниже).
  5. В поле Аутентификация выберите OAuth, если сервис требует вход, иначе Нет. Нажмите Создать.
  6. В новом чате откройте + → Приложения/Коннекторы и включите коннектор.

Название раздела у OpenAI менялось (Connectors → Apps/Plugins). Если пункт называется иначе, ищите «режим разработчика» в настройках. В рабочих пространствах Business/Enterprise администратор должен разрешить пользовательские коннекторы.

Авторизация

При первом использовании ChatGPT откроет окно входа в сервис. Войдите и подтвердите доступ.

Как подключить unbrowser by Unchained к Cursor

Адрес сервера, который нужно вставить в настройках Cursor:

https://unbrowser--protostatis-dev.run.tools

Быстрый способ: нажмите кнопку Открыть в Cursor ниже, Cursor сам предложит добавить сервер.

Вручную:

  1. В Cursor откройте Settings → Cursor Settings → MCP и нажмите Add new global MCP server.
  2. Вставьте JSON, скопированный ниже, в открывшийся файл ~/.cursor/mcp.json (для одного проекта: .cursor/mcp.json в корне проекта).
  3. Сохраните файл. Сервер появится в списке MCP; при необходимости пройдите авторизацию по кнопке рядом с ним.

Авторизация

Если сервис требует вход, рядом с сервером в списке MCP появится кнопка авторизации.

Список инструментов сервера (32)

Технические названия из tools/list. Нужны только разработчикам.

