Do not assume that companies are willing to put ALL of their intellectual property into your hands. Even if you would not be some startup where any sysadmin could steal and sell my data any time without you even noticing it, you will get hacked just like everyone else that stores interesting data. The data you have access to is absolutely perfect for the global data blackmailing gangs. As soon as you are successful, you will have every black hat hacker and their dog knocking on your doors.
Yep, makes a lot of sense. We architected our system to be easy to self-host & open-source in the future for this very reason, though we decided to launch with hosted because it's easier to improve and iterate.
Understood. Not my startup, but I would have started the other way round.
Businesses that would be willing to pay (a lot) for such a benefit often will be very conservative. In Germany the majority of medium sized businesses using SAP for example still refuse to be moved to SAP's cloud instead of on-premise.
C-Level types typically are not worried putting their email credentials etc into Outlook cloud and getting hacked this way. They are used to "everything is in the cloud". However, as soon as you mention, depending on the type of business "patents", "sales contacts", "production plans" C's will change their mind.
In Germany, where I am originally come from, all of these businesses are worried about their trade secrets to end up in China, and rightly so.
As self-hosting is very complex you could either make good money with consulting (but this means setting up tech teams in all target markets around the globe, using actual competent humans), or by selling it as a plug&play appliance. With that appliance simply being a rack server with a suitable GPU installed.
And again, for your business strategy the long-term risk of pretty much everyone trying to hack you on a daily basis appears too high to me. You might not have on your radar how serious industry spionage is. You will definitely have a fake utility company worker coming into your offices, trying to plug in a USB keylogger into some PC while nobody is looking.
As an example, proven strategy: Find targets internet uplink. Cut it. Customer calls ISP for help. You then send a fake ISP technician that arrives before the real one does. You put a data exfiltration dongle between the modem and the LAN. You then fix the cut outdoor line. Customer is happy that you have fixed it. Later the actual ISP guy arrives. Everyone will be a bit confused that the problem was already fixed, but then agree that it's probably just the ISP once again having screwed up their resource management. Works pretty much every time.
Onyx.app has a self hosted option. I just did the docker setup yesterday. It’s not a great home user option imo but seems like it’s functional for enterprise.
Just had a quick look - while they have that self-hosting option, they still assume you will use a cloud LLM. I started digging because I got confused of them not mentioning any GPU when it comes to resource requirements. There is some documentation on using it fully self-hosted including the LLM, but the emphasis here is on "some".
To be clear: I am looking at this from a CEO perspective, not a "I will play with it in my spare time" nerd one.
controlcore.io was brought to the market for the same exact reason. Not AI Powered, but to control AI and its interactions with your Data, APIs,. Applications etc. And yes, we just give our service as a self-hostable solution. However is the encryption and SOC compliance be, we want our clients to know that none of their internal data or interaction transaction leave their control.
Why do you think Glean and other top company GPT startups who are doing this for longer and have more resources cannot get this level of performance (helpful and accurate)? What makes your approach different and not easy to replicate?
Great question. We are actually surprised more than anything that existing products aren't better. Other companies' algo work DID get disrupted by reasoning models, because newer models don't care about what the top 3 search results are, they're really great at reading hundreds of documents and picking up signal from noise.
I can't speculate exactly on the work that other companies have done, but my guess is we were way more focused on the type of questions people actually ask to each other. We know there's still a lot more you can do to continue to improve it, and so it's up to other folks if they do it too.
Lastly, a few ways we want to be distinguished from all the other offerings are: 1) super easy to setup, 2.) very developer friendly
People can usually tell if an answer isn't helpful, but not always that it isn't accurate. Depending on the context, 85% accurate might not be good enough.
Yep, the other commenter is right--85% is helpful AND accurate. I'd love for you to give it a try and see if 85% is not good enough though. There's always more to push on quality and the more real feedback we get the better we can prioritize what people need.
I've been using Grapevine at my company for the last couple weeks. One of the coolest features is that it proactively answers questions (with citations!). Not everyone thinks to tag the bot but it often surfaces the relevant answer and document and saves everyone some time.
Interesting to see this now launching, when most companies have their own customGPT solution and MCP makes headway towards decoupling the frontend layer.
