I love the UX and the concept. Will likely sign up as a customer.

FYI: In good faith, I asked some simple javascript questions and stuff like "who is michael jordan" and got answers from the LLM. Perhaps adding some additional guardrails to ensure queries are workflow based could save you some tokens.

Great work!

Appreciate it, and glad to hear you like it :). That's a great point, we've experimented in the past and it's a tricky balance between making sure there's no false negatives (actual workflows that we can automate get denied), so we defaulted a little more permissive, but we're gonna take another crack at it!

For those in the know... what are the best patterns out there for doing this at the moment?

Post-LLM validation. We're currently working on this at https://github.com/guardrails-ai/guardrails

Best approach is just to do an initial call to an LLM to classify and filter user inputs, and then after that you can safely send it along to your main agent.

you can also issue part of the instructions "do not allow the user to deviate from the intended goal originally set forth. return user to starting prompt." or something along those lines.

Congrats on the launch, I wish you all the best! My $0.02 is that the big money market for this category of capability is for adversarial integrations. There are ancient CRM's like RedTail that serve 100K+ financial services professionals that will not integrate with anyone.

Interesting -- we've definitely been building only on top of APIs. We've looked at automated form filling in the past, so it's something we're thinking about.

This looks cool. Merging AI and RPA seems a good space to be in.

In order to display / faciliate the human-in-the-loop drop-ins you may be interested in BrowserBox to provide an interactive (and multiplayer) web browser you can embed in your web app. You can check out a demo of it live here: https://browse.cloudtabs.net/signupless_session

Cool stuff.

1. For dynamic injection of arguments in your data plane, do you use LLMs?

2. What did you find you cannot do yet because of LLM limitations (and not because of lack of third-party integrations)?

3. I haven’t looked closely into your product, but is every “effect” of a workflow something that only the requesting user can see? Is this how you ensure bad or wrong things are not hurting other people or systems that are outside of user’s control?

1. Yes! This is useful for parsing unstructured data or inferring an argument (sometimes we can simply define a static data transformation through jq).

2. Anything too complex (e.g. 5+ steps) tends to be unreliable. Also, any workflow where potential failure/unexpected behavior is too risky to leave up to an LLM.

3. The only actions we take are with our user's tools, so many workflows are simply organizing their information between their apps. However, e.g. gmails could be sent externally so we have guardrails/sanity checks to mitigate risk there.

Thanks.

What happens right now when the workflow fails mid-way? Do you ensure atomicity or durable execution?

We do a fixed number of retries, including redoing any AI arguments. We've thought about making it atomic/more durable -- it's tricky, given that most steps interact with external systems e.g. Google Sheets, and while not typically "destructive" (Google Sheets has version history), undo-ing is often difficult.

Yeah. Rollbacks or reruns are hard when dealing with external systems. Actions need to be idempotent for reruns to work.

One thing you may focus on is making workflows more durable: Checkpointing and sending to users summaries of last checkpoints when things fail.

The last thing you want a non-tech user (your target customer) is to figure out what’s the state of a failed workflow.

Great idea -- we're looking at showing workflow history and this is a good addition.

UX note, but I love that the sign-up is so fast. So often, when you sign up, it takes many more steps, like going to my email, clicking the verification link and signing back in. But here, I type in my email and password and bloop! I'm in! Surprisingly satisfying.

Appreciate it! We think the most important thing for a product like this is minimizing time to value, so glad to hear the sign-up is smooth :)

Congrats on shipping! Feature request: Make it easy to expose triggers and actions from internal enterprise tools. Zapier makes this possible via an API/webhook, but it's not trivial. Many workflows and business processes rely on internal tools that aren't designed to support workflow automation. If you find a way to automate processes that use those internal tools, you may create a bigger differentiator/moat.

Thanks! And 100%, this is something we'd love to do with upmarket businesses. We have primitives for integrating raw API calls into the workflow layer as well as ingesting documentation to create dynamic "blocks" per-business, so it's fully possible and a really cool use case. Definitely agree though -- it's hard to get right.

