KitewellRelease statusComparison
How Kitewell compares
RPA suites, AI agents, browser-automation libraries, and app connectors each solve part of this well. Here is what each does, where Kitewell differs, and when another is the better choice. Facts about other products come from their own pages, checked on the date shown.
Checked as of
At a glance
Scroll the table sideways to see every column.
| Category | Where it runs | A routine run | When a screen changes | People in the loop | What it works on |
|---|---|---|---|---|---|
| Kitewell | Your own Mac or Windows PC | Repeats website and desktop actions that worked, without asking the model. Reading a page and checks written in words ask it each run. | AI finds the new way and keeps it for later runs. The run record shows what it fixed. | Any step can wait for approval or be sent back with a note, from the app, a notification, or Claude and ChatGPT. | Websites, desktop apps, Excel, email, scripts, APIs, and coding agents |
| Desktop RPA suites | Windows PCs (Power Automate for desktop). UiPath in its cloud or on your premises. | Runs the flow by its recorded selectors. AI steps in when an element is not found. | AI self-healing: UiPath Healing Agent, Power Automate (preview, runtime only), Automation Anywhere's fallback | People review suggested fixes: UiPath recommendations, Power Automate repair, Automation Anywhere in Debug mode | Desktop apps and web pages |
| Japanese RPA | WinActor: licensed per PC (node-locked). BizRobo!: desktop to server editions. | — | AI help for error questions (WinActor 7.7). AI that analyzes robots (BizRobo! AI Apps). | — | — |
| AI agents that operate a computer | An environment you provide (OpenAI, Anthropic). A browser extension (NEC cotomi Agent). | The model chooses each mouse and keyboard action, and your code carries it out. cotomi Agent reproduces a procedure shown once. | The model sees the screen again at each step. | Vendors advise that a person confirm consequential actions. | Browsers and desktops you provide. Gemini: browsers and mobile. cotomi Agent: work on the web. |
| Browser-agent developer tools | Hosted cloud browsers (Browserbase, Browser Use), or a browser your own code starts | Replays cached actions or generated code without the model | Falls back to the agent. Skyvern regenerates its code automatically. | Skyvern lists human-in-the-loop on Enterprise | Websites |
| App connectors | n8n can be self-hosted, on premises or in a private cloud. | A trigger, then actions in connected apps | — | Approval steps: Zapier on paid plans, n8n before an AI agent uses a tool | Apps with integrations: 10,260+ (Zapier), 2,366 (n8n) |
Each row sums up the facts in its section below, where every fact links to its source. A dash means the sources checked for this page do not cover it, not that the product lacks it.
Desktop RPA suites
UiPath, Power Automate, and Automation Anywhere replay recorded or designed flows, and each now uses AI to recover when a screen changes.
UiPath
Healing Agent works in two modes. In self-healing mode it applies fixes during a run and continues; in recommendation mode it suggests fixes that developers can review and apply.
Source: UiPath blog · checked
Basic starts at $25 a month, is hosted in UiPath's European region, and has limited users and robots. Self-healing UI automation is included in Enterprise and available for purchase on Standard; both are priced through sales.
Source: UiPath pricing · checked
On premises, Healing Agent in Automation Suite 2.2510 needs AI Center's Computer Vision skill and an LLM, either UiPath's or your own through the AI Trust Layer.
Source: UiPath Documentation · checked
UiPath ranked first in Japan's RPA market by revenue share for the eighth year in a row, in ITR Market View for fiscal 2024.
Source: UiPath newsroom · checked
Where Kitewell differs
Kitewell is one app on the computer you already have, with nothing else to host. AI builds the steps from your words, and fixing a changed screen is part of every plan, including the free one. The fix is kept for later runs and shown in the run record.
Choose UiPath instead when…
You run a large automation program with a center of excellence, many robots under central governance, and a budget for an enterprise platform. UiPath's scale and its track record in Japan fit that better.
Power Automate
Self-healing (preview) in Power Automate for desktop covers desktop and browser automation. When an element or window is not found, AI picks the most likely element and recaptures its selector so the run continues. It is runtime-only, Microsoft notes it might occasionally identify an unintended element, and it is included at no extra cost for organization premium accounts.
Source: Microsoft Learn · checked
Self-healing uses GPT-4.1 mini and Claude Sonnet 4.5. Administrators must enable Anthropic as a subprocessor and, outside US regions, allow moving data across regions. It is not available in government or sovereign clouds.