Data seems to be stored outside of the customers control, so this will be a difficult sell for many companies.
In terms of what businesses we're targeting: we wanted to provide either 1.) a turnkey solution for a company GPT for all the people who don't have it yet, or 2.) a higher quality company GPT for people who do have an internal solution.
Our sense is that ~70% knowledge companies at large still don't have a custom GPT yet, and that of the people who do, our system can be more performant because we're spending more effort than their internal team is. There's a lot of details we've solved on data ingestion and search algo that improved our accuracy dramatically, and things breadth of data connectors is the kind of thing that is expensive for an internal team but worth it if you're providing the service at large.
Not sure what you mean by "data seems to be stored outside of the customers control," but fortunately I think many SaaS apps that were trying to lock down customer data from themselves are walking that back a little bit.
Please look into the Zero Data Retention policies of the subprocessors that you are using. For example, Open AI does not include files as falling under their ZDR [1], thus utilizing OAI as an LLM solution inherently adds unnecessary risk of data exposure that many enterprise clients do no want to onboard. Also, you have to think about those companies obligation to their clients/customers when it comes to data security, along with the risk of IP being exposed to 3rd party systems that they do not have control over, when they make their decisions when it comes to utilizing various business solutions.
The Gather work product is unfortunately going into maintenance mode. We still have a strong team working on the core AV and performance, and the business is very decent and more than supports that team (and we still use the Gather product heavily ourselves).
However, it didn't reach the growth trajectory we needed, so a majority of the company will be working on Grapevine + new products instead.
Ah to be clear, the team is staffed well enough on Gather that the experience will be equal, if not better, than what you've been using already. So we haven't been telling customers to migrate (and unfortunately, there aren't many good alternative products right now).
I've seen Gleam and Onyx, and I think the real problem is that there is a lot of garbage coming in. If you want to solve the problem, you need to find a way to clean the information coming in. And if you've cleaned the information coming in, you have a lower need to have an LLM answer the question.
Our experience, especially with the most recent reasoning models, is that the LLM's are a lot better now at sifting through the garbage. So if you last gave these products a try more than a month ago, I would try them again.
(Additionally, there are a lot of details that do make a big difference in data processing / search algo too, which have taken our own internal accuracy on hard questions from 30% => 80%+)
No self-hosted version yet. You would need to trust us in the same way you trust your other SaaS apps that host company IP (e.g. Slack, Notion, Github, etc.)
I get that's a big ask from a startup. If it helps, we are a company that's been around for 4+ years and have built a work tool (https://gather.town) used for 100k+ people for their daily work, Sequoia-backed, are SOC II certified, and go way beyond that for the security considerations for this product.
YCS19 - so 6 years. Obviously you did well enough to survive. Interesting that the pivot is something so basic after all this time. Kind of interested in the story there.
2020-2022: We built https://gather.town, which during COVID blew up across every use-case possible: conferences, birthday parties, weddings, universities. It was a good business during COVID but eventually started to shrink.
2022-2025: We built Gather for remote workers, which was a long grind into in Audio/Video, performance, and making a game-interface that was good for work but replicated the parts of in-person work people enjoyed. It's a decent business, but didn't match our ambitions with how we wanted to change work for the better.
2025+: We have lots of ideas for how we can make work a lot better with AI. The general theme is, "can we make work as fun as a video game?" Idea being: video games are super similar to work at its core, and AI can both 1.) dramatically change how people need to spend their days, and 2.) help you "game design" someone's work day.
The Grapevine system is the first tablestakes layer people need to have for us to build the products we're excited about. Surprisingly, "company context" was not as good as we thought despite it being such an obvious, big business opportunity. So while I agree it's "basic," it does seem necessary, and is also not the full-scale of what we want to achieve still :)
Is the "ChatGPT" brand name becoming a generic term, like Baid-Aid or Kleenex?
There is the ChatGPT product, operated by OpenAI, Inc, which you can access via their web site or their API. OpenAI does publish gpt-oss as an open-weights model. I suppose you could argue that gpt-oss is "a ChatGPT," though I'd normally think of it as "a large language model." Much like Claude, DeepSeek, Qwen and so on are other large language models.
