HARPA AI

Dispatch browser work to your own Chrome, Edge, Brave, or Opera through the HARPA AI GRID API. Agents reach pages behind a real logged-in session that a headless scraper cannot.

4 actions API key available

Platform Documentation ↗ Capability data verified 2026-08-01

An account moves to research stage in the CRM and the enrichment flow starts

Agent calls Search Google through the operator's own Node, so results match what a signed in user sees

Agent calls Scrape Page on the shortlisted URLs and receives full page markdown

Agent confirms a Node was actually online rather than accepting an empty response as an empty page

Agent calls Run AI Prompt against the page with the page placeholder to answer one specific question

Agent writes the findings to the CRM record and posts the summary to the deal channel

Rep approves the Node, command name and inputs before any saved AI Command runs unattended in their session

What This Integration Enables

HARPA AI is the odd one out among web connectors, and its oddity is the entire value. HARPA's cloud does not execute anything. It queues work for a Node, which is your own Chrome, Edge, Brave or Opera running the HARPA extension with Remote-Control Browser enabled. The page is fetched by that browser, in that session, with those cookies and those permissions. That is how it reaches an intranet page or a vendor portal you are already signed in to without anyone minting a service credential, and it is why a Google search here returns what a real signed in user sees rather than what a datacenter proxy is shown.

The consequence deserves saying plainly rather than being buried under the benefit. Work dispatched through this connector runs as the person who owns that browser. Their access is the agent's access. FlowRunner's whole argument is that an agent operating with a person's authority should be an agent that knows when to stop and ask them, which is what human-in-the-loop means here, and this is the connector where that argument is least abstract. Two practical constraints follow from the same architecture and are worth planning around: if no Node is online, a synchronous call returns nothing useful and an asynchronous one waits up to 30 days, so the machine needs to stay powered on. And calls block until Nodes finish, up to the timeout, with HARPA returning whatever it has rather than a timeout error, so a flow checks that content actually came back rather than treating an empty response as an empty page.

Without FlowRunner

Cloud scrapers hit the wall Pages behind a corporate sign in or on an intranet are simply unreachable

Anonymous search results Results collected through a scraping proxy are not the results the team is looking at

Research happens in tabs Someone opens the page, reads it, and retypes what they found into the CRM

With FlowRunner

Your session does the work Pages the operator is already signed in to are readable without provisioning a second credential

The same results the team sees A search runs in the real browser, so what the flow records matches what the rep would find

Findings land in the record Page content and the answer to a specific question are written back where the deal lives

Use Case Scenarios

Account research inside the rep's own session

An account moves to research stage. The agent runs Search Google through the rep's Node, using ordinary search operators to narrow to the company's own domain, then calls Scrape Page on the shortlisted URLs and gets full page markdown back. Because the pages load in a browser already signed in to the industry publications and vendor portals the team subscribes to, the paywalled analysis the rep would have read manually is available to the flow. Findings are written back to the account record in HubSpot with the source URL beside each one.

Reading an internal page no cloud scraper can reach

An operations team keeps a rota, an inventory view or a status board on an intranet host that resolves only inside the corporate network. No cloud scraper can see it, and standing up an API in front of it is a project nobody has budget for. The agent calls Scrape Page against the internal URL through a Node running on a machine inside that network, extracts the specific elements it needs with CSS or XPath selectors rather than pulling the whole page, and posts the resulting summary to Slack on a schedule.

Asking one question about one page

Sometimes the requirement is not a dataset, it is an answer. The agent calls Run AI Prompt against an open page with the page placeholder inside the prompt, asking something narrow such as whether a supplier's terms page still states the same notice period. HARPA replaces the placeholder with what the browser can see and returns the answer. The result goes into Notion with a timestamp, so the next person to ask can see when it was last checked and against which URL.

Human-in-Loop Highlight

Run AI Command is the gate on this connector, and the reason is written into the field list. Its Inputs parameter exists so that the command's prompts are answered up front, which the connector documents as running unattended rather than pausing for a human. That is a deliberate removal of a pause, executing a multi-step automation, in a live signed in browser, under a real person's identity. It gets worse in a useful way. HARPA publishes no endpoint listing available AI Command names, so a flow refers to a command by free text copied out of the extension. It also publishes no response schema, so the success shape has to be mapped empirically on the first real call. The flow cannot introspect what it is about to run and cannot reliably tell from the response whether it did what was intended.

So the FlowRunner agent stops before the first unattended execution and every time the target changes. It posts to the owner of that browser: "About to run the AI Command named supplier-terms-update on Node ws-mark-laptop against portal.acme.com, and to auto-answer its prompts with these two inputs. This runs in your signed in session under your account. Confirm, or run it once with the prompts left for you to answer?" Read operations through Scrape Page and Search Google need no such ceremony. A saved command that can act inside somebody's session does, because the person whose authority is being spent is the person who should spend it.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via preferred channel

Human decides

Agent resumes with decision

Agent Capabilities

4 actions

Web

Scrape Page Opens a URL in one of your Nodes and returns the page content, either as full page markdown or as specific elements picked out by CSS, XPath or text selectors. Because a Node is your own browser running the HARPA extension, the page is fetched with your real session, so pages behind a login you already hold work without extra credentials. A webhook URL delivers results asynchronously instead of holding the connection open.
Search Google Runs a Google search through one of your Nodes and returns the results page. Because the search runs in your own browser rather than through a scraping proxy, it reflects the results a real signed in user would see. Google search operators work as normal, so a query can be narrowed by site or by title.

Automation

Run AI Command Runs one of HARPA's built in or your own custom AI Commands against a page, meaning the multi-step automations you build in the extension, such as summarising an article or filling a form. Inputs answer the command's prompts up front so it runs unattended rather than pausing for a human. Result Parameter selects which HARPA variable is returned, using dot notation for nested values. HARPA publishes no endpoint listing available command names, so they are copied from the extension.
Run AI Prompt Sends a free form AI prompt to a page opened in one of your Nodes, with no saved AI Command required. The page content is referenced inside the prompt with a placeholder that HARPA replaces with what the browser can see. The quickest way to ask a one off question about a page from inside a flow.

Frequently Asked Questions

What can FlowRunner do with HARPA AI?

FlowRunner agents can run Scrape Page, Search Google, and Run AI Command in HARPA AI, plus 1 more action.

Does connecting HARPA AI to FlowRunner require OAuth?

No. HARPA AI connects to FlowRunner with an API key, no OAuth flow required.

Can HARPA AI trigger a FlowRunner workflow automatically?

HARPA AI doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.

Start building with HARPA AI

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us