oassis — web scraping, crawling and browser control for AI agents

Measured conclusion: Deep-tested. L2 outcome: tools_ok.

Scrape to markdown, map and crawl sites, drive a real browser. Pay per call, no API key, no signup.

Add to your agent

Copy this into your MCP client config:

{
  "mcpServers": {
    "oassis — web scraping, crawling and browser control for AI agents": {
      "url": "https://api.oassis.dev/mcp",
      "transport": "streamable-http"
    }
  }
}

Measured tools (14)

• web_scrape read-only

  Processes a page and returns every output you ask for at once: markdown, html, links, screenshot, PDF, accessibility tree, elements by selector, AI-structured data, and `controls` (what can be clicked). One call, and a partial failure does not void the rest. From 1 credit ($0.002) per output. A url pointing at a PDF, Word, Excel or CSV file is converted to markdown instead, with no browser, for $0.002. Reference for the oassis API: https://oassis.dev/openapi.json

  Parameters:

  • pdf — object — Options for the `pdf` format: `format` (a4, letter, legal…), `landscape`, `printBackground`, `scale` 0.1 to 2.
  • url — string — Page to process.
  • html — string — Raw HTML instead of `url`, for a page you already have. Nothing leaves for the network.
  • json — object — For the `json` format: a `prompt` describing what to pull out, a JSON `schema` to fill, or both.
  • wait — object — When to consider the page loaded: `until` (load, domcontentloaded, networkidle0, networkidle2), `selector` to wait for, `timeout` in ms.
  • block — object — What not to load, which makes a render faster and cheaper. `resourceTypes` takes image, font, stylesheet, media… and `urlPatterns` are regular expressions, not globs: `\.svg$`, not `*.svg`.
  • links — object — Options for the `links` format: `visibleOnly`, `excludeExternal` to keep only internal links.
  • maxAge — integer — Accept an answer up to this many milliseconds old. A cache hit costs $0.0002 instead of the format price. Leave it out to force a fresh render.
  • formats — array — Outputs you want in the same response. `controls` is the map of what can be clicked; `elements` needs `selectors`; `json` needs `json.prompt` or `json.schema`.
  • request — object — How to make the request: `headers`, `cookies`, `userAgent`, and `auth` for basic authentication.
  • binaryAs — string — How a screenshot or a PDF comes back. `base64` is the only value today.
  • controls — object — Options for the `controls` format: `visibleOnly`, and `limit` (1 to 1000) to cut the map.
  • viewport — object — The window to render in: `width`, `height`, `deviceScaleFactor`.
  • selectors — array — CSS selectors for the `elements` format.
  • screenshot — object — Options for the `screenshot` format: `fullPage` for the whole page, `type` (png, jpeg, webp), `quality` 1-100 (not with png), `omitBackground`, `selector` to capture one element, and its own `viewport`.

  Hints: read-only yes · destructive no · idempotent no · open-world yes

• web_session_open

  Opens a browser session and returns the map of controls. Opening costs 2.5 credits plus the requested outputs; the first minute is included. Further started minutes cost 2 credits with an account key. Use web_act to interact and web_session_close when done. API key and credits required. Reference: https://oassis.dev/openapi.json

  Parameters:

  • url (required) — string — Page to process.
  • pdf — object — Options for the `pdf` format: `format` (a4, letter, legal…), `landscape`, `printBackground`, `scale` 0.1 to 2.
  • json — object — For the `json` format: a `prompt` describing what to pull out, a JSON `schema` to fill, or both.
  • wait — object — When to consider the page loaded: `until` (load, domcontentloaded, networkidle0, networkidle2), `selector` to wait for, `timeout` in ms.
  • block — object — What not to load, which makes a render faster and cheaper. `resourceTypes` takes image, font, stylesheet, media… and `urlPatterns` are regular expressions, not globs: `\.svg$`, not `*.svg`.
  • links — object — Options for the `links` format: `visibleOnly`, `excludeExternal` to keep only internal links.
  • maxAge — integer — Accept an answer up to this many milliseconds old. A cache hit costs $0.0002 instead of the format price. Leave it out to force a fresh render.
  • formats — array — Outputs you want in the same response. `controls` is the map of what can be clicked; `elements` needs `selectors`; `json` needs `json.prompt` or `json.schema`.
  • request — object — How to make the request: `headers`, `cookies`, `userAgent`, and `auth` for basic authentication.
  • binaryAs — string — How a screenshot or a PDF comes back. `base64` is the only value today.
  • controls — object — Options for the `controls` format: `visibleOnly`, and `limit` (1 to 1000) to cut the map.
  • viewport — object — The window to render in: `width`, `height`, `deviceScaleFactor`.
  • selectors — array — CSS selectors for the `elements` format.
  • screenshot — object — Options for the `screenshot` format: `fullPage` for the whole page, `type` (png, jpeg, webp), `quality` 1-100 (not with png), `omitBackground`, `selector` to capture one element, and its own `viewport`.