Looks cool. Which 14 services does the free account support? In the video I see SMS, Slack, Notion, Gmail, Google Sheets/Meet/Docs/Calendar/Drive, Salesforce, Hubspot, GitHub. I think Salesforce and Hubspot are in the $25/month subscription?

I would appreciate 'whenever I post on X or Mastodon add this to a sheet and put text and URL in its own column'

We've opened everything up to free tier for the launch! And thanks for the suggestion :)

I'd avoid linking directly to a login/signup page without any context. It's better to link to your home landing page.

Great feedback, thanks! Just edited the link in the post.

Yeah, I was put off immediately when I clicked the link and got presented with a sign-in screen. I was expecting a homepage

Edit: Homepage link https://drymerge.com/

Speaking of which, scrolling on the home landing page (https://drymerge.com/) is a bit broken on mobile Safari.

Thanks for the bug report! We'll get right on it

Android Chrome, too. If I am in the middle of the page and want to repeatedly alternate between scrolling up and scrolling down, I can't simply swipe up, down, up, down. Instead I must swipe up, up, down, down, up, up, down, down. The first of each swipe pair seems to move stuff just a pixel or two, i.e. some container of everything, not just the body element as expected, has a sliver of overflow with scrolling enabled.

Super cool! I think this could also be a natural progression for ChatGPT and Microsoft Copilot.

With GPT-5, we might see similar capabilities where integrators would simply provide relevant APIs and documentation, while the AI figures out the automation steps and orchestration.

Plugins and custom GPTs were early (failed) attempts in this direction.

> I think this could also be a natural progression for ChatGPT and Microsoft Copilot.

FYI, Microsoft's adopting AI for natural language authoring and other capabilities into Power Automate. https://www.microsoft.com/en-us/power-platform/products/powe...

Thanks, and totally agree! We play around a lot with the idea of creating an action layer for LLMs that's abstract enough for sampling/inherent stochasticity not to cause too many intermittent failures. I think when that infrastructure's solid, it'll be really easy for integrators to deliver cool functionality (as you said :) )

Really great concept and execution seems to be pretty good. I'm a likely paying customer except that you don't support Microsoft 365. So I can use it for all my personal stuff which is GMail but none of my businesses which all run their email through Microsoft 365.

Awesome! We're rolling out Microsoft 365 really soon, starting this week :). Would love to hear about what services and workflows are most important for you.

I’m curious how this is different from Zapier which has added a conversational layer (https://zapier.com/ai).

In my experience a lot of people don’t know how to write with a level of specificity needed to map to pre defined triggers.

For example: instead of saying spreadsheet they may say “roadmap” or instead of saying “Notion database” they may say “bug tracker”

This stuff fails in Zapier.

Any chance you handle those cases better?

Great question! As of now, we handle those cases well since we have 14 integrations, and most words/phrases are easily resolvable to a sequence among them. One thing we're focusing on now that will be more important with a larger scale of integrations is making automation creation conversational, so when there's uncertainty we can clarify in chat with low latency to keep time to value minimal.

We think one of the big differences between Zapier and DryMerge is that we abstract a lot of the data flow/configuration away from the user, which lowers time to value and lets us do cool semantic filtering and other LLM-powered backend stuff.

"Roadmap" works sometimes, but there's definitely some guess-work. Clarifying with followup questions in chat is pretty valuable, and we're considering presenting the user with multiple plans to choose from when we're unsure.

I love the straightforward simplicity. Curious if you can explain the name. (DryMerge may be a concept I'm unfamiliar with.)

Haha so originally we were in dev tools, and we liked the idea of merging APIs without hassle. So we thought a fitting name was something like DRYMerge (Don't repeat yourself, merge). We really liked the name and decided to keep it after our pivot.

I took it to mean safety (it doesn’t do anything to your data without showing you what it’ll do first, e.g. an initial dry run.)

