> Any call the agent makes to an external connector is reviewed by an LLM judge to prevent dangerous operations.
A few basic questions: how reliable is this judge since it’s based on an LLM? What additional measures can an admin or someone with more technical knowledge take to tighten this further if needed? Are the rules in or used by this judge visible to the users or an admin?
Maybe that's a hot take, but it's probably still safer than how things are often done in reality when employees want to "get something done" and sign up for random services, download their own software or other variants of shadow IT.
This posture is applicable while we ease out of what I'm calling the "golden era" (loose restrictions on access with limited consequence) but one of the reasons I am pedantic and bearish about solutions like this is that they too have a short lifespan that we can already estimate the expiration date of.
We assume LLMs can always be fooled, so the judge is just one step of the verification process. The deterministic layers, per-tool db roles, read-only-by-default integrations, query timeouts, are the ones an admin can configure. The judge's rules aren't exposed to admins today. I like your idea of letting admins add their own rules or knowledge to the judge, we should add that.
One of the core issue here is LLMs are probabilistic not deterministic, even with set of rules or context , llms are inherently built to make best guess, so my question is how you are ensuring that LLMs are following rules?
You're right, and we don't. The judge is best-effort screening not enforcement. Enforcement is deterministic with per-tool Postgres roles, proxy-injected creds, host allow-list, human approval on destructive writes. Those hold whether the judge is right or wrong.
I like the security-first posture. What's your view on how foundation models will or won't evolve into this space? Like will CC steamroll this in 2 years when it can natively build connectors and run them from within the desktop app? Not saying you won't have an ongoing edge, I just want to understand the thesis better so I can learn. Cool product!
Thanks! We believe the building aspect is already commoditized and we're experimenting with leaning into this with BYOA (bring-your-own-agent). Our bet is that the labs will stay focused on the model and won't do the more unglamorous governed deployment layer needs (credential scoping, per-tool db, audit, etc), but time will tell.
Don’t know if anyone else is feeling this, but I’m getting a sense they’re gona have to go in to this layer. Not enough money to warrant the investment in the model only approach…
I’m playing with this exact concept but going down the path of a Claude code plugin that is optimized for configuring lowdefy apps because of that framework’s unique approach where you don’t write code you generate yaml configs that then drive the rendering of an app. So far it’s working pretty well, still pressure testing it. The audit control and permissions management you have is great, especially around the connectors. Nice job!
Thanks! We've actually have been experimenting with the config generation approach as well. One tradeoff we noticed was it turns the system into more of an app builder where the ceiling of what you can do is lower compared to a coding agent. But it's much faster than codegen. How dynamic have you found Lowdefy compared to codegen?
Congratulations with your launch! I suspect the product market fit for tools like these will be huge. Especially for SMBs with just 10s of employees in the office, lacking the budget for SAP consultants. However I think the agent creating an app is adding unnecessary complexity. Users want answers or insight in data, why build an app for that if the agent can provide it directly?
Thanks! Agree that for one-off questions, chat is the right interface. We’ve found apps win when the workflow is repeated across a team and they almost become “templates” that seed ideas for other coworkers.
Me and my co-founder are building something similar but it started out as a tool to help build Lovable style prototypes for your existing product. Think Product Managers building an idea without having to figure out how to install git, VS Code, 17 other tools and then try share a URL.
Initial talks with customers resonated more with the internal tooling idea though. It's flexible enough to support both.
We're earlier along than this product though so there's a fair few wrinkles in our product.
Awesome launch and good pricing. We built an equivalent version in our org that takes HTML/JSX files and stores/serves them like an S3 bucket would. Works fabulous and we considered turning it into a SaaS product, met a real need in our org. Wishing you guys the very best of luck!!
I am currently building something similar for my company. We are enabling people to host apps with python backends as well, so that people can do more complex tweaking of data for scientific workflows. The out-of-box solutions keep users constrained to an environment that doesn't allow things like importing weird packages on the back and front end for custom apps. The security implications are challenging, but I can mostly just keep the apps hosted in a highly constrained environment.
