Day 1 — you do the work
Run the task once, for real, in your own tools. Riplics watches and drafts the skill: “Qualify an inbound lead” — inputs, steps, success test.
Teach AI agents by demonstration
Turn the way you work into AI agents.
Riplics AI watches you do the work once — chasing invoices out of an inbox, qualifying a lead, refilling the same spreadsheet — and turns your demonstration into a skill: the goal, the inputs, what success looks like. Your agents take it from there — on schedule, browser included, API when clicking is the slow way, receipts when you ask.
Agent working
Process incoming invoices
Reading attachments, validating fields, and keeping the ledger current.
Detect email and PDF attachment
Extract invoice fields
Write record and request approval
47
invoices processed
Archived, recorded, and ready for review.
The actual builder
This is the screen, not a mock-up.
A competitor-pricing watcher, open in our own editor. The pills between the steps are real typed values being handed forward — extracted arrives at the dedup step, which passes new_items on to the records table.
yes, the whole thing →The minimap in the corner is the rest of it. Steps are cards, values are pills, and the line between two cards is the actual handoff — that is the whole mental model.
Show. Learn. Delegate.
Train agents by demonstration, not by building workflows.
A recording here isn’t a macro. Riplics reads what you were trying to do — the goal, the inputs, how success is judged — and keeps your click-path as just one way to get there. When the button moves, the agent still knows the job. When the tool has an API, it skips the clicking entirely.
Day 2 — you delegate it
Confirm the card and hand the skill to an assistant. It runs on demand or on schedule, and checks its own result before calling the job done.
Day 10 — twelve skills
Each taught the same way: by doing. Your assistant decides which skill fits, feeds one’s output into the next, and asks before anything irreversible.
Day 30 — an AI workforce
Skills are shared across your team; assistants delegate to each other and keep shared notes. Your processes become their skills.
The unglamorous parts, built in
Dry runsMonthly spend capsApproval before sendDedup across runsAudit trailsCSV export
Popular integrations
Connect the tools your team already uses.
Honestly, “Gmail in, Sheets out” is most workflows’ whole story. The rest are here for the day a run needs to post to Slack or file something in HubSpot.
How it works
A real run, minute by minute.
04:27:03Schedule fires — "EU tender watch"nobody is awake, nobody needs to be
04:27:08Browser opens the tender portal, searches "marketing"
04:27:3119 listings extracted: buyer, contact, procedure, estimated value
04:28:12Same extraction reused for 5 more keywordsone saved workflow, called six times
04:29:44Merged to 25 rows — 3 already seen last week, droppeddedup means nothing emails you twice
04:29:45Quality gate: columns aligned, nothing suspect — cleared to send
04:29:52Styled report + CSV delivered. Every step in the audit trail.
And when the portal restructures — it will — the run diagnoses the drift, repairs its own selectors where it can, and keeps a snapshot of the page it saw so you can see exactly what changed. If a step still can’t be repaired, the agent falls back to the step’s goal: it knows what the click was for, so it finds another way — including going through the tool’s API instead of its screens. Data that looks wrong gets blocked, not sent. Automation demos are easy; it’s the Tuesday 4am runs that are hard, so that’s what we spend our time on.
The platform
The parts you’d otherwise end up building yourself.
Workflows call workflows
The tender watch above calls one saved extraction six times with six keywords. Outputs come back namespaced, so keyword four’s rows never bleed into keyword five’s.
Starts when the work starts
A schedule, a webhook, or the moment an email lands. By the time your first step runs, the PDF invoice from the attachment is already fields.
Records that outlive the run
Runs write into tables — leads, invoices, tenders — with provenance on every row. CSV export for the day your accountant asks.
Nothing sends itself
Gate a step and the run pauses with the real email, the real payment, the real post on screen. Approve, reject, or let it expire — the audit trail remembers which you chose.
Spend has a ceiling
Dry-run before anything touches the real world. Cap a month at $20 and a runaway loop stops at $20.
Your team, same picture
Workflows, records, files, and durable knowledge are shared per team, with roles. When something ran, what it read, what it sent — anyone can trace it, six months later.
Usage-based pricing
Pay for the work that actually runs.
Every charge lands in one audit trail. Set a monthly spend cap, inspect each run, and see AI, browser, and integration usage separately.
AI models
Model cost × 3.33
Token usage is priced from the model’s underlying cost, so smaller models stay inexpensive and larger models remain transparent.
Browser compute
$0.002 / minute
Metered while a cloud browser is running. The minimum browser charge is $0.02 per run, including short jobs.
Platform integrations
$0.01 / credit
Paid integration tools use 1–20 credits per call depending on the operation. Free lookup tools remain free.
$5 in starter credit is included with a new account.No surprise bundles: your balance, lifetime spend, service breakdown, and individual charges stay visible in Credits & Usage.Start building
Pick the process you’re most tired of.
Riplics AI is invitation-only while it’s young — we onboard teams one at a time and read every mail ourselves. Tell us what you want to stop doing by hand.