navigateFetch a URL with Chrome-fingerprinted HTTP using the active profile. Parses HTML, seeds the JS DOM, returns BlockMap inline. With `exec_scripts: true`, extracts inline AND external <script> tags from the parsed HTML, fetches externals in parallel (8s per-fetch timeout), eval's them in document order in QuickJS (with shims for setTimeout/fetch/etc.), then settles the event loop and fires DOMContentLoaded + load. `<script async>` is honored: async scripts execute after the sync queue. When `--policy=blocklist` is set, tracker URLs are blocked at script-fetch time (see scripts.policy_blocked in the result). Returns a `scripts` summary with inline_count, external_count, async_count, policy_blocked, executed, errors. Auto-extract: when the page embeds JSON-bearing <script> tags (density.json_scripts > 0 — covers application/json, application/ld+json, text/x-magento-init, text/x-shopify-app, etc.), navigate auto-runs `extract()` and returns the result as the `extract` field. Saves a round trip on the common case where the data the JS would have rendered is already sitting in the HTML — JSON-LD article schemas on news sites, __NEXT_DATA__ page state on Next.js apps, json_in_script product blobs on Magento/Shopify, GitHub RSC payloads, etc. Capped at 16 KB inline; over that limit `extract` returns a stub with strategy/confidence/size_bytes/hint and the agent should call `extract()` explicitly to retrieve the full payload. Pages with no embedded JSON get extract:null and pay zero extra cost. Tool advice: navigate also returns `tool_likelihoods` plus `tool_recommendations`, derived from concrete page signals (structure/headings, selector hints, density, embedded data, network captures, challenge state, and script pathology) so agents can pick the next tool without guessing. Auto-solve: Reddit's JS proof-of-work challenge (provider: reddit_js_challenge) is transparently solved — the challenge is detected, the GET solution URL is computed (solution = hex_value + hex_value), and the real page is returned in one navigate call. challenge:null on the result means the real page was served. Subsequent navigations in the same session carry the clearance cookie and skip the challenge entirely.
queryRun a CSS selector against the current page's parsed DOM. Returns matching elements as [{ref, tag, attrs, text, text_chars, text_truncated}]. Element refs (e:NN) are stable handles for use with click/type/submit. Selector engine supports tag, id, class, attribute matchers (=, ^=, $=, *=, ~=), all four combinators (descendant, >, +, ~), pseudo-classes (:first/last/nth-child including An+B formulas, :first/last/nth-of-type, :only-child/of-type), :not(), and :has().
query_debugDiagnose why a CSS selector did or did not match. Returns matched_count, sample matches, DOM summary counts, selector hints (top tags/classes/data attrs/ids), and actionable hints for selector_miss, thin_shell, or embedded_json. Use this when query() returns [] and you need to distinguish a bad selector from an empty/browser-rendered DOM.
textGet the textContent of the FIRST element matching the selector (default: body). Note: on Wikipedia/MDN/news sites, the first <p> is often a hatnote or image caption, not the lead paragraph — prefer `text_main` for reading the page's primary content.
text_mainGet the textContent of the page's main content area, excluding chrome (header/nav/footer/aside). Tries <main>, then [role=main], then a single <article>, then falls back to the longest non-chrome subtree. Use this for reading article body / docs page / blog post content.
text_cleanReturn chrome-stripped, JSON-stripped, whitespace-collapsed text from a selector or the best content root. Drops script/style/noscript/svg and page chrome (nav/header/footer/aside) plus obvious hidden widgets and repeated boilerplate.
find_textFind localized text matches and return [{ref, tag, attrs, before, match, after, text}]. Ranks article/main/content matches above nav/header/footer boilerplate.
text_aroundReturn cleaned surrounding text around an element ref or the best ranked text match. Returns {ref, before, match, after, text}.
query_textFind elements by visible text content. Returns the smallest/deepest element whose textContent matches the needle, with chrome (header/nav/footer/aside) skipped. Anchor-promotion: a span/strong/etc. inside an <a> resolves to the anchor (so click() targets the actionable element). Right tool when CSS selectors are unstable (React-rendered pages with hashed class names) but the visible label is reliable — e.g. find a 'Sign in' button without knowing its class.
blockmapRecompute the BlockMap for the current page. Use after eval'd JS or click/type modifies the DOM. Same shape as the inline blockmap from navigate.
page_modelRender the current page into semantic, task-discoverable JSON objects. Reconstructs page structure as search_form, nav_link, article_card, course_card, model_card, product_card, table, answer_block, and limitation objects with actions, normalized fields, goal-based scoring, and provenance. Prefer this as the first planning tool after navigate when raw links/text are too wide.
route_discoverFind page-owned navigation/search routes for a goal. Returns ranked visible links, forms with controls/query_url previews, and inferred URLs derived from page-owned routes plus goal terms. Use before guessing URLs manually.
discoverHigh-level cheap-first information discovery. Optionally navigates to a URL, runs light JS, merges DOM routes, inferred form/query URLs, and network JSON routes into one ranked graph with provenance plus route-level escalation hints. Use this when the task is to find where information lives before extracting it.