Separately I haven't yet seen a great security governance model for LLM integrations. At an enterprise level I'd like to govern rules such that, for example, if someone gives an LLM access PII information or proprietary data, then it shouldn't have access to a slack integration or the internet. Controlling this at the employee or team level doesn't make sense, as an employee may have reason to make separate use of both. From a security perspective I want them to be able use LLMs with different types of acccess, but not necessarily the same agent at the same time. Furthermore, they could ideally chain together agents with different permissions in specific orders. For example they could have a workflow where an agent can reach out to the internet, and then have a separate one that can read/write to slack, and then have a third that can interact with PII data. If they tried to wire together agents in the opposite order, it should get denied.
We ended up splitting permissions per-tool for data and per-person for connections, for basically this reason. Every tool gets its own db role that can only see its own schema, and integration credentials live in a broker the tool's code never touches. So one tool can't read another tool's data, and nothing the model writes can leak a credential.
The ordering thing is harder and currently we don't do it. Per-tool permissions at best ban a combination, they can't say "internet then Slack is ok, the reverse isn't." DeepMind's CaMeL paper (arxiv.org/abs/2503.18813) is the best take I know on this, it tracks data flow between tool calls and enforces policy there. Still an open problem as far as I can tell.
Thanks! That's the best kind of validation. If you ever take Prized for a spin, would genuinely love to hear how it compares to what you built, especially on the data access side.
Very cool. What model is used for the judge? We use setoku for Claude-built internal tooling but the design doesn’t allow writes so that there’s no inference cost on the server. Having the judge check for danger is a neat design.
Today it's GPT-5.6 Luna sitting inline on the egress broker, with Terra re-judging anything Luna flags as ambiguous. We are still experimenting though.
I did build a very similar product putting emphasis on integrations and security. It was tough to churn anyone out from Replit, Lovable and then CC and Codex.
When I talked to Lovable team, they told me Lovable is used primarily for prototyping, not internal tools and they are having a rough time trying to change the use case.
Thanks for the feedback. From what we've gathered so far, non-eng are using CC or Codex to build one-off internal tools, but they're deploying them on their personal Vercel accounts and sending the url over Slack. That's the workflow we want to replace with a reasonable level of security and auditability. We're still trying to gauge interest in using the CLI to build and having everything deploy on Prized, we're working out whether that's valuable.
The difference is Retool still basically trusts whoever (or whatever) is building the app, we don't. The agent writing the code never sees a credential and each tool gets its own Postgres role. Even if it generates something dumb it physically can't touch data outside its scope.
nice this is awesome, what i've been building internally for a while now. lovable-esque internal platform with an agentic chat building experience, preview window, our internal auth, internal standards, sandboxed agent devcontainers on k8s. Even down to the shared experience -- these are shared tools for everbody, team ownership structure, not individual. easily forkable, auto-genned product docs etc etc. Hope you can monetize it! Look into private endpoint connectivity for big corps that need the internal access but don't want to self host.
Thanks! The private endpoint tip is very much on point. We've already hit two prospects where the blocker was that they don't want full self-host but their security groups won't take generic cloud egress. It's on the list.
You can do this in lovable too now. But i dont think it gets used a lot. I think it is not very unique, i have seen bunch of providers, some are selling ai workers, some are selling internal tools in different shape and form.
I think the moat for internal tools is not where you think it is. Also, i see self host option, it is a nightmare too, no one will do it just for internal tools now. It was valid before AI era.
I am also seeing non engineering staff actually building tools with claudecode, paas like railway and stuff. People have become more efficient with engineer tools.
Fair, the generation part is commoditized. What we sell is the trust layer around company data. No credentials in the builder's hands, scoped access per tool, forkability etc. CC + Railway works until someone pastes a prod key into an env var.
