See it work
Real recordings of Kitewell's chat window, unedited and sped up evenly: one
request in plain words, then everything the AI and the app did.
How it works
- AI builds it. Describe the job in plain words. The assistant lays out
every step for you to check, works the site in a browser you can watch, and
tries the draft without sending anything out. Nothing runs until you apply
it.
- Your computer runs it. On a schedule, from a webhook, or from your
phone. It stops for your OK wherever you put a gate.
- AI maintains it. When a site or app changes, AI finds the new way,
keeps it, and the run shows what it fixed. When a run fails, AI tells you
why.
Browser automation, built by AI and maintained by AI
Tell Kitewell the job in plain words. AI signs in, clicks, types, and collects
in a browser you can watch and take over, and writes the steps from what
worked. Later runs repeat them, and when the site changes, AI finds the new
way and keeps it.
| Step | What the browser did | Result |
|---|
| Open page | 127.0.0.1:4320/billing/invoices | Completed |
| Do | Type [email protected] into the email field | Replayed |
| Do | Type %password% into the password field | Replayed |
| Do | Click Sign in | Recovered |
| Make sure | text "Invoices" | Completed |
| Collect | Every invoice issued since the last weekday | Completed |
From a real run against a stand-in supplier portal, as Kitewell shows it.
The portal had moved its sign-in button, so AI found the new one and kept it;
the next run replayed it.
- Behind any sign-in. Sign-ins stay in Kitewell's own browser profile on
the computer. Sites that ask for a client certificate work too.
- Passwords stay out of the model. A password is named in the step and
typed from the computer's secrets; the model sees only the name.
- It shows what it fixed. Each run marks the steps it repeated and the
ones AI fixed, with screenshots.
type: graph
schedule: "0 9 * * 1-5"
secrets:
- name: PORTAL_PASSWORD
ref: harbor-supply/password
llm:
model: kitewell-agent-computer
steps:
- id: collect
name: Collect new invoices
action: browser.run
with:
url: https://portal.harborsupply.example/billing/invoices
variables:
password: ${env.PORTAL_PASSWORD}
do:
- act: If a sign-in form is shown, type [email protected] into the email field and %password% into the password field, then sign in
- expect: The invoice list is shown
- extract:
instruction: Every invoice issued since the last weekday, with its number, date, item, and amount
schema:
type: object
properties:
invoices:
type: array
items:
type: object
properties:
number: {type: string}
date: {type: string}
item: {type: string}
amount: {type: string}
browser:
profile: harbor-supply
- id: enter
name: Enter them in Ledger
depends: [collect]
action: computer.run
with:
variables:
invoices: ${steps.collect.outputs.invoices}
do:
- launch: {command: open, args: [-a, Ledger]}
- act: For each invoice in %invoices%, add a new bill from Harbor Supply Co. with its number, date, and amount, but do not save
- expect: The new bills are listed with their amounts
- screenshot: before-save
approval:
prompt: Save these bills in Ledger? Check the before-save screenshot.
- id: save
name: Save them in Ledger
depends: [enter]
action: computer.run
with:
do:
- act: Save the new bills
- expect: The bill list shows the new bills as saved
- id: summary
name: Email me a summary
depends: [save]
action: mail.send
with:
mailbox: [email protected]
to: [email protected]
subject: Invoices entered in Ledger
message: |
Entered in Ledger today:
${steps.collect.outputs.invoices}
See this example in full →
How website automation works →
Coding agents: Claude Code, Codex, and 13 more, on your schedule
Put the coding agents you already use to work while you're away, on the
sign-in you already have.
Claude Code · Codex · Cursor · Gemini · GitHub Copilot · Aider · Amp · Cline
· DeepSeek · Droid · Goose · Kiro · OpenCode · Pi · Qwen, and any other CLI
as a custom command.
- Chain them. One writes, another reviews. Run them side by side, or once
per repository from a sheet.
- Keep going at the limit. When an agent hits its usage limit or fails,
the next one you listed takes over.
- Nothing ships unchecked. A gate holds the push until you approve. Send
it back, and the agent runs again with your notes.
