Browser tasks, recorded once.An API, ready to run again.
Describe a task you already do, in a chat. The copilot works through the real page, learns the steps, and turns them into one job your own systems can call. The recorded steps run as they are, without AI thinking them over again, so you get the same steps and the same result every time. There is no code to write or maintain.
- 01Openapp.northwindfreight.com
- 02FillorderId = NW-100423
- 03ClickCheck delivery status
- 04ExtractStatus, carrier, estimated arrival
The console above is a sample. You find the real URL in the app once you publish an automation.
BetaReplays the recorded steps exactlyREST / MCP support
The whole thing in 45 seconds.
How a daily screen task becomes one callable job after you describe it once in a chat. The film replays the work on the admin page and what happens when the job is called.
The screens, code, and figures in the film are samples. Yours will show the sites you use. The film has no sound.
Build it, hand it off, fix it, improve it.
Jinba Browse covers more than the build. It stays with you through the daily use that follows. Walk through the four stages.
Build it just by talking.
- Say what you want done in a chat. No settings screens, no code.
- The copilot operates the real page and puts the steps together as it goes.
- If something is off, fix it in the same conversation.
Describe it once. It learns the steps.
Teaching it takes no settings screens and no programming. You describe the work you do by hand today, in a chat.
Describe the task you already do
Say something like “I want the delivery-status check I do every morning taken care of.” The copilot opens the real page and works through it, and every action it takes is kept as a step. When it needs something only you know, such as a password, it asks you then and there.
Name it and publish it
Give the recorded work a name such as “Check delivery status” and publish it. It is now one automation you can call at any time.
Then just call it
Call it whenever you need it. Pass the part that changes each time, such as an order number, and it runs the recorded steps and answers in seconds.
Let's build “Check delivery status” for real.
A check you used to do by hand, one order at a time every morning, becomes something your system calls after you describe it once. Here it is, replayed as it happens.
- You say it once: “Check the delivery status of order NW-100423.” The copilot enters the order number on the usual page, presses Check delivery status, and reads the status, carrier, and arrival date.
- Those steps are published as one automation named “Track a shipment”. The only thing you pass each time is the order number.
- From then on, your system calls it over the REST API. Pass order number NW-100871 and the same page runs by itself; when the run finishes, you get the status, carrier, and arrival date. No AI runs during the call.
The screens, code, and figures above are samples. Yours will show the sites you use.
Learned once, the same every time.
AI does its thinking while the copilot learns the task. After that, every published run retraces the recorded steps: which page to open, what to enter where, and what to press.
The same result every time
Call it today or next month and it follows the same recorded steps. It will not do something different depending on the day.
More calls, no more AI cost
Recorded steps run without AI. Published calls add no AI model charges.
It never gets it wrong in silence
When the site changes and the steps no longer work, it does not return a plausible answer. It tells you the steps need to be redone.
Compared with handing every call to AI
When AI handles every call
- On every call, AI looks at the page and works out the next action on the spot
- You wait longer for results while it thinks
- The more you call it, the more you pay for AI
- The same request sometimes takes different steps and gets a different answer
- It can miss that the site changed and still return a plausible answer
With Jinba Browse
- The thinking happens once, while the copilot learns the task
- With no thinking time, a run takes only as long as the page takes to load and answer
- Recorded steps run without AI, so calls add no AI charges
- Recorded steps take the same path and give the same answer, whenever you call
- When the steps stop working, it stops and tells you instead of making up an answer
Turn these tasks into APIs
If you do a screen task the same way every day or every week, you can usually hand it over.
Check delivery statuses
Record signing in to a carrier's tracking portal and looking up an order number. From then on, pass an order number and get the status, carrier, and expected arrival date back as JSON.
Get quotes from internal tools
Record signing in to an internal system or partner portal that has no API, entering the conditions, and getting a quote. Other systems can then call it as an API.
Collect billing and usage reports
Record signing in to an admin console, choosing a date range and downloading or viewing a report. Call it through the API when you need it.
Monitor prices and stock
Record pulling prices and stock levels from a competitor's product pages. Every call visits the same pages with the same steps, so your monitoring breaks less often.
How does it compare?
Here is how Jinba Browse compares with the three usual ways to automate browser work.
Scroll sideways to see every column.
| Jinba Browse | Custom scripts | Traditional RPA | AI on every call | |
|---|---|---|---|---|
| How you build it | Describe the task in chat | Inspect the page and write code | Build a sequence in a dedicated tool | Write instructions |
| AI thinks on every call | No | |||
| Same result every time | Same steps each time | |||
| When the target site changes | Detect the change and stop | Can break without notice | Can break without notice | May return a different answer without notice |
| Cost that grows with each call | Run time only | None | Depends on the contract | AI charges grow |
| Best suited to | Websites that require sign-in | Sites with a stable structure | Routine internal business systems | Tasks that need judgment on each run |
Frequently asked questions
Do calls really add no AI cost?
Recorded steps replay without AI, so published calls do not add AI charges. Copilot use while creating or editing an automation can incur AI charges.
What happens when a site changes its layout?
If a recorded step cannot find its target, the run stops with an error. Use Create repair proposal, or record the step again and republish.
Where are sign-in details and API keys stored?
Sign-in details, API keys, and other sensitive values are encrypted when stored and never included in the recorded steps. Values are injected only at run time. The login info screen never shows a value. A run never reads a value back from the page. When a run fails, the page structure and screenshot it saves hide the value of every input field. A value that the site itself shows on the page or puts in a URL is not hidden, and neither is the live screen of a running browser.
Can I record a site that needs a login or form input?
Yes. During recording the copilot pauses at a sign-in form or any field that needs your input, and you type the value into the browser yourself. Passwords are not kept in the recording, and later runs use the value stored as login info.
Can I call an automation without REST?
Yes. A published automation is a REST API you call with an API key in the Authorization header. It also supports MCP (Model Context Protocol), so AI assistants such as Claude Desktop and Cursor can call it too.
Can I build complex automations that combine many steps?
The basic unit is one task, one automation. Steps include repeat, download, upload, dialog, and list extract, and you combine them by talking to the copilot. Start with a single task and add to it as you need.
Build your first automation now
No credit card required. Start by describing one task you already do, in a chat.