Dumpling AI
AIReach the tools an AI agent needs behind one Dumpling AI key: web scraping and crawling, Google search, news and places, document and image extraction, and video work. Agents research and extract without a connector per task.
What This Integration Enables
Dumpling AI is a toolbelt rather than a product with an opinion. It bundles the things an agent keeps needing and rarely gets from one place. That covers fetching and crawling web pages, Google search across web, news and places, and structured extraction from a page using a schema you supply. It also covers reading documents and images, analysing audio and video, generating images, and running your own Dumpling agents. All of it sits behind one key. The practical effect is that adding a capability to a workflow stops being a procurement conversation. You do not go find a scraping vendor because the flow needs one page, and you do not add an OCR contract because one supplier still sends scans.
That matters for how quickly a team can actually assemble something. FlowRunner's Agent Factory is built on the premise that composing capabilities should take minutes, and a toolbelt connector removes the slowest part of that: the wait for access. What it does not remove is judgment. Several of these operations spend real credits at a rate the caller controls, and several of them reach out to systems that belong to other people. Those are the points where FlowRunner's human-in-the-loop model earns its keep, and where an agent should be asking rather than proceeding.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Structured data off a page without writing a parser
A partnerships team tracks a directory of vendors whose listing pages all look different. Instead of writing a selector set per layout, the agent calls Extract From Page with a JSON Schema describing the fields it wants, and gets back an object in that shape along with a screenshot of the page it read. The screenshot is what makes the result reviewable: when a value looks wrong, someone can see the page the agent saw. Extracted rows land in Airtable with the source URL attached.
Reading a supplier document that arrived as a photo
An invoice arrives as a phone photo rather than a PDF. Convert Document To Text is the cheap path when raw text is all you need, but a photographed receipt needs reasoning, so the agent uses Extract From Image with a prompt and JSON Mode on, and gets back the vendor, total and date as machine-readable fields. Because the action takes several images at once, a batch of a dozen receipts is one call. Anything the model returns with low confidence or missing fields goes to a person rather than into the ledger.
A meeting recording turned into commitments
A customer call recording is dropped into a folder. The agent calls Extract From Audio with a prompt asking for decisions and commitments, and gets back the extracted output plus the duration and credits consumed. Cost scales with length, so the flow trims dead air first. The result is posted into the account's thread in Slack, where the account owner confirms the commitments before any of them become tasks that somebody is measured against.
Human-in-Loop Highlight
Crawl Site is the operation that needs a person in front of it, and not for the reason people expect. Depth controls how many links deep the crawler follows and Limit caps the page count, and those two numbers directly drive both the credit burn and the volume of traffic you send at a site that belongs to someone else. There is no undo on either. You cannot un-spend the credits and you cannot un-appear in a partner's server logs. So the agent does not choose those numbers on its own. It runs a cheap Search Web pass first to see how large the target actually is, then posts a plan: "Seed https://example.com/docs, depth 3, page cap 200, roughly this many credits at 10 per page. Target is a partner property. Approve, reduce the cap, or restrict to the docs subtree?" The research owner sets the ceiling, the crawl runs inside it, and the difference between a useful research job and an incident is a thirty-second decision made before the walk rather than an invoice read after it.
Agent Capabilities
19 actionsWeb
4- Scrape Page Fetches a single web page and returns its content as markdown, HTML or a screenshot, with the page title, the final URL after redirects and page metadata. JavaScript is rendered by default so single-page apps work, and cleaning strips navigation and boilerplate so the output is ready for a model. Used as the default fetch step when you want the whole page.
- Crawl Site Walks a site or sitemap from a starting URL and returns the content of every page it visits, plus the credits consumed. Depth controls how many links deep it follows and Limit caps the page count, so both drive cost directly. Used for corpus building, with the scope agreed before the run.
- Extract From Page Pulls structured data off a single page using a JSON Schema you supply, returning the extracted object and a temporary screenshot of the page it read. Used instead of Scrape Page whenever the shape of the answer is known, so results stay consistent across pages with different layouts.