Closest comp. I'd argue the "AI builds the app part" is fast becoming table stakes. The unsolved problem is what happens after the app exists and that's what we actually built. Anyone in the org can fork a coworker's tool and rebind it to their own data scope. Enforcement also happens at a different layer. From what we've gathered, Superblocks controls who can run an app and bakes guardrails in when the app is generated, but every app still queries through a shared database credential. On our side, each tool gets its own database role that can only see its own data, and credentials live in a proxy the code never touches. So even when the AI writes bad code, it can't reach anything it shouldn't.
Congratulations on the launch. I’m very bullish on this concept of allowing SMB’s and startups to roll their own software and internal tools. Many many moons ago I built an entire inventory tracking and CRM for a very small company in PHP (circa 2006). It ran the entire company with orders, customers, devices, and revenue. It was a monstrous app, and took forever. Today, could be build in a matter of days and much better.
Along the same lines. I got so fed up with Quickbooks self employed (god awful software and company) for my own business I wrote a replacement using Cursor. It tracks expenses (categorizing each in proper IRS categories) and estimating quarterly taxes. Uses Plaid to pull in business bank accounts and credit cards and has income and expense reports. It’s 100x better than Quickbooks.
> Any call the agent makes to an external connector is reviewed by an LLM judge to prevent dangerous operations.
A few basic questions: how reliable is this judge since it’s based on an LLM? What additional measures can an admin or someone with more technical knowledge take to tighten this further if needed? Are the rules in or used by this judge visible to the users or an admin?
It's not reliable, and presenting it as-such shows a lack of understanding of basic security posture.
Don't take my word for it, watch the news in <6 months and we'll certainly see it get bypassed.
Maybe that's a hot take, but it's probably still safer than how things are often done in reality when employees want to "get something done" and sign up for random services, download their own software or other variants of shadow IT.
Completely agree.
This posture is applicable while we ease out of what I'm calling the "golden era" (loose restrictions on access with limited consequence) but one of the reasons I am pedantic and bearish about solutions like this is that they too have a short lifespan that we can already estimate the expiration date of.
We assume LLMs can always be fooled, so the judge is just one step of the verification process. The deterministic layers, per-tool db roles, read-only-by-default integrations, query timeouts, are the ones an admin can configure. The judge's rules aren't exposed to admins today. I like your idea of letting admins add their own rules or knowledge to the judge, we should add that.
One of the core issue here is LLMs are probabilistic not deterministic, even with set of rules or context , llms are inherently built to make best guess, so my question is how you are ensuring that LLMs are following rules?
You're right, and we don't. The judge is best-effort screening not enforcement. Enforcement is deterministic with per-tool Postgres roles, proxy-injected creds, host allow-list, human approval on destructive writes. Those hold whether the judge is right or wrong.
I like the security-first posture. What's your view on how foundation models will or won't evolve into this space? Like will CC steamroll this in 2 years when it can natively build connectors and run them from within the desktop app? Not saying you won't have an ongoing edge, I just want to understand the thesis better so I can learn. Cool product!
Thanks! We believe the building aspect is already commoditized and we're experimenting with leaning into this with BYOA (bring-your-own-agent). Our bet is that the labs will stay focused on the model and won't do the more unglamorous governed deployment layer needs (credential scoping, per-tool db, audit, etc), but time will tell.
Don’t know if anyone else is feeling this, but I’m getting a sense they’re gona have to go in to this layer. Not enough money to warrant the investment in the model only approach…
I’m playing with this exact concept but going down the path of a Claude code plugin that is optimized for configuring lowdefy apps because of that framework’s unique approach where you don’t write code you generate yaml configs that then drive the rendering of an app. So far it’s working pretty well, still pressure testing it. The audit control and permissions management you have is great, especially around the connectors. Nice job!