I can't speak for other people, but our strong opinion about how companies should work is to reduce the massive amount of chores and tedium that exist in our work days today.
With the company GPT, we want to tackle things like: 1.) having to answer a repeated question from a colleague, 2.) answering questions to coworkers that are purely informational, and eventually 3.) things like standup updates, written updates to leadership on status, etc.
I think human interaction at work is one of the most valuable experiences if you're lucky enough to have good colleagues and interesting work. But I think they should almost entirely be around creativity, decision-making, debate, etc. rather than sharing information that exists elsewhere.
Your product is a bot that SPEAKS as an alternative to OTHER PEOPLE talking.
> reduce the massive amount of chores and tedium that exist in our work days today.
See... is that really a strong opinion? Like look what I asked you. How should COMPANIES WORK?
I'm being a little funny about this. I guess my point is that, there are a lot of Grapevines, including all the companies whose technologies you use. Paul Graham invests in all of them, and so he will be fine. But what about YOU? A soft and friendly Enterprise Sales tone... like give me a strong opinion. People at Google and Meta have better sales teams and technology. But they don't have strong opinions. Do you get it now?
Offer self-hosted and I would buy.
Do not assume that companies are willing to put ALL of their intellectual property into your hands. Even if you would not be some startup where any sysadmin could steal and sell my data any time without you even noticing it, you will get hacked just like everyone else that stores interesting data. The data you have access to is absolutely perfect for the global data blackmailing gangs. As soon as you are successful, you will have every black hat hacker and their dog knocking on your doors.
Yep, makes a lot of sense. We architected our system to be easy to self-host & open-source in the future for this very reason, though we decided to launch with hosted because it's easier to improve and iterate.
Understood. Not my startup, but I would have started the other way round.
Businesses that would be willing to pay (a lot) for such a benefit often will be very conservative. In Germany the majority of medium sized businesses using SAP for example still refuse to be moved to SAP's cloud instead of on-premise.
C-Level types typically are not worried putting their email credentials etc into Outlook cloud and getting hacked this way. They are used to "everything is in the cloud". However, as soon as you mention, depending on the type of business "patents", "sales contacts", "production plans" C's will change their mind.
In Germany, where I am originally come from, all of these businesses are worried about their trade secrets to end up in China, and rightly so.
As self-hosting is very complex you could either make good money with consulting (but this means setting up tech teams in all target markets around the globe, using actual competent humans), or by selling it as a plug&play appliance. With that appliance simply being a rack server with a suitable GPU installed.
And again, for your business strategy the long-term risk of pretty much everyone trying to hack you on a daily basis appears too high to me. You might not have on your radar how serious industry spionage is. You will definitely have a fake utility company worker coming into your offices, trying to plug in a USB keylogger into some PC while nobody is looking.
As an example, proven strategy: Find targets internet uplink. Cut it. Customer calls ISP for help. You then send a fake ISP technician that arrives before the real one does. You put a data exfiltration dongle between the modem and the LAN. You then fix the cut outdoor line. Customer is happy that you have fixed it. Later the actual ISP guy arrives. Everyone will be a bit confused that the problem was already fixed, but then agree that it's probably just the ISP once again having screwed up their resource management. Works pretty much every time.
Onyx.app has a self hosted option. I just did the docker setup yesterday. It’s not a great home user option imo but seems like it’s functional for enterprise.
Just had a quick look - while they have that self-hosting option, they still assume you will use a cloud LLM. I started digging because I got confused of them not mentioning any GPU when it comes to resource requirements. There is some documentation on using it fully self-hosted including the LLM, but the emphasis here is on "some".
To be clear: I am looking at this from a CEO perspective, not a "I will play with it in my spare time" nerd one.
controlcore.io was brought to the market for the same exact reason. Not AI Powered, but to control AI and its interactions with your Data, APIs,. Applications etc. And yes, we just give our service as a self-hostable solution. However is the encryption and SOC compliance be, we want our clients to know that none of their internal data or interaction transaction leave their control.
Why do you think Glean and other top company GPT startups who are doing this for longer and have more resources cannot get this level of performance (helpful and accurate)? What makes your approach different and not easy to replicate?
