Linkup
AISearch the web with Linkup, an API built for grounding language models, returning ranked pages with full content or written answers with citations. Agents answer from current sources instead of stale model memory.
What This Integration Enables
Most search APIs were built for people, and then got pointed at models. Linkup was built the other way around. Search Web can return ranked pages with full content, a written answer with its sources cited, or a JSON object matching a schema you supply, and it can be restricted to specific domains or a date window before it runs. That last part is what makes it usable in a regulated process: the sourcing is a decision you make up front, not something you audit afterward. Fetch Page handles the other half, turning a single URL into clean markdown a model can read without an HTML parser in between.
There is a design detail worth naming, because it says what kind of tool this is. Linkup charges no credit when a search returns nothing relevant. An empty result is treated as a legitimate outcome rather than something to paper over. That is the same instinct behind FlowRunner's human-in-the-loop model: an agent that admits it has nothing is more useful than one that fills the gap. The searches that come back thin become automation exceptions a person looks at, and everything else keeps moving.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Grounded answers for a prospect question
A question arrives that a sales engineer cannot answer from memory and the model cannot answer from training data, because the fact changed last quarter. The agent runs Search Web in Deep mode, restricted to the vendor's own documentation domains and the last twelve months, and gets back an answer with its citations. It then calls Fetch Page on each cited URL and checks the claim is still present on the live page rather than only in a cached index. The draft goes to the sales engineer as a claim list, each line carrying its source. They approve, and the answer is sent.
Long-form competitive research handed off and collected later
A quarterly market review would take an analyst most of a week. The agent submits it with Start Research Task at a reasoning depth matched to the deadline, then polls Get Research Task on a backoff rather than holding a connection open for twenty minutes. When the status reaches completed, the agent persists the output immediately, because Linkup retains results for a bounded period only, and files the sourced findings into Notion alongside the previous quarter's version for comparison.
A batch of questions run through one queue
An enrichment backlog contains several hundred URLs to read and questions to answer. Instead of firing them one at a time, the agent groups them with Submit Task Batch, which accepts a mix of search, fetch, research and extract work in a single call. It then watches List Tasks, which reports how many tasks are pending or processing against the account's concurrency limit, and uses that number to decide when to queue the next group instead of guessing. Completed results flow into Google Sheets for the team that requested them.
Human-in-Loop Highlight
The gate here is not the search. It is the citation. A grounded answer fails in a specific and quiet way: the search succeeded, the sources are real, and one of them does not actually say what the answer claims it says. So before any Linkup-sourced answer leaves the building, the agent posts it to the reviewer as a mapping, not as prose: "Claim: the Business tier includes SSO. Source: vendor pricing page, published 14 March, re-read just now with Fetch Page. Claim: the audit log retains 90 days. Source: changelog entry, published 2 August. Approve the answer, or reject the second claim?" The reviewer is judging whether each source supports its sentence, which takes a minute, instead of judging a paragraph's plausibility, which takes an afternoon and usually does not happen. Once approved, the answer goes to the prospect and the source list is retained with it, so the same question next quarter starts from evidence rather than from someone's recollection.
Agent Capabilities
12 actionsSearch and Fetch
2- Search Web Runs a synchronous web search built for grounding a model, returning ranked results with full page content, a written answer with cited sources, or a JSON object matching a schema you supply. Fast handles focused lookups in under a second, Standard searches agentically across several sources, and Deep runs multiple iterations for coverage. Restrict to specific domains or a date window when the sourcing has to be defensible.
- Fetch Page Fetches one web page and returns it as clean markdown ready for a model. Optionally renders the page's JavaScript first, returns the raw source alongside the markdown, and extracts every image with its alt text. Used to verify a cited claim against the live page, or to read a single known URL without running a search at all.
Deep Research
3- Start Research Task Submits a deep research task and returns immediately with a task ID while the agent searches, reads and reasons across many sources in the background. Reasoning depth trades latency for thoroughness. Used when the question is worth twenty minutes of an agent's time and none of an analyst's.
- Get Research Task Retrieves one research task by ID so a flow can poll it after submission. Status moves from pending through processing to completed or failed, and the output stays empty until it finishes. Used as the wait step in a long research workflow.
- List Research Tasks Returns a paginated list of the account's research tasks with their status, input and output. Used to reconcile a batch of submitted runs, or to audit what an account has been spending research credits on.
Structured Extraction
3- Start Extract Task Submits an extraction task that crawls from a seed URL and pulls out many rows of structured data described in plain English. Returns a task ID immediately. Used for list-building work where the target shape is known but the page layouts are not.
- Get Extract Task Retrieves one extraction task by ID for polling. When it completes, the output carries the row count, the credits consumed, and a result URL pointing at a newline-delimited JSON file that is valid for 24 hours. Used as the collection step, with a download that cannot be deferred.
- List Extract Tasks Returns a paginated list of the account's extraction tasks with their status and output metadata. Used to find a result URL that has not been downloaded yet, before it expires.
Asynchronous Queue
3- Submit Task Batch Queues between 1 and 100 mixed asynchronous tasks in one call, each with its own type of search, fetch, research or extract. Used when a flow has many questions or URLs to process and should not hold a connection open for each one.
- Get Task Retrieves any asynchronous task by ID, whatever its type, with the output shape following the task's type. Used as the single polling step in a flow that submitted mixed work.
- List Tasks Returns a paginated list of every asynchronous task on the account, filterable by type and status, along with a quota block showing how much work is pending or processing against the concurrency limit. Used to pace a large batch rather than queuing more than Linkup will run in parallel.
Account
1- Get Credit Balance Returns the remaining credit balance on the account in US dollars. Used as a cheap connection check, and as the guard step before an expensive Deep search or a long research run, since running out of credit surfaces as a rate limit response rather than an obvious payment error.
Frequently Asked Questions
What can FlowRunner do with Linkup?
FlowRunner agents can run Search Web, Fetch Page, and Get Credit Balance in Linkup, plus 9 more actions.
Does connecting Linkup to FlowRunner require OAuth?
No. Linkup connects to FlowRunner with an API key, no OAuth flow required.
Can Linkup trigger a FlowRunner workflow automatically?
Linkup doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with Linkup
$100 in credits. No card required. Connect in minutes.