AC
task
JP
fyi
RS
later
Reads the request
A forwarded thread from AP. It pulls out the goal, the deadline and both attachments, then writes a plan you can edit.
Private beta · Cohort 01 opening
Give it a job.
Not a prompt.
thinkling is a general agent that operates a real computer — browser, spreadsheets, email, internal tools, the terminal — to finish end-to-end knowledge work. You describe the outcome. It does the hours, shows every step, and stops when something needs a human.
No card. No demo call. We onboard in the order you sign up.
0tasks run per week in closed alpha
0median unattended run length
0runs finished without a handoff
0apps reachable out of the box
Live demo · simulated
This is the part that usually
This is the part that usually
takes you all afternoon.
One task, four surfaces. thinkling reads the brief, opens what it needs, and works until the job is done — while narrating every action so you can audit it later.
supplier-portal.acme-vendors.com/onboarding/new
tab 3 / 5
Vendor onboarding
Step 2 of 4 · Company details
Submit & continue
Vendor 9 of 14 submitted
Q3_reconciliation.xlsx
=XLOOKUP(A7, bank!$A:$A, bank!$D:$D, "MISSING")
| Invoice | Vendor | Ledger | Bank | Δ | Status |
|---|
rows 0/218
matched 0
flagged 0
recalc 0.4s
Q3-variance-memo.docx
● drafting
0 words
Q3 payables variance — summary for review
● live
tool: browser.click
ctx 41.2k
sandbox: isolated
Illustrative reconstruction of an alpha run. Every action above maps to a real tool call in the trace.
One agent · your whole stack
A task doesn't live in one app.
A task doesn't live in one app.
Neither does thinkling.
One run. Five surfaces. Fourteen minutes. The context carries across all of them — you never re-explain anything at the handoff, because there is no handoff.
Northbend LLC
88-•••4021
Net 45
ap@north…
Logs in and clicks
Fourteen vendor records re-submitted through a four-step form on a portal that has never had an API.
| Invoice | Ledger | Δ |
|---|---|---|
| INV-4471 | 12,400.00 | 0.00 |
| INV-4473 | 28,015.00 | 1,000.00 |
| INV-4474 | 7,240.10 | 0.00 |
| INV-4475 | 15,880.00 | missing |
| INV-4476 | 2,145.75 | 0.00 |
Works inside the file
218 invoices matched against the bank export with real formulas. Six deltas surfaced, nothing quietly rounded away.
open12 → 2
auto-cleared10
Reaches what nobody exposed
The legacy ERP is just another window — screen, keyboard, mouse. Ten exceptions cleared, two deliberately held back.
Q3 payables variance
14 vendors · 218 invoices
2 need approval
reviewHands back finished work
A memo with the exceptions on top, artifacts filed where they belong, and a full replay sitting behind all of it.
one task · 5 surfaces · 0 integrations built · 1 human review
total 14m 08s — the same job took a person most of a Thursday
Capabilities
Built to work like a colleague,
Built to work like a colleague,
not like a chat box.
Most agents can answer. thinkling can act — on a real machine, over long horizons, with the judgement to stop and ask when the stakes are high.
Computer use
A full desktop in an isolated VM. It moves the mouse, types, opens native apps, drags files between windows and reads what's actually on screen — including the internal tool nobody ever built an API for.
Screen · keyboard · mouse · filesystem
Browser that finishes forms
Logs in, handles multi-step flows, uploads documents, survives a dropdown that only renders on hover, and pauses for you at MFA instead of guessing. Sessions are scoped, credentialed, and revocable.
Auth-aware · multi-tab · human-in-the-loop
Documents & data
Spreadsheets with real formulas, decks that follow your template, contracts read line by line. It produces the file, not a description of the file.
xlsx · docx · pptx · pdf · csv
Long-horizon planning
Runs that last hours, not turns. It decomposes the goal, keeps state across restarts, retries what failed and reports what it couldn't do.
Durable state · resumable · self-correcting
Verifiable by design
Every run leaves a replayable trace: screenshots, tool calls, diffs, the reasoning behind each decision. Approve, roll back, or hand a section to a teammate.