Great question. We are actually surprised more than anything that existing products aren't better. Other companies' algo work DID get disrupted by reasoning models, because newer models don't care about what the top 3 search results are, they're really great at reading hundreds of documents and picking up signal from noise.
I can't speculate exactly on the work that other companies have done, but my guess is we were way more focused on the type of questions people actually ask to each other. We know there's still a lot more you can do to continue to improve it, and so it's up to other folks if they do it too.
Lastly, a few ways we want to be distinguished from all the other offerings are: 1) super easy to setup, 2.) very developer friendly
">85% of answers are helpful & accurate"
People can usually tell if an answer isn't helpful, but not always that it isn't accurate. Depending on the context, 85% accurate might not be good enough.
Yep, the other commenter is right--85% is helpful AND accurate. I'd love for you to give it a try and see if 85% is not good enough though. There's always more to push on quality and the more real feedback we get the better we can prioritize what people need.
95% of answers could be accurate. Combined with the 85% that are helpful and you have “85% of answers are helpful & accurate.”
I've been using Grapevine at my company for the last couple weeks. One of the coolest features is that it proactively answers questions (with citations!). Not everyone thinks to tag the bot but it often surfaces the relevant answer and document and saves everyone some time.
Thanks for using it! Any feedback on what could be made better? And did you guys try any of the alternatives before using Grapevine?
Interesting to see this now launching, when most companies have their own customGPT solution and MCP makes headway towards decoupling the frontend layer. Data seems to be stored outside of the customers control, so this will be a difficult sell for many companies.
What type of businesses are you targeting?
In terms of what businesses we're targeting: we wanted to provide either 1.) a turnkey solution for a company GPT for all the people who don't have it yet, or 2.) a higher quality company GPT for people who do have an internal solution.
Our sense is that ~70% knowledge companies at large still don't have a custom GPT yet, and that of the people who do, our system can be more performant because we're spending more effort than their internal team is. There's a lot of details we've solved on data ingestion and search algo that improved our accuracy dramatically, and things breadth of data connectors is the kind of thing that is expensive for an internal team but worth it if you're providing the service at large.
Not sure what you mean by "data seems to be stored outside of the customers control," but fortunately I think many SaaS apps that were trying to lock down customer data from themselves are walking that back a little bit.
Please look into the Zero Data Retention policies of the subprocessors that you are using. For example, Open AI does not include files as falling under their ZDR [1], thus utilizing OAI as an LLM solution inherently adds unnecessary risk of data exposure that many enterprise clients do no want to onboard. Also, you have to think about those companies obligation to their clients/customers when it comes to data security, along with the risk of IP being exposed to 3rd party systems that they do not have control over, when they make their decisions when it comes to utilizing various business solutions.
[1] https://platform.openai.com/docs/guides/your-data#zero-data-...
Was a huge fan of gather.town, is this the official notice that it's going into maintenance mode?
The Gather work product is unfortunately going into maintenance mode. We still have a strong team working on the core AV and performance, and the business is very decent and more than supports that team (and we still use the Gather product heavily ourselves).
However, it didn't reach the growth trajectory we needed, so a majority of the company will be working on Grapevine + new products instead.
This is sad to hear. We loved Gather but the technical issues stacked up and we made the choice to kill it around 2 weeks ago.
We've still been searching for a proper replacement for go-karting. Our team greatly enjoyed that little mini game.
A thought for any lurking vibe-coders.
Is there a recommended migration path ? or another product that you would suggest here?
Ah to be clear, the team is staffed well enough on Gather that the experience will be equal, if not better, than what you've been using already. So we haven't been telling customers to migrate (and unfortunately, there aren't many good alternative products right now).
I've seen Gleam and Onyx, and I think the real problem is that there is a lot of garbage coming in. If you want to solve the problem, you need to find a way to clean the information coming in. And if you've cleaned the information coming in, you have a lower need to have an LLM answer the question.
Our experience, especially with the most recent reasoning models, is that the LLM's are a lot better now at sifting through the garbage. So if you last gave these products a try more than a month ago, I would try them again.