- Started by GitHub. An issue or pull request can start the run through a
webhook, with no open ports.
type: graph
schedule: "0 2 * * 1-5"
# Replace this with a clone of your repository that nobody works in, and
# make test with the command that runs your tests.
working_dir: /absolute/path/to/repo
steps:
- id: test
name: Run the tests
run: |
git switch main >&2
git pull --ff-only >&2
if make test >&2; then echo passed; else echo failed; fi
output:
result: {from: stdout}
- id: fix
name: Fix what failed
depends: [test]
preconditions:
- condition: ${steps.test.outputs.result}
expected: failed
action: harness.run
with:
provider: kitewell-agent-claude
fallback:
- provider: kitewell-agent-codex
prompt: |
The test suite on main fails. Create or reset a branch named
fix/nightly-tests from main, run make test, and find the cause. Fix
the code; change a test only when the test itself is wrong, and say
so. Run the tests until they pass, then commit. Do not push. End with
what failed, why, and what you changed.
output: NOTES
approval:
prompt: Push fix/nightly-tests and open a pull request?
- id: push
name: Push and open the PR
depends: [fix]
run: |
git push -u origin fix/nightly-tests --force-with-lease
gh pr create --fill --base main --head fix/nightly-tests
See this example in full →
Each vendor's terms apply. Coding agents in Kitewell →
Beyond the browser
- Desktop apps with no API. A model
reads the screen and works the app. Later runs replay on Windows by the
app's own buttons and fields, and on macOS from the screen.
- Excel, row by row. Runs once
per row of the workbook you already have and writes each result back beside
it. Merged headers, era dates, and full-width digits included.
- Email. Find, read, file, and send
mail in Microsoft 365, Gmail with an app password, or any IMAP mailbox.
- Scripts, servers, and APIs.
Commands, Docker, SSH rollouts a few machines at a time, and any OpenAPI,
with decision steps that choose the next path.
It starts on its own
- Schedules. Daily, on weekdays,
or any schedule, in your time zone. Starts a sleeping computer missed are
caught up when it wakes.
- Webhooks. GitHub, Stripe, Slack,
or any service starts a workflow on your computer through Kitewell's relay.
Signatures are checked, no port is opened, and a request waits up to 7 days.
- From anywhere. From ChatGPT or Claude
on your phone, from Claude Code or Codex over MCP, or with a REST call.
Approve and follow runs from your phone. Know why one failed.
- Approve or send back. Any step can wait for your OK. Approve it, or send
it back with a note, from the run, a notification, or ChatGPT and Claude on
your phone.
- Runs in your browser. Dagu Console
shows each run of a synced project and the computer it ran on. Open it on
your phone. Included in every plan.
- Alerts. Slack, Microsoft Teams, PagerDuty, email, or a desktop
notification when a run fails, waits for you, or misses its schedule. With
Personal and Team.
- When a computer goes quiet. Dagu Console emails you when a computer
stops reporting for 10 minutes: asleep, offline, or Kitewell quit. Turn it
on for computers meant to stay on.
- Failure diagnosis. AI reads a failed run and says whether to try again,
step in, or change the workflow.
- Run record. Every step's output, screenshots, and decisions. Retry from
the step that failed.
- Test first. Run one step with real inputs before you save, and try a
draft without sending anything out.
- Notes that stay. The assistant keeps what it learns in knowledge pages,
for you and whoever takes over.
ChatGPT Plus and Pro can read runs but not approve them.
Run it on every computer. Watch them all from one place.
- Every computer, one view. Dagu
Console lists each computer's status, Kitewell version, and today's runs.
One alert per workflow, however many computers run it, closed by the next
success on any of them.
- Sync and roles. Share projects
with Viewer, Editor, and Admin roles. Each computer keeps its own sign-ins
and secret values.
- Releases. Promote workflows from
Dev to Staging to Prod, with a diff and rollback.
- Robot ledger. Export a ledger of every workflow for audits, one row
each.
- Shared computers (Enterprise). Computers your organization owns, such
as a VM in your own Azure or AWS, run workflows day and night with nobody
signed in. Nothing stops when a laptop sleeps or its owner moves on.
- Build from a browser (Enterprise). Sign in with your company account,
then build and try workflows on the computer they run on while you watch
its browser. Approve fixes from your phone, with no remote desktop.