Source: Microsoft Learn · checked
Repair with Copilot (preview) shows the selector Copilot found, and a person applies it once or for every run. It is available only in the US region, and Copilot is optimized for English.
Source: Microsoft Learn · checked
Power Automate for desktop runs on Windows 10, Windows 11, and Windows Server, and needs an active Internet connection.
Source: Microsoft Learn · checked
Using Power Automate for desktop with a Microsoft account is available at no extra cost.
Source: Microsoft Learn · checked
Where Kitewell differs
Kitewell runs on macOS as well as Windows and needs no Microsoft 365 tenant. A fix for a changed website is kept and used on later runs, and you choose the model, such as Anthropic, OpenAI, or Gemini.
Choose Power Automate instead when…
Your company is standardized on Microsoft 365, your flows run on Windows, and your IT team accepts the cloud AI and data settings that self-healing needs. Power Automate is already in that stack, and desktop use with a Microsoft account costs nothing extra.
Automation Anywhere
When an object is not detected within the timeout, Generative Recorder's generative-AI fallback looks for it. It then recommends updating the automation, which can be accepted in Debug mode but not in Run mode.
Source: Automation Anywhere Documentation · checked
A Vision Fallback combines OCR, object detection, and vision-language models. Automation Anywhere cites a 60% improvement in fallback efficacy.
Source: Automation Anywhere Community · checked
Where Kitewell differs
Kitewell keeps the new way a site or app needs as soon as it works, with no separate step to accept it, and the run record shows which step AI fixed. One person can start on their own Mac or PC.
Choose Automation Anywhere instead when…
You already run many bots on Automation 360 and want AI fallback across them without moving anything.
Japanese RPA
WinActor and BizRobo! are widely used in Japan, and their newest AI features help people build and maintain robots.
In Nork Research's March 2025 survey of 1,300 Japanese companies with annual revenue under ¥50 billion, NTT-AT, Usac System, and the Open group ranked at the top for RPA adoption, and domestic vendors held most of the top 10 (reported by IT Leaders).
Source: IT Leaders · checked
WinActor
WinActor 7.6, on sale from July 17, 2025, added scenario templates drafted by generative AI and refined through dialogue, and integration with image-recognition AI (reported by Impress Cloud Watch).
Source: Impress Cloud Watch · checked
WinActor 7.7, released September 1, 2026, added AI help for questions about operation and errors, and generative-AI support for converting VBScript libraries to Python.
Source: WinActor (NTT-AT) · checked
WinActor Manager on Cloud 4.0, available from February 16, 2026, adds an MCP server so AI agents can run existing WinActor scenarios (reported by Impress Cloud Watch).
Source: Impress Cloud Watch · checked
A node-locked license costs ¥1,098,680 for the full version and ¥300,080 for the runtime version, tax included (reported by Impress Cloud Watch, July 2025).
Source: Impress Cloud Watch · checked
Where Kitewell differs
In Kitewell, AI builds the steps from your words, and when a website or app changes, AI finds the new way and keeps it. The same workflow can also handle email, Excel row by row, and coding agents.
Choose WinActor instead when…
Your WinActor scenarios run fine and your team knows how to maintain them. Keep them; with WinActor Manager on Cloud 4.0, AI agents can even call them.
BizRobo!
BizRobo! AI Apps uses AI to analyze robot files and write their specifications, map running robots to business flows, and draft a robot's wireframe from a task described in chat.
Source: BizRobo! (Open, Inc.) · checked
BizRobo! comes in editions from desktop to server.
Source: BizRobo! (Open, Inc.) · checked
Where Kitewell differs
Kitewell starts on one computer, free for one project, and someone who is not an RPA developer can describe the job in words. When a website changes, AI finds the new way and keeps it.
Choose BizRobo! instead when…
Your BizRobo! robots run reliably on a server and your team maintains them. Keep them, and look at Kitewell only for the work they don't cover.
AI agents that operate a computer
These agents look at the screen and decide each click as they go. The vendors' APIs leave the environment, the schedule, and the safeguards to the developer.
Gartner expects AI agents to complement, not replace, RPA and business process automation in the near term, and notes that in frequently run processes an AI agent can cost about 30 times as much as an RPA bot (Gartner's Saikat Ray, via @IT, January 2026).
