StealthGPT
AIUse the StealthGPT API for AI writing, humanization, and AI-content detection. Agents draft and screen copy in the same workflow that will publish it.
What This Integration Enables
StealthGPT bundles three things that are usually sold by three different vendors: it writes, it rewrites text so it reads as more human-written, and it scores text for how likely it is to be flagged as machine-generated. That combination is not neutral in the market, and it is worth being plain about what the connector actually exposes rather than talking around it. Write Or Humanize Text does both jobs depending on one flag: set Rephrase to false and pass a brief to write from scratch, set it to true and pass raw source text to rewrite what already exists. Detect AI Content runs the same engine as the StealthGPT web detector, accepts up to 3,000 words, and returns a score from 0 to 100 where lower reads as more human. What you do with that score is a decision your organisation makes. FlowRunner's contribution is making sure a person makes it.
Operationally, the connector is built for long work. Synchronous generation is available and fine for short pieces, but anything substantial belongs in the asynchronous shape. Start Stealth Agent Run queues a multi-step job under an Academic, SEO or Social preset, with optional fact-checking and image generation passes, and returns a run id to poll with Get Stealth Agent Run. Start Humanization Run and Start Document Humanization Run do the same for rewriting, the latter fetching a PDF, DOCX, TXT or Markdown file up to 4 MB from a public URL and extracting the text server-side, which means scanned or image-only PDFs will not work. Every long action accepts a webhook URL, but that is a per-request callback rather than a subscription, so this connector has no triggers and flows poll or receive the callback into a separate entry point. Get Word Balance reports the prepaid words remaining and consumes no credits, which makes it the free pre-flight check before any batch.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Long-form production that does not hold a step open
An approved brief starts a run. The agent calls Start Stealth Agent Run with the SEO preset, the fact-check pass on and image generation off, and gets a run id immediately. A polling flow calls Get Stealth Agent Run until the status is completed, failed or cancelled, then reads the finished markdown, the word count and the credits charged. The draft is written to Contentful or WordPress as a draft, never as a published post, and the editorial team picks it up from there. The Academic and Social presets take the same shape, with Social requiring a target platform on the call.
A detection score attached to every piece, before anyone argues about it
Whatever the source of a draft, the flow runs Detect AI Content over it and stores the score in the same record as the text, along with the date it was measured. Pieces are batched under the 3,000 word limit so long articles are scored in sections rather than truncated. The value of doing this inside the workflow is not the number. It is that the number exists at a fixed point in the process, measured the same way for every piece, instead of being produced later by whoever is making an argument.
A document rewritten without going through a person's desktop
A file lands in Google Drive. The agent passes its public URL to Start Document Humanization Run, which downloads and validates the file, extracts the text server-side and queues the rewrite, returning a run id. Get Humanization Run collects the result once complete, along with its human-likeness score and the credits charged. The output always comes back as text regardless of the input format, so the flow is responsible for putting it back into whatever shape the destination needs. Get Word Balance runs before the batch so a large document set fails at the start rather than in the middle.
Human-in-Loop Highlight
The action that stops an agent here is Detect AI Content, and not because it is expensive or irreversible. It is because of what a score turns into when it points at a person. Detect AI Content returns a number between 0 and 100. It is a probability estimate produced by one vendor's engine over at most 3,000 words, and it is not evidence of anything about who wrote the text. The moment a flow lets that number drive a consequence for a named individual, a rejected freelance invoice, a flagged submission, a note in someone's file, the workflow has quietly promoted an estimate into an accusation, and an accusation is not something you take back by rerunning the check. So FlowRunner does not let the agent close that loop. The agent measures, records the score and the date, groups the outliers, and hands the editor the text alongside the number: "Three of eleven submissions from this contributor scored above the threshold you set. The score is a probability, not proof, and two of the three are under 900 words where the engine is least reliable. Review the drafts, or hold the batch?" A person reads the writing and decides. This is the digital andon cord in its most literal form: the agent has hit something it is not qualified to judge, so it stops the line rather than guessing. Human-in-the-loop exists precisely for the decisions where being wrong lands on somebody.
Agent Capabilities
9 actionsWriting
2- Write Or Humanize Text Generates new content from a brief, or rewrites supplied text so it reads as human-written. Set Rephrase to false and pass a topic or instructions to write from scratch; set it to true and pass only the raw source text to rewrite existing writing. Returns the finished text with a 0 to 100 human-likeness score and the word cost of the request. Runs synchronously, so long documents belong in Start Humanization Run instead.
- Generate Blog Article Produces a complete SEO-oriented article from a topic in a single call, optionally embedding relevant images. Choose a target length and whether the article returns as HTML or Markdown so it can go straight into a CMS. Longer sizes consume proportionally more words from the account balance.
Detection
1- Detect AI Content Scores a piece of text for how likely it is to be flagged as AI-generated, using the same engine as the StealthGPT web detector. Accepts up to 3,000 words per request and returns a score from 0 to 100 where lower reads as more human. Used as a recorded measurement in a publishing process, or to measure the effect of a rewriting pass.
Stealth Agent
2- Start Stealth Agent Run Queues a long-form multi-step job and returns immediately with a run id to poll. The agent plans and writes under one of three presets: Academic for essays and research-style documents, SEO for search-oriented articles, and Social for LinkedIn or Medium posts, which requires a target platform. Optional passes fact-check the draft and generate images. A webhook URL can be supplied to be notified on a terminal state instead of polling.
- Get Stealth Agent Run Returns the current state of a Stealth Agent run. While queued or running the response reports progress; once complete it carries the finished markdown, the word count produced and the credits charged. Poll until the status is completed, failed or cancelled.
Humanization Runs
3- Start Humanization Run Queues an asynchronous rewriting job for a block of text and returns a run id straight away, which avoids request timeouts on long pieces. Send only the raw source text with no instruction wrapper, so the whole input is treated as material to rewrite. A webhook URL can be supplied instead of polling.
- Start Document Humanization Run Fetches a PDF, DOCX, TXT or Markdown file from a public URL and queues it for rewriting, returning a run id. The text is extracted server-side and the result always returns as text, so scanned or image-only PDFs will not work. The file must be 4 MB or smaller, and the download is validated before it is forwarded.
- Get Humanization Run Returns the current state of an asynchronous rewriting run started from text or from a document. Once completed, the response carries the output, its 0 to 100 human-likeness score and the credits charged. Poll until the status is completed, failed or cancelled.
Account
1- Get Word Balance Returns the prepaid words left on the account and, where metered billing is enabled, the pay-as-you-go status and any usage still pending billing. Call it before a large batch to fail early, or on a schedule to alert when the balance runs low. This call consumes no credits.
Frequently Asked Questions
What can FlowRunner do with StealthGPT?
FlowRunner agents can run Write Or Humanize Text, Detect AI Content, and Generate Blog Article in StealthGPT, plus 6 more actions.
Does connecting StealthGPT to FlowRunner require OAuth?
No. StealthGPT connects to FlowRunner with an API key, no OAuth flow required.
Can StealthGPT trigger a FlowRunner workflow automatically?
StealthGPT doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with StealthGPT
$100 in credits. No card required. Connect in minutes.