Trace · replay · rollback · approvals
Memory that compounds
The second time you ask for the month-end close, it already knows your chart of accounts, which vendor always sends a broken PDF, and that Rina signs off on anything over $10k. Team memory is shared, inspectable, and editable — you can delete a fact and it stays deleted.
vendor quirks
approval chains
file conventions
tone of voice
tool logins
what went wrong last time
How it works
Three moves. That's the whole product.
Describe the outcome
Plain language, a Slack thread, a forwarded email, or a recurring schedule. No workflow builder, no nodes, no if/else canvas to maintain.
It plans, then works
thinkling drafts a plan you can edit, spins up a clean workspace, and executes — asking only when it hits something irreversible or ambiguous.
You review the trace
Deliverables land where you already work. The full replay sits behind them, so review takes minutes instead of redoing the task yourself.
Evals
Measured on work, not trivia.
We score on task suites drawn from real back-office workloads — each one graded on whether the artifact a human asked for actually got produced, correctly, unattended.
Multi-step web tasksportal forms, filings, procurement
87.4%
Spreadsheet reconciliation218-row ledger vs. bank export
93.1%
Document generationmemo, deck and contract redlines
81.0%
Native desktop appslegacy ERP, no API available
72.6%
6-hour unattended runscompleted without human rescue
68.9%
thinkling preview build
internal suite v3 · n=1,200 tasks · figures are illustrative placeholders
Reach
Connected where there's an API.
Connected where there's an API.
Hands-on where there isn't.
Native connectors for the tools with good APIs — and a real browser and desktop for the ninety percent of enterprise software that doesn't.
Plus anything reachable through a browser, an SSH session, or a screen. If a person can do it on a laptop, it's in scope.
Trust
An agent with your credentials
An agent with your credentials
needs to earn them.
Scoped credentials
Per-task grants with expiry. It never holds a standing key to everything, and every use is logged against the run.
Isolated workspaces
Each run gets a fresh VM that is destroyed when it ends. No shared state, no residue between customers.
Your data stays yours
No training on customer data, ever. Configurable retention down to zero, with regional processing on request.
Stop conditions you set
Spend caps, blocked domains, and mandatory approval on anything irreversible — payments, deletions, sends.
SOC 2 Type II — in progressGDPR-ready DPASSO / SCIM on enterpriseSelf-host available for design partners
Questions
The ones you were about to email us.
What actually is thinkling?
A general-purpose agent with a computer. You hand it a job that would take a capable person a few hours — reconciling accounts, filing across portals, chasing down data, assembling a report — and it does the job on a real machine, then hands back the artifact plus a full trace of how it got there.
How is this different from an AI chat assistant?
A chat assistant produces text. thinkling produces outcomes: a submitted form, a corrected spreadsheet, a filed expense report, a signed-off memo. The interface matters less than the fact that something in your systems is genuinely different afterwards.
Is it going to do something irreversible?
Not without you. Payments, deletions, outbound sends and anything you mark as sensitive require explicit approval. It also stops on its own when confidence drops, rather than improvising — a run that pauses and asks is a feature, not a failure.
Which models does it run on?
Frontier models for planning and judgement, smaller specialised ones for perception, extraction and repetitive steps. The routing is ours to tune; you see the cost, the latency and the outcome, not the plumbing.
What does it cost?
Pricing is per completed task with a monthly floor, so you pay for finished work rather than tokens burned. Design partners in the first cohort get preview pricing locked for a year. Exact numbers go out with your invite.
When do I get access?
We're onboarding in signup order, in small batches, so each new team gets real attention during setup. Tell us what you'd point it at first — the closer that is to what we're already good at, the sooner you'll hear from us.
Contact
Who's behind it,
Who's behind it,
and how to reach us.
Company
Fast Control
The company building thinkling.
Based in
Paris, France
Working across European and US hours.
dat@thinkling.dev
Questions, partnerships, or anything else.
Cohort 01 · 72 seats left
Stop doing the
boring half of your job.
Join the waitlist. We onboard in order, in small batches, and we'll ask you exactly one question when it's your turn: what should it do first?
You're on the list.
We'll email you when your batch opens. Nothing else — no newsletter, no drip.
position · calculating…