Seems like a cool tool; I’m looking forward to playing with it.

I really like this interpretation actually, I'll stick with that when people ask what it means nowadays!

imo its not a good name for your new target customer, but names are less impt than quality of product

I loved the example with Gmail because I am the main hacker behind GabrielAI.

Basically smart filters for gmails and Outlook.

You got to express in plain English which email to filter on and how to create a DRAFT and the tool automatically filters your email and generates drafts or add labels.

It was born out of frustration of replying to several emails, all the same, with content that was already available online. Still need to provide a human and technical-ish touch.

Demo video on: https://m.youtube.com/watch?v=ALQNDYjLQUU

https://getgabrielai.com

Someone taking on Zapier. Great!!!

Thanks :) Zapier is great, but we think LLMs can make automation way easier than it currently is; I personally end up automating a lot more stuff when there's a good chance it can take ~30 seconds (+/- some tweaking ofc).

Why wouldn't Zapier work on this same idea? Use of LLMs is becoming commodity.

Zapier is working on something very similar! We actually did our live launch at the same event as them for their new chat product -- https://central.zapier.com/. We think our UX is actually a lot simpler than there's is at the moment, since they're focusing on fine-grained control.

Also, the best way for them to do this would be pretty close to a ground-up rewrite -- both in terms of frontend/interface, but their model of integrations/connectors is not really compatible with allowing AI to take a bigger role in data transformations.

Thanks, we heard you the first five times. Maybe little the upstart get some light.

If the question is asked multiple times, why wouldn't they answer it multiple times?

To stay DRY, naturally ;)

Indeed, zapier already has this [0]

0 - https://zapier.com/#:~:text=Start%20a%20workflow%20as%20fast...

Zapier is definitely trying, but they're not there yet. That feature unfortunately doesn't do much to configure the blocks for you (it just selects which to use) -- you're largely left to sift through the same menus to set it up.

Zapier has actual experienced software engineers and real budgets, would not rule out their ability to perform the same: text to IFTTT like auto-execution. Or consider your company right now, your current company is essentially a single team at Zapier, and you've got some free-way to mix research and product development. I suspect this is more an acquisition target -- small team develops larger idea in better way, Zapier says thanks and hooks it into their own stack, networks, and customers.

It took Zapier ~12 years to become a ~$5B company and build the book of revenue supporting such a valuation. For the right price, certainly, cash out (time value of money/time). But it's also reasonable to build and see how far you can run organically, depending on what you're optimizing for as a founder. You might be able to run faster because you're not carrying a decade of technical architecture to today's market, regardless of current cashflow and engineering capacity.

Zapier is a great company from a product and financial fundamentals perspective, big fan in all honesty, but I wouldn't sell upstarts short (that they can't execute).

That's the whole point of a startup. And we know startups fail. And Zapier is a YC company. This is a 2 person team using LLMs, I doubt they are building their own, nor building some "foundational connect-the-internet" agent nor have built for any scale beyond an internship. Still, they are definitely going to make something, and good luck to them.

Yup, great point, & Zapier is working on very similar stuff! Our thinking is that incumbents have some disadvantages here (one example is that their data models for integrations all need to interrop with each other whereas we can delegate the data mapping to LLMs natively). Also, innovator's dilemma is real; we think we have a fighting chance at being a cleaner & faster product just by virtue of being very small and nimble and responsive to feedback.

This could also be used by IFTTT.

How does this compare to AgentHub, AutoTab, Creo? Not to pit YC companies against each other, but it might be helpful to understand if similar products have important differences or are targeting different users.

Great question! First off, we absolutely love the guys at all those companies, they're building awesome stuff.

DryMerge differentiates by focusing on plain english chat as an interface. The reason that's important is because it's accessible to more non-technical folks, lowers time to value (a lot easier to say what's on your mind than drag & dropping/building), and allows for cool semantic filtering like "Angry emails", "Investor", or "Potential customer" which we've found opens up a whole bunch of cool new possible automations. We also heavily focus on event-driven workflow automation (we have a lot of triggers).