  Hints: read-only no · destructive no · idempotent no · open-world yes

• web_act

  Runs actions against an open session and returns the resulting state. Actions: {navigate}, {click:{ref}}, {type:{ref,text,clear}}, {select:{ref,value}}, {press}, {scroll}, {wait}, {back}. The `ref` is the one `controls` gave you. It stops at the first failure and tells you where. $0.0005 (0.25 credit) per action, plus requested outputs and any outstanding browser time. API key and credits required. Reference for the oassis API: https://oassis.dev/openapi.json

  Parameters:

  • sessionId (required) — string — The session to act on, from `web_session_open`.
  • actions (required) — array — What to do, in order: click, type, select, scroll, hover, press, navigate, wait. One to twenty-five per call.
  • pdf — object — Options for the `pdf` format: `format` (a4, letter, legal…), `landscape`, `printBackground`, `scale` 0.1 to 2.
  • url — string — Page to process.
  • json — object — For the `json` format: a `prompt` describing what to pull out, a JSON `schema` to fill, or both.
  • wait — object — When to consider the page loaded: `until` (load, domcontentloaded, networkidle0, networkidle2), `selector` to wait for, `timeout` in ms.
  • block — object — What not to load, which makes a render faster and cheaper. `resourceTypes` takes image, font, stylesheet, media… and `urlPatterns` are regular expressions, not globs: `\.svg$`, not `*.svg`.
  • links — object — Options for the `links` format: `visibleOnly`, `excludeExternal` to keep only internal links.
  • maxAge — integer — Accept an answer up to this many milliseconds old. A cache hit costs $0.0002 instead of the format price. Leave it out to force a fresh render.
  • formats — array — Outputs you want in the same response. `controls` is the map of what can be clicked; `elements` needs `selectors`; `json` needs `json.prompt` or `json.schema`.
  • request — object — How to make the request: `headers`, `cookies`, `userAgent`, and `auth` for basic authentication.
  • binaryAs — string — How a screenshot or a PDF comes back. `base64` is the only value today.
  • controls — object — Options for the `controls` format: `visibleOnly`, and `limit` (1 to 1000) to cut the map.
  • viewport — object — The window to render in: `width`, `height`, `deviceScaleFactor`.
  • selectors — array — CSS selectors for the `elements` format.
  • screenshot — object — Options for the `screenshot` format: `fullPage` for the whole page, `type` (png, jpeg, webp), `quality` 1-100 (not with png), `omitBackground`, `selector` to capture one element, and its own `viewport`.

  Hints: read-only no · destructive no · idempotent no · open-world yes

• web_scrape_batch read-only

  A batch OF SCRAPES: reads a list of urls YOU give it (2 to 50, from any sites) and returns a jobId. It discovers nothing on its own — for that use web_crawl. API key and credits required. Poll it with web_batch_status.

  Parameters:

  • urls (required) — array — The pages of this batch. Two to fifty, no duplicates: a batch charges per url.
  • json — object — For the `json` format: a `prompt` describing what to pull out, a JSON `schema` to fill, or both.
  • wait — object — When to consider the page loaded: `until` (load, domcontentloaded, networkidle0, networkidle2), `selector` to wait for, `timeout` in ms.
  • block — object — What not to load, which makes a render faster and cheaper. `resourceTypes` takes image, font, stylesheet, media… and `urlPatterns` are regular expressions, not globs: `\.svg$`, not `*.svg`.
  • links — object — Options for the `links` format: `visibleOnly`, `excludeExternal` to keep only internal links.
  • maxAge — integer — Accept an answer up to this many milliseconds old. A cache hit costs $0.0002 instead of the format price. Leave it out to force a fresh render.
  • formats — array — Outputs you want in the same response. `controls` is the map of what can be clicked; `elements` needs `selectors`; `json` needs `json.prompt` or `json.schema`. A screenshot or a PDF is not available here: every result is kept until you collect it. Ask for those one url at a time on web_scrape.
  • request — object — How to make the request: `headers`, `cookies`, `userAgent`, and `auth` for basic authentication.
  • binaryAs — string — How a screenshot or a PDF comes back. `base64` is the only value today.
  • controls — object — Options for the `controls` format: `visibleOnly`, and `limit` (1 to 1000) to cut the map.
  • viewport — object — The window to render in: `width`, `height`, `deviceScaleFactor`.
  • selectors — array — CSS selectors for the `elements` format.

  Hints: read-only yes · destructive no · idempotent no · open-world yes

• web_batch_status

  Checks a batch of scrapes: status, how many are done, and the results. Free for the owning account. Pass `cancel: true` to stop it.

  Parameters:

  • jobId (required) — string — The batch, from web_scrape_batch.
  • limit — number — Results to return (default 50).
  • cancel — boolean — Stop pending pages and refund pages that never started. An in-flight read may finish and is billed; failed reads are refunded.

  Hints: read-only no · destructive no · idempotent yes · open-world no

• web_map read-only

  Every url of a site, fast and cheap: its sitemap plus, optionally, the links on the page. Use it BEFORE crawling, to see what is there and decide what is worth reading. 1 credit ($0.002) with `includePage: false` (no browser at all), 1.5 credits ($0.003) with the page. Reference for the oassis API: https://oassis.dev/openapi.json

  Parameters:

  • url (required) — string — Page to process.
  • limit — integer — Urls to return. Up to 5000.
  • search — string — Keep only urls containing this text.
  • includePage — boolean — Render the page too, to catch urls the sitemap does not list. `false` uses no browser at all, so it costs less: $0.002 against $0.003.
  • excludePaths — array — Discard urls whose path contains one of these. Up to twenty. It wins over `includePaths`.
  • includePaths — array — Keep only urls whose path contains one of these. Up to twenty.
  • includeSubdomains — boolean — Treat subdomains of the same site as part of it.

  Hints: read-only yes · destructive no · idempotent yes · open-world yes

• web_crawl read-only

  Follows a site's links and reads every page. Returns a jobId; poll it with web_crawl_status. API key and credits required; `limit` controls the maximum pages charged. Use web_map first if you only need the urls. Reference for the oassis API: https://oassis.dev/openapi.json

  Parameters:

  • url (required) — string — Page to process.
  • json — object — For the `json` format: a `prompt` describing what to pull out, a JSON `schema` to fill, or both.
  • wait — object — When to consider the page loaded: `until` (load, domcontentloaded, networkidle0, networkidle2), `selector` to wait for, `timeout` in ms.
  • block — object — What not to load, which makes a render faster and cheaper. `resourceTypes` takes image, font, stylesheet, media… and `urlPatterns` are regular expressions, not globs: `\.svg$`, not `*.svg`.
  • limit — integer — Pages this crawl may read, up to 200. Charged up front; what it does not read comes back.
  • links — object — Options for the `links` format: `visibleOnly`, `excludeExternal` to keep only internal links.
  • maxAge — integer — Accept an answer up to this many milliseconds old. A cache hit costs $0.0002 instead of the format price. Leave it out to force a fresh render.
  • formats — array — Outputs you want in the same response. `controls` is the map of what can be clicked; `elements` needs `selectors`; `json` needs `json.prompt` or `json.schema`. A screenshot or a PDF is not available here: every result is kept until you collect it. Ask for those one url at a time on web_scrape.
  • request — object — How to make the request: `headers`, `cookies`, `userAgent`, and `auth` for basic authentication.
  • binaryAs — string — How a screenshot or a PDF comes back. `base64` is the only value today.
  • controls — object — Options for the `controls` format: `visibleOnly`, and `limit` (1 to 1000) to cut the map.
  • maxDepth — integer — How far from the starting page to follow links. 0 reads only the starting page.
  • viewport — object — The window to render in: `width`, `height`, `deviceScaleFactor`.
  • selectors — array — CSS selectors for the `elements` format.
  • excludePaths — array — Discard urls whose path contains one of these. Up to twenty. It wins over `includePaths`.
  • includePaths — array — Keep only urls whose path contains one of these. Up to twenty.
  • includeSubdomains — boolean — Treat subdomains of the same site as part of it.