RPA was the right idea. Building it by hand was the problem.
| Hand-built RPA | Kitewell |
|---|
| A specialist records and scripts every click | AI builds it from your words |
| A changed screen means someone repairs the robot | AI finds the new way and keeps it |
| A robot PC and a license per robot | The computer you have, free to try for 14 days |
| One tool per kind of system | Websites, desktop apps, Excel, and email in one workflow |
| Hard to see what a robot did | Every step, screenshot, and approval in the run record |
How Kitewell compares →
Your computer, your keys, your files
- Secret values are encrypted on the computer, and never synced,
exported, or backed up.
- Your choice of AI: Anthropic, OpenAI, Gemini, OpenRouter, Kitewell's
included model, or a local model that keeps everything inside your network.
- No lock-in. Workflows are plain YAML on the open-source
Dagu engine. Export a project any time.
- Quiet in the background. Runs from the menu bar or tray, with nightly
backups, edit history, and engine updates that roll back if they fail.
- What leaves, exactly. A synced project sends run status and the names of
failed steps, never logs, output, or secret values.
- Six languages. The app speaks English, Japanese, Chinese, Korean,
French, and German.
Everything that leaves, and when →
Examples
All examples →
Pricing
| Trial | Personal | Team |
|---|
| Price | $0 for 14 days, no card | $20 a month (¥3,000 in Japan) | $99 a month per computer (¥14,800 in Japan); a year is ten months' price |
| For | Trying Kitewell on your own work | One person's own work at the computer, paid work included | An organization of up to 25 people |
| Computers | Up to 3 | 1 | 1 to 100 |
| People | Up to 25 | 1 | Up to 25, at no extra cost |
| Projects | Up to 15, with 100 workflows each | Up to 5, with 100 workflows each | Up to 15, with 100 workflows each |
| API keys and connected apps | Any number | Any number | Any number |
| Runs with nobody signed in (Windows) | Included | | Included |
| Runs and alerts in Dagu Console | Included | Included | Included |
| Kitewell's alerts, webhooks, MCP Events | Included | Included | Included |
| Shared projects with roles | Included | | Included |
| Included model credits | 100 once | 500 a month | 500 a month per computer |
| Credit packs | | 1,000 for $12 | 1,000 for $12 |
Prices exclude tax. Every plan needs a Dagu Cloud account. Signing in from
Kitewell starts the trial; after it nothing runs until you subscribe, and
projects stay read-only, to open and export. Personal runs while you are
signed in. An organization that runs Kitewell for shared work, on several
people's computers or on a computer set aside for it, needs Team, which you
subscribe to in Dagu Console. Credit packs
never expire and are used after the month's credits. For more than 25 people
or 100 computers, shared computers, the workspace in the browser, SSO,
audit-log export, closed-network sync, or an SLA, Enterprise is arranged under
a contract; email [email protected] for a
quote. Automation agencies resell Kitewell at $69 a computer a month through
the partner program. AI provider usage is
separate.
Pricing details →
Install
Kitewell runs on macOS 13 or later, on Apple silicon and Intel, and on Windows
10 version 1809 or later and Windows 11, on x64. Website automation uses
Chrome, or Edge on Windows.
Download the installer from the download page
or GitHub Releases, then
follow Your first workflow.
To use Kitewell with nothing to install, open
app.kitewell.com and sign in. Each project made
there runs in Kitewell Cloud, with your
computer closed.
Documentation
Support
Kitewell is made by Descarty, Inc.
About this repository
This repository holds the kitewell.com website, customer documentation,
release notes, and public releases. The application source is not published
here.
Website development
Requires Node 24 and npm. From this checkout:
cd website
npm ci
npm run dev
Before pushing:
npm run check
npm test
npm run build
The website uses Astro for marketing pages and Starlight for documentation.
Docs are Markdown under website/src/content/docs/docs/; the nested docs
segment preserves the public /docs/ route. Published Markdown notes live in
releases/. The build checks internal links, including documentation anchors.
The README's animated loops and the repository's social preview card live in
.github/readme/. The card renders from social-preview.html at 1280×640:
"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" --headless=new \
--hide-scrollbars --window-size=1280,640 --force-device-scale-factor=1 \
--screenshot=.github/readme/social-preview.png \
"file://$PWD/.github/readme/social-preview.html"
Upload it under the repository's Settings → General → Social preview.