Source: @IT · checked
OpenAI computer use
With the computer use tool, the model returns structured mouse and keyboard actions, and your application carries them out in a browser or desktop.
Source: OpenAI API docs · checked
OpenAI's guide says to use an isolated browser or VM with an allow list, confirm consequential actions, and treat screen content as untrusted.
Source: OpenAI API docs · checked
Where Kitewell differs
Kitewell is a finished app around the model. It works out a website or desktop task once, repeats what worked without asking the model while the screen still matches, and adds schedules, approvals, and a run record.
Choose OpenAI computer use instead when…
You are building computer use into your own product and will run and secure the environment yourself.
Anthropic computer use
With the computer use tool, your application carries out each action in an environment you control and returns the results, such as screenshots, to Claude. Claude does not connect to that environment directly.
Source: Claude Docs · checked
The tool is generally available on the Claude API and Google Cloud, and in beta on Amazon Bedrock and Microsoft Foundry.
Source: Claude Docs · checked
Anthropic recommends a dedicated virtual machine or container, an allowlist of domains, and a person confirming decisions with meaningful real-world consequences.
Source: Claude Docs · checked
Where Kitewell differs
Kitewell can use Claude as its model and can run Claude Code as a step on your own sign-in. Around the model it adds replay of what worked, schedules, approvals, and a run record.
Choose Anthropic computer use instead when…
You are a developer adding computer use to your own agent or product and want the API directly.
NEC cotomi Agent
NEC announced cotomi Agent on August 18, 2026. It is sold through Microsoft Marketplace at ¥15,000 per user per month excluding tax, plus fees for a dedicated tenant (reported by Impress Cloud Watch).
Source: Impress Cloud Watch · checked
It automates work on the web. After an experienced employee shows a procedure once, it can reproduce it, and it installs as a browser extension (reported by Impress Cloud Watch).
Source: Impress Cloud Watch · checked
Its agent technology, cotomi Act, scored 80.4% on the WebArena benchmark, above the human task success rate of 78.2% (reported by Impress Cloud Watch, December 2025).
Source: Impress Cloud Watch · checked
Where Kitewell differs
Kitewell works desktop apps as well as websites, on your own computer. Team is priced per computer rather than per user, and one project is free.
Choose NEC cotomi Agent instead when…
Your work happens in the browser, you buy through Microsoft Marketplace, and you want an agent from a large Japanese vendor that learns from a veteran's demonstration.
Gemini 2.5 Computer Use
Gemini 2.5 Computer Use launched in public preview through the Gemini API in Google AI Studio and Vertex AI on October 7, 2025. Google says it is optimized for web browsers, shows promise for mobile, and is not yet optimized for desktop OS-level control.
Source: Google blog · checked
Developers can instruct it to refuse, or to ask the user for confirmation, before certain high-stakes actions.
Source: Google blog · checked
Where Kitewell differs
Kitewell works desktop apps as well as websites, on Windows and macOS, and can use Gemini as its model.
Choose Gemini 2.5 Computer Use instead when…
You are building a browser or mobile agent into your own product on Google Cloud.
Browser-agent developer tools
These tools let developers have AI work a site once, then replay what worked and fall back to AI when the page changes.
Skyvern
After an agent run succeeds, Skyvern can generate code from the actions it took, and later runs execute that code instead of asking the model. If the code fails because the page changed, Skyvern falls back to the agent and regenerates the code; this is on by default.
Source: Skyvern docs · checked
Plans are Free, Hobby at $29 a month, Pro at $149 a month, and Enterprise at a custom price. Human-in-the-loop is listed on Enterprise only.
Source: Skyvern pricing · checked
Where Kitewell differs
Kitewell is an app rather than a developer platform. Any plan can stop a step for a person's approval, and the same workflow can include desktop apps, Excel, and email.
Choose Skyvern instead when…
You are coding browser automation into your own product, want hosted browsers and an API, and need many sessions at once.
Stagehand and Browserbase
Stagehand caches act() results. On a cache hit, Browserbase's server returns the result with no LLM inference; if the recorded selector no longer resolves, Stagehand falls back to full inference.
Source: Stagehand docs · checked
The cache lives on Browserbase's servers. With a local browser the cache option has no effect, and every call runs inference.