  Hints: read-only yes · destructive no · idempotent no · open-world yes

• web_crawl_status

  Checks a crawl: status, pages read, discovered and still queued, and the pages themselves. Free for the owning account. Pass `cancel: true` to stop it.

  Parameters:

  • jobId (required) — string — The crawl, from web_crawl.
  • limit — number — Pages to return (default 50).
  • cancel — boolean — Stop pending pages and refund pages that never started. An in-flight read may finish and is billed; failed reads are refunded.

  Hints: read-only no · destructive no · idempotent yes · open-world no

• web_search_exa read-only

  Search the web with Exa's index: a query instead of a url, for when you do not know where to look. Returns title, url and a snippet per result. To read the pages, pass the urls to web_scrape_batch. The engine is named because the price is Exa's, passed through with no markup and read from its own payment challenge on every call — the quote is read at execution time and is not a guaranteed fixed price. API key and credits required over MCP.

  Parameters:

  • query (required) — string — What to search for.
  • limit — integer — Results to return. Up to 50.
  • since — string — Only results published after this ISO date.
  • domains — array — Only these domains, subdomains included: `example.com` also matches `docs.example.com`.
  • snippets — boolean — Text alongside each result. On by default: a bare url is rarely enough.
  • excludeDomains — array — Never these domains, subdomains included.

  Hints: read-only yes · destructive no · idempotent no · open-world yes

• web_feedback

  Tell us an answer was good or bad. FREE. Use it when a result is wrong — empty markdown, a control map missing a button, data that does not match the page — with the url or the jobId so it can be reproduced. It is the only way we learn that we read a page badly: our logs cannot tell that apart from a page that is simply like that. Reference for the oassis API: https://oassis.dev/openapi.json

  Parameters:

  • verdict (required) — string — Did the answer do the job or not.
  • url — string — The page that came out wrong: usually the fastest way to reproduce it.
  • route — string — Which tool or endpoint it is about, as you called it.
  • comment — string — What you expected and what you got.
  • reference — string — The jobId or sessionId it happened on, so it can be looked up.

  Hints: read-only no · destructive no · idempotent no · open-world no

• web_task_start

  Starts a research task: a question answered by reading public pages, with sources, progress and cost. Requires an API key and credits. The maximum budget is reserved, actual execution cost is charged, and unused credits are released. Returns a task id: check it with web_task_status.

  Parameters:

  • task (required) — string — What to find out, in plain words. Be specific about what to compare.
  • mode — string — Only research is available: read public pages and answer.
  • urls — array — Pages to read first, up to 5. Public http or https urls only.
  • limits — object — Bounds on cost, time and steps.

  Hints: read-only no · destructive no · idempotent no · open-world yes

• web_task_status read-only

  Returns a task's status, progress, sources, limitations, result and billing. Only your own tasks.

  Parameters:

  • id (required) — string — The task id, from web_task_start.
  • access_token — string — Task-scoped access token returned by web_task_start; not an API key.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• web_task_cancel

  Stops a task using its owning API key, or its legacy access_token. No new operations start. Cancellation is free; unused account credit reservations are released. Cancelling a finished task changes nothing.

  Parameters:

  • id (required) — string — The task id, from web_task_start.
  • access_token — string — Task-scoped access token returned by web_task_start; not an API key.

  Hints: read-only no · destructive no · idempotent yes · open-world no

• web_session_close

  Closes the browser session and stops its time. HTTP equivalent: POST /web/v1/session/{id}/close. No additional close fee; outstanding browser time is settled when applicable. If you do not close it, it closes itself after a minute without use.

  Parameters:

  • sessionId (required) — string — The session to close, from web_session_open.

  Hints: read-only no · destructive no · idempotent yes · open-world no

Publisher website

Measured by MCPYP probing. Untested fields are omitted, never estimated. Listing is always free; MCPYP never proxies calls or stores user tokens.