Tejas Kumar

I Should Join Anthropic

This post is about the day Anthropic launched Claude Duet with the exact words on my startup’s homepage and why I’ve come out of it thinking I should join them.

On Tuesday, 21 October, I boarded a plane at 20:23 Berlin time to go speak at CityJS Athens. 46 minutes later Anthropic announced Claude Duet and I saw it the next day while scrolling X. 45 seconds into the launch video, the screen says this:

A lavender screen from the Claude Duet launch video with the words “You keep the final say” in black serif type

That’s my line!! It’s the headline on daiku.ai, the startup I’ve been building since 4 October, and it has been since 7 October:

The daiku.ai homepage with the headline “Your AI does the clicking. You keep the final say.” above a paragraph saying Daiku is an add-on for Claude, ChatGPT or any AI you use

I said “they copied me” out loud, then thought something a lot darker (more on that below).

What is Claude Duet?

Claude Duet is a new app from Anthropic where Claude works as a proactive agent: it watches your connected apps for what needs you, does the tasks you give it in its own browser on a virtual machine while you’re away, and asks you before it sends, pays, signs or deletes anything. It runs on Claude Opus 5.5 and you can’t pick another model.

Right now it’s a research preview on iOS, Android and claude.ai, by waitlist, only for people on the Pro and Max plans in the US (Anthropic’s help page has the details). Anthropic’s post announcing it on X had over 9 million views 2 days later, which for an app you can only get from a US waitlist is wild. I’m on the waitlist now (I signed up from Germany so I’ll probably be waiting a while).

Daiku said “You keep the final say” 14 days before Claude Duet

The Daiku repo has the receipt: commit e6385883, at 01:24:08 on 7 October (Berlin time), added this line to the brand book in BRAND.md:

Hero line: Your AI does the clicking. You keep the final say. The landing hero since 2026-10-07, when the owner asked for a site built to convert.

The owner is me. A git date is easy to fake. The Wayback Machine isn’t: it saved daiku.ai on 10 October with “Your AI does the clicking. You keep the final say.” as its headline, 11 days before Duet.

Daiku and Claude Duet in October 2026 (Daiku's git history, the Wayback Machine, my Gmail, Anthropic)
Daiku: 789 commits
Oct 4 to Oct 23: 19 days
You keep the final say, on daiku.ai
Oct 7 to Oct 23: 16 days
Daiku tells you when something needs you
Oct 18 to Oct 23: 5 days
Cloudflare for Startups gives Daiku $10,000 in credits
Oct 21
Anthropic announces Claude Duet
Oct 21

Look at the 21st: at 10:42 that morning an email from Cloudflare for Startups said Daiku had $10,000 in credits for the next 12 months. And at 21:09 that same night, Anthropic shipped Duet!!

Did Anthropic copy me?

It sure looks like it: the same 5 words on a full screen of their launch video, 14 days after they went up as my headline. But there’s a wild twist. I wrote Daiku’s homepage copy with DeepSeek V4.1 Flash in OpenCode. That’s where “You keep the final say” came from.

Anthropic, meanwhile, says it “designed it around Claude Opus 5.5”, and Karan Ahuja, who leads Duet, says he’s “been using Claude Duet to build Claude Duet”. So Duet was made with Claude and my line was made with DeepSeek.

To be fair to Anthropic, asking before an agent acts was their idea before it was mine. Claude in Chrome has asked “before taking high-risk actions like publishing, purchasing, or sharing personal data” since August 2025. And Anthropic wrote about keeping people in control of agents in April. The idea is old. It’s the exact 5 words that got me.

Did DeepSeek distill Claude?

Anthropic says DeepSeek distilled Claude (distillation means training your own model on another model’s answers). In February, in Detecting and preventing distillation attacks, Anthropic said DeepSeek, Moonshot and MiniMax had over 16 million exchanges with Claude through about 24,000 fraudulent accounts, over 150,000 of them DeepSeek’s.

In September, Anthropic’s threat report went further. It says “DeepSeek rerouted requests from users that were attempting to use one of DeepSeek’s models through third-party or Anthropic coding harnesses, like Claude Code, the Claude Agent SDK, or OpenCode” and “Selected tagged users then had their requests relayed to Claude Opus.” That’s over 12.1 million exchanges in 14 days in July.

Wait, what?! Anthropic is saying some people who typed into OpenCode thinking they were talking to DeepSeek were talking to Claude Opus. It even says their data “was likely routed to Anthropic without the knowledge or consent of DeepSeek’s customers”.

I typed into OpenCode thinking I was talking to DeepSeek!! I couldn’t find DeepSeek answering any of this and I have no way of knowing whether my own DeepSeek calls ever came back from Claude. But if they did, maybe Anthropic didn’t copy me at all. Maybe Claude wrote my line through DeepSeek, then wrote Duet’s line for Anthropic, and said the same 5 words twice, 14 days apart.

