0813 | AI Agent Launches on Product Hunt: CodeBurn, BearDrive, Chat Agent, Grok Bot

Mia: Welcome to ProductHunt Daily, I'm Mia, and this is Bri. Today we've got a packed lineup of launches, starting with two very different takes on the same AI problem.

Milo: Thanks Mia, I'm Milo. We're also covering a codeless chat agent, AI teammates for real work, and a free way to run AI models on your own machine.

Mia: Plus two app builders taking aim at the same gap, a self-hostable voice AI platform, an agentic workflow automation launch, and even a credit card picker for Safari.

Milo: A lot to get through, so let's dive into today's Product Hunt briefing.

Mia: Two launches this cycle go after the exact same problem from opposite ends: AI coding spend shows up as one big opaque bill, and neither side can see where it went. CodeBurn attacks it for the solo developer, and Cohesor for the enterprise. Let's start with CodeBurn, because it's free, runs locally, and is open source. The pitch is that the answer already existed — every AI coding tool already writes session files to your disk. Claude Code, Cursor, Codex, Copilot, and about forty tools in all. Maker Resham says nobody was ever reading the logs. So CodeBurn does, and breaks every token and every dollar down by task, model, project, and even pull request.

Milo: It's a nice trick because there's no account and no uploads — everything stays on your own machine, and it's MIT-licensed. The maker claims more than 150,000 developers use it. And the surfaces cover just about everything, from a command-line tool to a desktop app for Mac, Windows, and Linux, a menu bar app on macOS, a panel extension for GNOME on Linux, a web dashboard served from your own machine, and even an MCP server, so an agent can check its own spending.

Mia: There's also an Optimize tab that hunts for waste — think cache bloat or the retry tax — applies a fix, and tracks how much it actually saved, with undo. And the community reaction is telling. One post describes spending more than 200 dollars a day on Claude Code with zero visibility before this. A GitHub comment praises what one user called the reconciliation guarantee. But the sharpest point came from someone with a strategy and venture background, who said what they really want is the cost per useful unit of AI work, not cost per token. And another commenter raised a real open question: with fan-out orchestration, where one task spins up five to ten parallel subagents and only a final step merges the output into a pull request — does CodeBurn group those child sessions under one parent task, or does it just show you several expensive, unattributed bills?

Mia: Next up is BearDrive, and the easiest way to picture it is an open-source shared folder for your team's AI agents. The founder, Snow of Runbear, frames it this way: agents already create real files locally — reports, decks, CSV exports, research — and BearDrive syncs the folder they already work in, so every file is born shared. That means versioned, attributed all the way down to the agent session, and reachable through links only the team can open.

Milo: So the workflow stays where it already is — it works with Claude Code, Cowork, Codex, Gemini CLI, and any local tool — and it's open source and self-hostable, with the managed service free during beta. The problem it's solving is that teams lose agent output. The company built it after everyone on the team used Claude Code daily, but agent work, in their words, lived and died on one person's laptop — so teammates kept re-researching things an agent had already done. Notion was a manual chore, Git was the wrong tool for research and ops artifacts, and Dropbox, as they put it, syncs bytes but tells you nothing about who made what or whether it's current.

Mia: The claimed fix: other agents read from the same folder seconds later at a real local path, with full change history — who, when, from which device — plus Slack links and a web view that needs no install. The makers say they run the whole company on it, including their strategy wiki, sales collateral, and research across CEO, marketing, dev, and support projects, and that they planned this very launch in it. They're also careful to flag what it is not: not an agent memory system, and not another workspace you have to migrate into.

Milo: And the community response included commenters describing shared per-customer setup as the workflow that sold them. So the throughline is simple: if your agents already write to a local folder and the output keeps vanishing with whoever ran it, BearDrive's bet is that the sync layer is what was missing.

Mia: Today's briefing centers on one launch: Chat Agent from Trigger.dev, which is billed as AI chat that keeps running after you close the tab. Founder James describes it as a backend for AI chat apps — every conversation gets its own machine that lives for the whole conversation, sleeps when nobody's typing, and wakes right where it left off, with no state management required from the developer.