Source: Stagehand docs · checked
Browserbase plans are Free, Developer at $20 a month, Startup at $99 a month, and Scale at a custom price. Asia Pacific is available on request for enterprise customers.
Source: Browserbase pricing · checked
Where Kitewell differs
Kitewell keeps what worked on your own computer and runs it in your own Chrome or Edge, with the sign-ins saved there.
Choose Stagehand and Browserbase instead when…
You are a developer writing browser automation in code and want hosted browsers that scale in the cloud.
Browser Use
workflow-use creates deterministic workflows that fall back to the Browser Use agent when a step fails. Its README says it is in very early development and not recommended for production.
Source: GitHub: browser-use/workflow-use · checked
Cloud browsers cost $0.02 an hour, and agents are billed at the model's cost plus 20%, plus browser time and traffic. There are no subscription plans.
Source: Browser Use pricing · checked
Where Kitewell differs
In Kitewell, replaying what worked and fixing it when a site changes is how every website step runs by default, with no code to write.
Choose Browser Use instead when…
You want an agent library to build your own automation on, or pay-as-you-go cloud browsers for your own code.
App connectors
Zapier and n8n connect apps through their integrations and APIs rather than by working a screen.
Zapier
Zapier says it connects 10,260+ apps, with no code required.
Source: Zapier · checked
Each Zap consists of a trigger and one or more actions.
Source: Zapier Help Center · checked
Human in the Loop's Request Approval action pauses a Zap so reviewers can approve, decline, or change data. It is on the Professional, Team, and Enterprise plans.
Source: Zapier Help Center · checked
Where Kitewell differs
Kitewell works the screens of websites and desktop apps that have no integration, on your own computer. A Zap can also start a Kitewell workflow through a webhook, on Personal and Team.
Choose Zapier instead when…
Every system involved has a Zapier integration. Connecting through integrations is quicker to set up and steadier than working a screen.
n8n
n8n can be self-hosted on your own infrastructure, on premises, or in a private cloud.
Source: n8n Docs · checked
n8n lists 2,366 integrations.
Source: n8n · checked
Human-in-the-loop for tools requires a person's approval before an AI agent runs specific tools, through n8n's chat, Slack, Microsoft Teams, Gmail, and other channels.
Source: n8n Docs · checked
Where Kitewell differs
Kitewell is a desktop app with no server to run, and it works the screens of websites and desktop apps as well as calling APIs.
Choose n8n instead when…
Your systems have APIs, and you want a visual workflow tool you can self-host for your team.
Where Kitewell falls short
Where another tool may serve you better, or where Kitewell has limits today.
- New, from a small companyKitewell is made by Descarty, Inc. Workflows are plain YAML on the open-source Dagu engine, and a project can be exported any time.Details →
- Slower than scripted RPA on fixed, high-volume entryA robot replaying a fixed script enters data faster. Kitewell waits for each page, and asks a model wherever a step reads a page or checks something written in words.Details →
- macOS desktop replay works from the screenOn Windows a replay finds the app's own buttons and fields. On macOS it needs the same screen as when the task was learned, and reading from the screen asks the model every run.Details →
- The computer has to be on and awakeKitewell runs while the computer is awake and you are signed in. It does not wake a sleeping computer; scheduled starts it missed run when the computer is back.Details →
- Website steps use Chrome or EdgeChromium or Brave works when you give its path. Building and fixing a website step takes an Anthropic, OpenAI, or Gemini model; small local models usually cannot drive a browser.Details →
- A cloud model sees what it works onWhen AI decides a desktop action, a screenshot of the whole screen goes to the model's provider; a website step sends the page's visible text and layout. Replayed actions send nothing, and with a local model the prompts stay inside your network.Details →
- Gmail connects with an app password for nowGoogle is still reviewing Kitewell's Gmail access, so a Gmail mailbox starts from an app password until then.Details →
- Team covers up to 25 peopleFor a larger organization, Enterprise is arranged under a contract.Details →
How these facts were checked
Each fact about another product links to the page it comes from, usually the vendor's own documentation, pricing, or announcement, and shows the date that page was opened to confirm it. Figures from news reports and surveys name who published them. Products change quickly, so check the linked page before you decide.
What this page says about Kitewell follows its documentation and the app as released. Product names are trademarks of their respective owners.
Something out of date or wrong? Write to [email protected] and it will be corrected.