Read any web page and get a written answer to your question about it. Renders the page in a real browser, extracts the content, and analyses it in one paid call. Answers only from the page, and says when the page does not contain the answer.
Full website audit in one call: accessibility (WCAG 2.1 A/AA via axe-core), on-page SEO, HTTP security headers and page speed / page weight, from a single browser load of one URL. Cheaper than four separate calls; if any part cannot run, nothing is billed.
Page speed and page weight audit of a web page from one real browser load: total bytes transferred, HTTP request count and DOM node count, measured rather than estimated, with findings when the page is heavy. Page-weight signals, not Core Web Vitals.
Security headers check for a website: audits HTTPS, HSTS, Content- Security-Policy (CSP), X-Content-Type-Options, X-Frame-Options / clickjacking protection, Referrer-Policy and CORS from the real HTTP response of any live URL. Findings with severity for what is missing. Header posture only, not a penetration test.
SEO audit of a web page (on-page SEO check) for any live URL or raw HTML: title tag, meta description, H1 heading structure, canonical link, OpenGraph and social tags, JSON-LD structured data, html lang attribute. Rule-by-rule findings with severity and what is missing or malformed.
Website accessibility audit (a11y, WCAG compliance check): WCAG 2.1 level A and AA conformance of any live URL or raw HTML, tested with axe-core in a real headless browser. Returns every violation with rule id, impact, help text and affected node count. Deterministic rules; a check that cannot run returns an error, never a pass.
Everything about one Base mainnet address in a single call: ETH balance, transaction count, and whether it is a contract or an externally owned account. Three chain reads folded into one answer.
A generic allowlisted read-only JSON-RPC call to Base mainnet: your method, your params. A JSON-RPC error (a revert reason, 'block not found') comes back as the result, not a failure -- it is the chain's own answer to exactly this call.
Run Python code in Google's own hosted sandbox and get the executed code, its output, and the outcome back. Not run in this service's own infrastructure.
Detect anomalies in a target table's time series against a history table, using AI.DETECT_ANOMALIES (TimesFM). Both tables share the same timestamp and value column names.
Forecast a time series in a BigQuery table with Google's pretrained TimesFM model (AI.FORECAST) -- no model to train. Point it at the table and the timestamp/value columns.
Execute a read-only BigQuery SQL query, including against Google's public datasets, and get the rows back as JSON. Every query is dry-run first and refused if it would scan more than the byte ceiling, so cost is bounded before anything runs.
Ask a question in plain language about any BigQuery table, including Google's public datasets, and get a written answer with the SQL and the rows it came from. Writes the query, runs it under a byte ceiling, and reads the result back as an analysis.
Fetch any web page and return its readable text, title, description and links, rendered in a real browser so JavaScript-built pages extract correctly. Falls back to direct fetch when rendering is unavailable.
Fetch any URL and get back exactly what the server sent: status code, headers, and body. Unlike extract.page, every status code is a result, not a failure -- a 404 or 500 from the target is delivered as one.
Answer a specific question about text you supply, using Gemini. Answers only from the material given and says so explicitly when the material does not contain the answer, rather than guessing.
Pull named fields out of text and return them as a JSON object with exactly the keys you asked for. Fields the text does not state come back as null instead of being invented.
A raw completion from a large language model: your prompt, your system message, your provider and model choice among what this deployment has configured (Gemini; Claude on Vertex once Model Garden access is enabled). Unlike llm.analyze, nothing is prescribed about the shape of the answer.
Search for places -- businesses, addresses, points of interest -- via Google's managed Maps Grounding Lite MCP server. Requires the operator to enable the Maps Grounding Lite API and set a key; unavailable until then.
Combine a live crypto spot quote with live prediction-market probabilities on the same subject and get them analysed together in one call. Market data and implied probabilities only -- not investment advice.
Live prediction markets and the probabilities they currently imply, from Polymarket. Search by topic or take the highest-volume markets. Returns each outcome as a percentage, not a raw price.
Live bid, ask and volume for one Coinbase product, from an independent market-data host to market.quote -- a genuine second source, not the same read twice.
Read any web page and get a written answer to your question about it. Renders the page in a real browser, extracts the content, and analyses it in one paid call. Answers only from the page, and says when the page does not contain the answer.
A research brief on a company from live web sources: what it does, its products, and anything notable, cited to [n] source numbers, with thin or conflicting evidence disclosed rather than papered over.
Read any web page and return exactly the fields you name as JSON -- pricing, contact, product names, whatever you ask for. Fields the page does not state come back null rather than invented. Built for pipelines that need a fixed shape.
Probe a customer-specified MCP server's initialize handshake and tools/list: whether it requires authentication, what protocol version it speaks, and which of its tools are not marked read-only.
Transcribe up to 60 seconds / 10MB of audio using Google Cloud Speech-to-Text. Longer audio is refused before payment; this is the synchronous API only.
Deterministic statistics over (x, y) points: ordinary least squares linear regression with standard errors and confidence intervals, a fitted normal distribution with quantiles and probability queries, exact Student t and Jarque-Bera p-values validated at your alpha, and predictions with prediction intervals. Points come inline or from a BigQuery table (two numeric columns; above max_rows the rows are chosen by fingerprint, never at random).
Check up to 10 specific claims against up to 4 specific source URLs. Each claim comes back SUPPORTED, CONTRADICTED or UNSUPPORTED, with the quote the verdict is based on.
Generate a short video from a text prompt using Veo. Off by default: enabled only once the operator has confirmed the configured Veo model resolves on this project (see the provider module for why).