And honestly, that scares me more than being copied. When every startup and every lab writes with models that might all be Claude underneath, we all end up with the same homepage, which is terrifying… who’s the original then?

“My dream is dead”

Back in Athens, I read the Duet announcement and my darkest thought went something like damn, they’re gonna eat me. I don’t fucking matter.

That thought is an old one for me. I was born with severe hemophilia A and by the time I was 8, the working assumption was that I would never be anything. On Wednesday that old thing haunted me again: one guy with a 17 day old startup, against a company that can put out an app and get 9 million views on one post.

7 minutes earlier I’d been telling my friend Faris that Cloudflare had just given Daiku $10,000 in credits and that it was “Crazy that people are taking this seriously”. At 14:00 Athens time I sent the Duet post to the people I tell everything, with 4 words: “My dream is dead”.

Faris Aziz, who runs ZurichJS and had spent that morning getting Daiku to help him review conference talk submissions on his phone, replied “RIP” and then “go work on their team for Duet” (hi Faris, this post is your fault).

Then David Khourshid answered. David made XState, runs Stately, and is one of my closest friends in the industry and in life. I even MARRIED THEM (I officiated his and Jenny’s wedding). Jenny and my wife Lea are close too. 11 minutes after my text, he wrote “Your dream is evolving by force”:

A WhatsApp chat: Tejas shares Claude’s post on X announcing Claude Duet with the words “My dream is dead”. David replies “Your dream is evolving by force”

A dream evolving by force is a story so old it’s in the Bible!! Joseph (yes, the Technicolor Dreamcoat one) got sold into slavery by his own brothers. Years later, with them standing in front of him, he told them “As for you, what you intended against me for evil, God intended for good” (Genesis 50:20). I’m not casting Anthropic as Joseph’s brothers, by the way: I doubt they know I exist.

David had seen this coming. On 7 October, when I first showed him Daiku, he signed up and wrote “This field is going to get really crowded soon though lol”, then “The proactive thing is really good, would love to see your thing do that”. Daiku got notices on 18 October (it tells you the minute something needs you) and 3 days later Anthropic launched “a proactive agent”.

That afternoon I went back to work on why Faris’s review session had stopped after 2 or 3 talks. By the evening I’d found it (Daiku never stopped him: ChatGPT did, after 287 browser steps for 16 ratings) and fixed it.

Why I’m building Daiku

I’ve wanted this since I saw Iron Man in 2008: an AI like J.A.R.V.I.S. that just does the thing on the computer for you. The models can do that now. The products still can’t (I could never get ChatGPT’s agent mode to work for me). On top of that, a full browser in the cloud for every person usually costs too much to sell to normal people.

By the grace of God, I’ve cracked the money part: a typical person on Daiku costs about $4 a month to run and even my own heavy use (54 hours of browser time this month) costs about $9.

But the reason I keep going is people. My friend Isaac is trying to build a business (isaac.fitness) and doesn’t even have a laptop. With Daiku, his AI drives a full desktop Chrome in the cloud for him. The name comes from Japan, where a daikō (代行) is a driver you call to drive you home in your own car: in Daiku, the AI is the driver.

My parents live in India. How cool would it be if they could just say “find the best flight to visit Tejas in Berlin and get the visa” and then it just happens?! Wow!

Then there’s trust. I’d never trust a for-profit non-profit like OpenAI with my logins and my data. So in Daiku, every person’s browser is encrypted with its own key and you sign in with a passkey instead of a password. Every yes is your face or fingerprint, signed over the exact site, action and amount (no AI can fake one of those). Your browser and its log are stored in the European Union (EU), wherever you live.

Daiku is what I want for me, my family and my friends: maximum safety and privacy, built in Europe. I wanted it and it didn’t exist.

In 19 days I’ve made 789 commits to Daiku (how I build it is in The 5 Levels of Agentic Coding). 33 people from 11 countries have joined the waitlist and I’ve let 18 of them in. 9 people besides me have had their AI drive their own Daiku browser. Across everyone, AIs have taken 20,122 actions in Daiku browsers and 96% of them worked. There have been 10 approvals so far (every single one a passkey) and people took the wheel themselves 74 times. Daiku earned $0 this month and cost me $92.60 and nights like 7 October, when the new headline went up at 01:24.

Of the people who connected an AI and said which one, 5 used Claude and 6 used something else:

The AI apps people other than me connected to Daiku, counted by person, 6 to 22 October 2026 (Daiku's analytics)
Claude
5
ChatGPT
4
Didn't say
4
OpenCode
1
Hermes Agent
1

Did Claude Duet kill my startup?