Thanks! We've actually have been experimenting with the config generation approach as well. One tradeoff we noticed was it turns the system into more of an app builder where the ceiling of what you can do is lower compared to a coding agent. But it's much faster than codegen. How dynamic have you found Lowdefy compared to codegen?
I know someone who could use this but a price point/plan between $0 and $100 would be a good way to help people get started.
Congratulations with your launch! I suspect the product market fit for tools like these will be huge. Especially for SMBs with just 10s of employees in the office, lacking the budget for SAP consultants. However I think the agent creating an app is adding unnecessary complexity. Users want answers or insight in data, why build an app for that if the agent can provide it directly?
You can reuse the app indefinitely, for cheap, and predictably. I would say these are huge advantages.
Thanks! Agree that for one-off questions, chat is the right interface. We’ve found apps win when the workflow is repeated across a team and they almost become “templates” that seed ideas for other coworkers.
Me and my co-founder are building something similar but it started out as a tool to help build Lovable style prototypes for your existing product. Think Product Managers building an idea without having to figure out how to install git, VS Code, 17 other tools and then try share a URL.
Initial talks with customers resonated more with the internal tooling idea though. It's flexible enough to support both.
We're earlier along than this product though so there's a fair few wrinkles in our product.
https://flexenv.com/
Awesome launch and good pricing. We built an equivalent version in our org that takes HTML/JSX files and stores/serves them like an S3 bucket would. Works fabulous and we considered turning it into a SaaS product, met a real need in our org. Wishing you guys the very best of luck!!
I am currently building something similar for my company. We are enabling people to host apps with python backends as well, so that people can do more complex tweaking of data for scientific workflows. The out-of-box solutions keep users constrained to an environment that doesn't allow things like importing weird packages on the back and front end for custom apps. The security implications are challenging, but I can mostly just keep the apps hosted in a highly constrained environment.
Separately I haven't yet seen a great security governance model for LLM integrations. At an enterprise level I'd like to govern rules such that, for example, if someone gives an LLM access PII information or proprietary data, then it shouldn't have access to a slack integration or the internet. Controlling this at the employee or team level doesn't make sense, as an employee may have reason to make separate use of both. From a security perspective I want them to be able use LLMs with different types of acccess, but not necessarily the same agent at the same time. Furthermore, they could ideally chain together agents with different permissions in specific orders. For example they could have a workflow where an agent can reach out to the internet, and then have a separate one that can read/write to slack, and then have a third that can interact with PII data. If they tried to wire together agents in the opposite order, it should get denied.
We ended up splitting permissions per-tool for data and per-person for connections, for basically this reason. Every tool gets its own db role that can only see its own schema, and integration credentials live in a broker the tool's code never touches. So one tool can't read another tool's data, and nothing the model writes can leak a credential.
The ordering thing is harder and currently we don't do it. Per-tool permissions at best ban a combination, they can't say "internet then Slack is ok, the reverse isn't." DeepMind's CaMeL paper (arxiv.org/abs/2503.18813) is the best take I know on this, it tracks data flow between tool calls and enforces policy there. Still an open problem as far as I can tell.
Thanks! That's the best kind of validation. If you ever take Prized for a spin, would genuinely love to hear how it compares to what you built, especially on the data access side.
Very cool. What model is used for the judge? We use setoku for Claude-built internal tooling but the design doesn’t allow writes so that there’s no inference cost on the server. Having the judge check for danger is a neat design.
Today it's GPT-5.6 Luna sitting inline on the egress broker, with Terra re-judging anything Luna flags as ambiguous. We are still experimenting though.
I did build a very similar product putting emphasis on integrations and security. It was tough to churn anyone out from Replit, Lovable and then CC and Codex.
When I talked to Lovable team, they told me Lovable is used primarily for prototyping, not internal tools and they are having a rough time trying to change the use case.
Hope my experience helps!