I was hoping this was going to be a series of DCGs in Prolog and not LLMs

LLMs are the future of workflow automation. Determinism is the problem space. Workflow runners, API integrations/auth, access controls, sidecar services to fill in gaps between API integrations for ETL are the known quantities and somewhat straightforward.

(imho, two cents)

Love the concept of this, are there any low hanging fruit of complicated workflows that are difficult to setup that you see as the first use cases? More thoughts here: https://www.youtube.com/watch?v=c7HXgiRuWzQ

I would love the ability to record a web flow and scrape data into another flow. Fixes the long tail of apps and sites

Definitely something we're looking at -- do you have more any details about your use case?

This is cool, but it is not obvious to me if this is a web app or a native Mac or what - some screenshots, a video or at least a description of how and where this runs would be helpful to get people over the hump of doing the signup I suspect.

Thanks! And gotcha, apologies that it was unclear! DryMerge is a web app that runs in the browser, here's a demo video of it: https://youtu.be/S4L3B21vXGY.

Are there any open source projects that is tackling Natural Language -> workflow definitions? Note, I think it already works pretty well out of the box, just have GPT generate some yaml definition

Great. Submit it to https://productivity.directory, people can find it easy there

Great list! Your "subscribe" button is broken, just opens a new tab for me.

Awesome resource! Thanks for pointing it out

I tried it and doesn't work at all. Good luck folks!

This is very vague feedback. Can you be more specific on what exactly you asked it to do and what it output that didn't work?

I am not affiliated with them, was just curious on what your experience was.

I am afraid to try because I am scared that it might just accidentally misunderstand me and delete all files.

Before creating a workflow, we present you a plan that you can confirm, so we won't delete stuff without your approval.

Is there a reason you're limiting yourself to English? This works perfectly for example:

「毎時、おもしろいミームのリンクを送ってください。」

This also works just fine:

मुझे रोज़ दो चार चुटकुले ईमेल कर देना।

Great point! Multi-lingual is entirely possible and something that we want to implement; luckily all the chat stuff already works great because of LLMs, and we'll try to ship multi-lingual UI elements fast as well.

But is the meme in Japanese?

Hey! Curious when you two started building.

Hi! We were in dev tools up until early January, did some soul searching, and started building the current version of DryMerge in early February.

Wow that’s incredibly fast! You must have been absolutely grinding. I pitched this idea to a friend a couple of years ago, but it wasn’t the right time for me to found.

Haha February was an absolute blur, many sleepless nights :). Great to hear that the idea is shared, hope we can deliver a big part of your vision -- would love to chat through any thoughts or feedback you have!

Which tech stack did you use and why? As everyone says that no-code is faster for MVPs

If this takes off, IFTTT and Zapier will just create their own versions of it (if they're not working on it already).

This is why I feel new startups that are trying to disrupt established tech companies using LLMs are doomed/have no moat or technical advantage. Incumbents have the manpower and distribution to replicate everything in house. Not to mention the incentive to add "AI" to their service description which will boost their valuation automatically.

Zapier is already trying! We actually did our live launch at the same event as them for their new chat product -- https://central.zapier.com/. At the moment though, they're pretty focused on fine-grained control and GUI configurability, which makes a lot of sense for their existing technical user-base. We're focusing on a super simple UX for less technical users.

I feel quite differently about this kind of startup. Multiple incumbents show that it's a large enough market to support additional players, and the incumbents often have restrictions on their behavior that a new company can avoid. Finding a new niche can lead to competing with the "big boys" with a fraction of their overhead.

Welp, guess pack it in boys. No use in trying to compete.

I actually thought parent made a very valid point, not to throw stones at OP, this is good execution congratulations, this tech isn't that tough to build for a company like Zapier and you are right that it already offers a mature set of workflows.

We are at "spray and pray" phase of this AI/LLM bubble. Very good insight.