(Additionally, there are a lot of details that do make a big difference in data processing / search algo too, which have taken our own internal accuracy on hard questions from 30% => 80%+)
Is this self-hosted? If not, am I to work under the assumption that my company's IP is worth ~$0?
No self-hosted version yet. You would need to trust us in the same way you trust your other SaaS apps that host company IP (e.g. Slack, Notion, Github, etc.)
I get that's a big ask from a startup. If it helps, we are a company that's been around for 4+ years and have built a work tool (https://gather.town) used for 100k+ people for their daily work, Sequoia-backed, are SOC II certified, and go way beyond that for the security considerations for this product.
Are slack/github/notion dumping company data into an unknown set of LLM APIs under the hood?
(My company uses Zulip/Gitea/Affine for data sovereignty reasons, but this kind of thing seems worse)
The name is way too long for people to type queries using it.
Fortunately, Grapevine is just the name of our system, we let people white-label their Slack bot when they actually set it up :)
YCS19 - so 6 years. Obviously you did well enough to survive. Interesting that the pivot is something so basic after all this time. Kind of interested in the story there.
2019-2020: We prototyped lots of telepresence ideas, our work shown here: https://siemprecollective.com/
2020-2022: We built https://gather.town, which during COVID blew up across every use-case possible: conferences, birthday parties, weddings, universities. It was a good business during COVID but eventually started to shrink.
2022-2025: We built Gather for remote workers, which was a long grind into in Audio/Video, performance, and making a game-interface that was good for work but replicated the parts of in-person work people enjoyed. It's a decent business, but didn't match our ambitions with how we wanted to change work for the better.
2025+: We have lots of ideas for how we can make work a lot better with AI. The general theme is, "can we make work as fun as a video game?" Idea being: video games are super similar to work at its core, and AI can both 1.) dramatically change how people need to spend their days, and 2.) help you "game design" someone's work day.
The Grapevine system is the first tablestakes layer people need to have for us to build the products we're excited about. Surprisingly, "company context" was not as good as we thought despite it being such an obvious, big business opportunity. So while I agree it's "basic," it does seem necessary, and is also not the full-scale of what we want to achieve still :)
Is the "ChatGPT" brand name becoming a generic term, like Baid-Aid or Kleenex?
There is the ChatGPT product, operated by OpenAI, Inc, which you can access via their web site or their API. OpenAI does publish gpt-oss as an open-weights model. I suppose you could argue that gpt-oss is "a ChatGPT," though I'd normally think of it as "a large language model." Much like Claude, DeepSeek, Qwen and so on are other large language models.
Well GPT = general purpose transformer
Generative pretrained transformer, I think? https://en.wikipedia.org/wiki/Generative_pre-trained_transfo...
It’s actually generally prime t-bones
> the day-to-day questions that actually blocked people
do you have a very strong opinion about how companies should work?
"No"
Okay, does Dario Amodei? He thinks more than half the workforce should "just" be replaced. That's a strong opinion! Do you see what I am saying?
I can't speak for other people, but our strong opinion about how companies should work is to reduce the massive amount of chores and tedium that exist in our work days today.
With the company GPT, we want to tackle things like: 1.) having to answer a repeated question from a colleague, 2.) answering questions to coworkers that are purely informational, and eventually 3.) things like standup updates, written updates to leadership on status, etc.
I think human interaction at work is one of the most valuable experiences if you're lucky enough to have good colleagues and interesting work. But I think they should almost entirely be around creativity, decision-making, debate, etc. rather than sharing information that exists elsewhere.
> I can't speak for other people
Your product is a bot that SPEAKS as an alternative to OTHER PEOPLE talking.
> reduce the massive amount of chores and tedium that exist in our work days today.
See... is that really a strong opinion? Like look what I asked you. How should COMPANIES WORK?
I'm being a little funny about this. I guess my point is that, there are a lot of Grapevines, including all the companies whose technologies you use. Paul Graham invests in all of them, and so he will be fine. But what about YOU? A soft and friendly Enterprise Sales tone... like give me a strong opinion. People at Google and Meta have better sales teams and technology. But they don't have strong opinions. Do you get it now?
what's your point, you ask for strong opinion without any context, what answer do you expect?
I don’t think anyone knows what you’re saying.