- Capture Screenshot Renders a page and returns a temporary URL to the image. Captures the full scrolling page or just the viewport, clips to a rectangle, waits for late-loading content and auto-scrolls to trigger lazy images, with cookie banners blocked by default. Used for visual evidence in a review step, and the URL should be downloaded promptly.
Search
5- Search Web Runs a Google web search and returns organic results with titles, links, snippets and positions, plus the featured snippet, related searches and People Also Ask entries when present. Scrape Results fetches the top results' full content in the same call. Used to ground an answer without a second round trip.
- Search News Runs a Google News search and returns articles with title, link, snippet, publication date, source outlet and thumbnail. Used with a date range filter and a schedule to build press monitoring or a competitor watch.
- Search Places Runs a Google Places search and returns local business listings for a query, scoped to a location. Used for prospect list building, address enrichment, or resolving the place identifier a business needs before its reviews can be fetched.
- Get Autocomplete Suggestions Returns the search suggestions Google would offer for a partial query, localised by country, language and place. Used as the cheapest way to mine real search demand before committing to full searches.
- Get Google Reviews Fetches Google reviews for a business identified by a search keyword, a place ID or a CID, sorted by relevance, recency or rating. Returns the aggregate rating, total review count and the individual reviews. Used to pull either representative or extreme feedback into a customer experience workflow.
Documents
5- Convert Document To Text Turns a document into plain text from a URL or a base64 payload, optionally restricted to a page range. Used when raw text is all that is needed, which is cheaper than asking a model to reason over the file.
- Extract From Document Reads one or more documents with a model and answers a prompt about them, returning the output plus page count, file count and credits consumed. JSON Mode returns machine-readable output. Used to compare or consolidate a related set of files in one call.
- Extract From Image Reads one or more images with a vision model and answers a prompt about them, covering OCR of a photographed receipt, describing a product shot, or classifying a screenshot. JSON Mode returns machine-readable output. Used to batch a set of images through a single step.
- Convert To PDF Converts a document into a PDF from a URL or base64 content and returns a temporary URL to the result. Used to normalise mixed file types before archiving, with prompt download since the URL expires.
- Merge PDFs Combines several PDFs into one in the order supplied and returns a temporary URL, optionally producing an archival PDF/A variant or an accessible PDF/UA document with your own metadata stamped on. Used to assemble a submission packet from parts a workflow gathered separately.
Media
3- Get YouTube Transcript Fetches a video transcript as text, optionally with timestamps grouped into fixed-size chunks, falling back to the first available language track and reporting which one it used. Used to bring recorded talks and demos into a research workflow.
- Extract From Video Analyses a video with a model and answers a prompt about it, returning the output alongside the duration and credits consumed. Used to summarise a recording, pull action items from a meeting capture, or classify user-generated content. Cost scales with length.
- Extract From Audio Analyses an audio file with a model and answers a prompt about it, returning the output plus duration and credits consumed. Used to transcribe and summarise calls, extract decisions from a voice note, or classify support recordings.
AI
2- Generate AI Image Generates an image with one of Dumpling's hosted models and returns the resulting image references. The input object is model-specific, so the model's own documentation defines the fields it accepts. Used for asset production inside a content workflow, with a permanent option for images that must outlive the default retention.
- Generate Agent Completion Runs one of your Dumpling AI agents over a conversation and returns its reply, the reasoning steps it took, every tool call it made, and the token and credit usage. A thread ID continues an existing conversation. Used to reuse agent logic already built in Dumpling, with the tool-call log available for review.
Frequently Asked Questions
What can FlowRunner do with Dumpling AI?
FlowRunner agents can run Scrape Page, Crawl Site, and Extract From Page in Dumpling AI, plus 16 more actions.
Does connecting Dumpling AI to FlowRunner require OAuth?
No. Dumpling AI connects to FlowRunner with an API key, no OAuth flow required.
Can Dumpling AI trigger a FlowRunner workflow automatically?
Dumpling AI doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with Dumpling AI
$100 in credits. No card required. Connect in minutes.