Claude Duet didn’t kill my startup (not yet, anyway).

Getting sherlocked means the platform you build on ships your product as a feature of its own. The word comes from 2002, when Apple shipped Sherlock 3 with Mac OS X 10.2 and it did what Karelia’s Watson app already did. For Claude users in the US on a Pro or Max plan, Duet sherlocks part of what I do.

For everyone else it’s a different thing. TechCrunch compares Duet with other agents like Meta’s Muse and OpenAI’s Dots. Daiku isn’t an agent. As I put it to David on 7 October, “my thing is not an agent, but it lives in your agent and enriches it with your personal context”.

Claude Duet vs Daiku

Claude Duet Daiku
Which AI drives it Claude Opus 5.5 only Claude, ChatGPT, Claude Code, Cursor, Codex, Gemini CLI, OpenCode, or any AI app that adds a connector by link
Who can use it Pro and Max subscribers in the US, by waitlist Invite only, from anywhere, with a browser online from 1 of 11 countries
Where your browser is kept Anthropic’s cloud computer The EU
Can you watch it and take over? No: “You can’t watch Claude Duet work in its browser or take over the browser” Yes, any time
How you say yes An approval in the app, and Anthropic’s “auto mode classifier” approves actions Your passkey: Face ID or a fingerprint, signed over the exact site, action and amount
Does it come to you first? Yes, that’s the point of it Notices, since 18 October. Otherwise it acts when your AI asks
Price Included in Pro and Max From €19 a month on top of the AI you already use

If Anthropic opens Duet up in Europe and lets you watch and take over its browser, most of what I’d tell a Claude user about Daiku goes away. Everyone on ChatGPT, Cursor, Gemini or OpenCode would still need something like it. So would anyone who’d rather have their own fingerprint decide what’s OK than a classifier.

Why I should join Anthropic

I’ve spent 19 days and 789 commits on the problem Duet has to solve: a browser in the cloud that an AI drives while you’re away, signed in to your life, that still asks you first. If I worked on Duet, I’d push for:

  1. Letting you watch it and take the wheel. Duet can’t show you its browser yet. People on Daiku took over theirs 74 times in 11 days so they clearly want to.
  2. A passkey for every yes. A fingerprint signed over the exact action is a yes no model can fake, even on the day the model is wrong.
  3. Duet in Europe. I’d build it from Berlin, for the people around me who can’t get it today.

Anthropic does take on small teams that built what it needed. It hired Humanloop’s team in August 2025 and bought Bun in December. In February, Vercept, a computer use startup, wound down its own product to “join Anthropic in pushing the frontiers of computer use”. Daiku could join with me or not. Either way, it’s what I want for my parents, for Isaac and for me.

But honestly, IBM is the best job I’ve ever had, with the best manager I’ve ever had. Roy Derks is incredible and a long time friend. We have a Betriebsrat (an elected works council with rights under German law). I have maximum safety and security to do the best work of my life and support my family, even start one! I want to become a father soon and this is the ideal environment for that. It would take a lot to get me to leave.

Questions

What is Claude Duet?

Claude Duet is Anthropic's proactive agent app, announced on 21 October 2026. Claude checks your connected apps for what needs your attention, works through tasks in a browser on its own virtual machine when you step away, and waits for your approval before it sends a message, makes a payment, signs a document or deletes anything. It runs only on Claude Opus 5.5, as an app for iOS and Android and on claude.ai.

Is Claude Duet out?

Claude Duet is out as a research preview, but only by waitlist and only for people on Claude's Pro and Max plans in the US. Team and Enterprise plans aren't included yet. Anyone can join the waitlist, but people outside the US, including the European Union (EU), can't use it yet.

Is there an alternative to Claude Duet?

Meta's Muse and OpenAI's Dots are the closest apps from other big AI labs. My startup Daiku is a different kind of alternative: instead of being another agent, it gives Claude, ChatGPT or any AI app that takes a connector its own browser, signed in to your sites and stored in the European Union (EU). It asks for your face or fingerprint before it sends, pays or deletes anything.

Did DeepSeek distill Claude?

Anthropic says it did. In February 2026 Anthropic said DeepSeek had over 150,000 exchanges with Claude through fraudulent accounts. Its September 2026 threat report says DeepSeek relayed some of its own customers' requests, from coding tools like OpenCode and Claude Code, to Claude Opus: over 12.1 million exchanges in 14 days in July. I couldn't find a response from DeepSeek.

What does sherlocked mean?

To get sherlocked is to have the platform you build on release your product as a built-in feature. The term goes back to Apple's Sherlock 3 in Mac OS X 10.2 (2002), which did the same job as the Watson app by Karelia Software.