network_extractParse captured JSON/API/network responses into semantic objects with fields, scores, matched query terms, and capture/path provenance. Use after navigate or activate when network_stores shows JSON/GraphQL/NDJSON captures and raw body_preview is too noisy.
extractAuto-strategy structured-data extraction. Tries JSON-LD (schema.org) → __NEXT_DATA__ → Nuxt → JSON-in-script (Magento, Shopify, BigCommerce custom-typed scripts) → OpenGraph/meta → microdata → text_main fallback, returns the highest-confidence hit as {strategy, confidence, data, tried}. Use this as the one-shot 'give me the data, you figure out how' call when you don't want to plan the strategy yourself. Pass strategy='json_ld' (or any of the names above) to force a specific extractor.
extract_tablePull a <table> into {headers, rows, row_count}. Headers come from <thead><th>...</th></thead> if present, else the first <tr>'s <th> cells. Each subsequent <tr>'s <td> cells become a row dict keyed by header (or 'col_N' if no header for that column). Right tool for pricing tables, specs, finance/listings tables — saves writing the per-cell mapping eval.
table_to_jsonAlias for extract_table with a first-table default. Pulls a table into {headers, rows, row_count}; selector defaults to 'table'. Use this when an agent expects a table-to-JSON convenience tool.
extract_listPull a repeated card pattern into [{...}, {...}]. Right tool for HN-style lists, search results, product grids — collapses per-site eval boilerplate. Field spec shapes: 'css selector' (text content), 'css selector @attr' (attribute), or ['css selector', '@attr'] (tuple form). If a sub-selector returns null, the field value is null.
extract_cardsAuto-detect repeated article/card/product/course/listing blocks and return normalized items [{title, price, condition, url, availability, snippet, meta, image_alt, score}]. Prefer this over extract_list when the page has semantically ambiguous recipe, course, product, or model cards and you do not already know field selectors. Optional selector scopes detection to known card nodes; kind can bias scoring (recipe, course, product, listing).
settleDrain the JS event loop: alternately runs queued microtasks (Promise resolutions) and fires expired setTimeout/setInterval callbacks, sleeping to the next deadline when only timers remain. Returns when the queue is empty OR max_ms elapses OR max_iters iterations complete. Defaults: max_ms=2000, max_iters=50. Use after seeding the DOM (or after eval'd code that schedules timers) to let pending callbacks run.
clickDispatch a click event on the element at `ref` (e.g. e:142, returned from query). If the element is <a href> and the click was not preventDefault'd, auto-follows the href via navigate (returns the full navigation result with new BlockMap). Otherwise returns {ok, ref, tag, follow: null}.
activateHigher-level action probe. Clicks an element by ref or visible action text, settles, and returns before/after URL, BlockMap/page_model summaries, network counts, hashes, and classification: navigated, dom_changed, network_changed, no_effect, or unsupported.
typeSet the value of an input/textarea (referenced by `ref`) and dispatch input + change events. Use before submit on form fields.
submitSubmit a form by gathering input/textarea/select values and navigating to the resolved action URL. Supports GET and application/x-www-form-urlencoded POST. Checked checkbox/radio values are serialized; multipart upload forms are not supported.
bodyReturn the raw HTML body of the last navigation. Use as a fallback when the BlockMap or selectors aren't enough — but the response can be large (often 100KB+).
evalRun arbitrary JavaScript in the embedded QuickJS runtime against the current page's parsed DOM. Returns the JSON-stringified result. Power tool — prefer query/text/blockmap when the CSS selector engine can express what you need. Canonical param is code; raw JSON-RPC also accepts script or expression aliases and errors if no code-like param is present.
cookies_setAdd cookies to the session jar. Each item is an object {name, value, domain, path?, secure?, http_only?, url?} or a raw Set-Cookie string. Used to replay clearance cookies (e.g. PerimeterX _px3) lifted from a real Chrome session, bypassing bot detection without running the challenge JS.
cookies_getReturn all cookies currently in the jar as [{name, value, domain, path, secure, http_only}]. Use this to export cookies to disk for a later session.
cookies_clearDrop all cookies from the jar.
report_outcomeBind a task outcome (success/failure/quality) to a previous navigation_id from a navigate() call. Used by the policy framework's outcome protocol — see docs/probabilistic-policy.md §4.5. v0 emits an outcome_reported NDJSON event for the navigation; no posterior updates yet. Drivers should call this once per agent task so future Bayesian phases (B/D-2) can attribute extraction success/failure to specific policy decisions.
network_storesReturn content-bearing fetch/XHR responses captured during navigate, ranked by likely content value. SPAs often keep their data in API responses (JSON, GraphQL, NDJSON, Next/Nuxt route data) that are cleaner than the rendered DOM — this tool surfaces them directly. Each entry has capture_id, URL, status, content-type, body_preview (truncated to 256 KB), body_bytes (full size), body_truncated flag, navigation_id, and a heuristic score. Bodies for trackers/ads/CSS/HTML/media are NOT captured. The navigate result already contains a top-5 summary scoped to that navigation; use this tool to get more entries, filter by host, or pull captures from a different navigation.
network_stores_clearDrop all captured network responses from the session's network store. Use this between unrelated navigations if you don't want earlier captures showing up in later network_stores calls.