Thanks for the feedback. From what we've gathered so far, non-eng are using CC or Codex to build one-off internal tools, but they're deploying them on their personal Vercel accounts and sending the url over Slack. That's the workflow we want to replace with a reasonable level of security and auditability. We're still trying to gauge interest in using the CLI to build and having everything deploy on Prized, we're working out whether that's valuable.
How does this differ from retool?
TBH I was a retool power user for years, but I feel like they completely missed the vibe code train.
It's so much more annoying to go drag boxes around and type out little JS snippits in every chart as opposed to just telling claude to make the chart.
Have you looked at their latest product? You describe your app in a text box and the AI builds it, with all the enterprise goodies like RBAC baked in.
The difference is Retool still basically trusts whoever (or whatever) is building the app, we don't. The agent writing the code never sees a credential and each tool gets its own Postgres role. Even if it generates something dumb it physically can't touch data outside its scope.
It's funded by YC!
https://www.ycombinator.com/companies/retool
So is retool. :)
nice this is awesome, what i've been building internally for a while now. lovable-esque internal platform with an agentic chat building experience, preview window, our internal auth, internal standards, sandboxed agent devcontainers on k8s. Even down to the shared experience -- these are shared tools for everbody, team ownership structure, not individual. easily forkable, auto-genned product docs etc etc. Hope you can monetize it! Look into private endpoint connectivity for big corps that need the internal access but don't want to self host.
Thanks! The private endpoint tip is very much on point. We've already hit two prospects where the blocker was that they don't want full self-host but their security groups won't take generic cloud egress. It's on the list.
I like this a lot. Do you handle scenarios where the data and context live in a data warehouse (Snowflake, etc)?
There are a lot of operational needs that exist downstream from dbt models that combine data from the various source systems.
Yes, we support Snowflake and ClickHouse today. And we're happy to integrate more, just say the word :)
Demo looks cool, congrats on the ship! Probably missed but is it possible to externally share/view the built artifact outside of prized?
Thank you! Yes, you can share it externally if it isn't connected to external data sources.
You can do this in lovable too now. But i dont think it gets used a lot. I think it is not very unique, i have seen bunch of providers, some are selling ai workers, some are selling internal tools in different shape and form.
I think the moat for internal tools is not where you think it is. Also, i see self host option, it is a nightmare too, no one will do it just for internal tools now. It was valid before AI era.
I am also seeing non engineering staff actually building tools with claudecode, paas like railway and stuff. People have become more efficient with engineer tools.
Fair, the generation part is commoditized. What we sell is the trust layer around company data. No credentials in the builder's hands, scoped access per tool, forkability etc. CC + Railway works until someone pastes a prod key into an env var.
Seems like Superblocks
Closest comp. I'd argue the "AI builds the app part" is fast becoming table stakes. The unsolved problem is what happens after the app exists and that's what we actually built. Anyone in the org can fork a coworker's tool and rebind it to their own data scope. Enforcement also happens at a different layer. From what we've gathered, Superblocks controls who can run an app and bakes guardrails in when the app is generated, but every app still queries through a shared database credential. On our side, each tool gets its own database role that can only see its own data, and credentials live in a proxy the code never touches. So even when the AI writes bad code, it can't reach anything it shouldn't.
Congratulations on the launch. I’m very bullish on this concept of allowing SMB’s and startups to roll their own software and internal tools. Many many moons ago I built an entire inventory tracking and CRM for a very small company in PHP (circa 2006). It ran the entire company with orders, customers, devices, and revenue. It was a monstrous app, and took forever. Today, could be build in a matter of days and much better.
Along the same lines. I got so fed up with Quickbooks self employed (god awful software and company) for my own business I wrote a replacement using Cursor. It tracks expenses (categorizing each in proper IRS categories) and estimating quarterly taxes. Uses Plaid to pull in business bank accounts and credit cards and has income and expense reports. It’s 100x better than Quickbooks.
Thanks! Custom beats generic if building it takes days instead of months.