Milo: The argument is that most chat agents are built on request/response endpoints, which hit timeouts and have no memory between turns — forcing developers to bolt on Postgres, Redis, and queue coordination. Chat Agent instead removes that API route entirely and works with the AI SDK developers already use: streamText on the server, useChat on the client.

Mia: And the maker-reported capabilities are the real substance here. No timeouts — production data showing that one in twenty turns runs longer than 36 minutes. If you refresh mid-response, streaming resumes where the browser stopped. Close the tab and come back days later, the conversation is preserved. Memory between turns is just variables. A sub-agent spawned four questions earlier still persists. Waiting on human approval doesn't get billed. And every turn is traced with prompts, tool calls, latency, and cost metrics, with nothing to set up.

Milo: Context from the founder: Trigger.dev is an open-source Apache 2.0 TypeScript platform for building and running AI agents and workflows, no timeouts, with a dashboard for every run. Community feedback was warm — one commenter notes the project was announced earlier this week and is open source, and another says they've been a Trigger customer since the early days and call the team A plus. So the practical shift is in how a developer thinks about a chat agent: not a single request that has to finish fast, but a long-lived machine you can walk away from.

Mia: And the last launch is Grok Bot, an early-beta product described as AI teammates that you can give real work to. The core idea: bots sign into your tools, use them the way a person would, and come back with finished work. You can message them like teammates, they keep context, multitask, learn your working style, and run on their own cloud computer.

Milo: The example tasks give you a real sense of the scope — running outbound overnight, coding expenses out of Gmail, scheduling a recruiting onsite, rebalancing paid ads, shipping a bug pack. Bots work in parallel, collaborate where it makes sense, and keep going around the clock even with your laptop closed. Show a bot a workflow once, and it saves that as a routine and runs it on its own the next time. The company claims bots get smarter over time, keep context, and learn from each other.

Mia: Pricing and availability: Grok Bot is included with Cursor Ultra at 200 dollars a month, billed monthly — and that includes the bot's own computer, sign-in to tools, scheduled routines, multi-platform access, and extended AI token limits. There's also Cursor Premium Teams at 120 dollars per seat per month, which adds centralized team billing and settings, a team marketplace for skills and plugins, shared usage analytics, and SAML and OIDC single sign-on. Users already on Cursor Ultra or SuperGrok can sign in and get Grok Bot included, and there's a macOS Apple Silicon download, with desktop and iOS mentioned.

Milo: The community conversation raised pointed questions rather than verified results. One commenter compared it to PPLX Computer, but noted Grok Bot adds bot-to-bot collaboration.

Mia: So the honest bottom line is that for a product this new, the takeaway is the ambition — delegating real, messy, multi-step work to an agent that keeps its own computer — while the details on how well it delivers remain open questions the community is still asking.

Mia: Let's start with something for people who like local AI but don't want to become junior system administrators. Unsloth has launched a desktop app on Product Hunt — a free, open-source application for running and training AI models entirely on your own machine. The maker describes it as the first desktop app to both run and train models, completely local, with native downloads for Mac, Windows, and Linux.

Milo: So the pitch is genuinely no-terminal. You can run large language models, image and video generation tools, and audio models, and there are no-code workflows for training and generating media. One commenter on the launch summed it up well — they said every version of Unsloth has gotten easier, but this desktop release feels like the most complete version yet, the first one they'd actually recommend to someone who wants local models without turning it into a weekend setup project.

Mia: And the same commenter flagged that everything runs fully offline, and there's a command called unsloth start that lets agent tools call your local models directly — so your models become useful beyond the app itself. Now, the product page lists some ambitious capabilities that aren't independently verified. Through that start command, agents like Claude Code and Codex can plug straight into a local GPU. And there's an OpenAI-compatible interface, so existing apps and scripts can connect to the models without special code.

Milo: On the image side, it supports tooling like MiniMax-H3, FLUX, and Z-Image, with video generation through Wan and LTX. To give you a sense of scale — on an Nvidia B200, they claim MiniMax-H3 generated a roughly high-definition, 124-frame video. Again, that's the maker's claim, not independently tested. The through-line here is the same one coming up again and again today: local AI, minus the terminal.

Mia: Two more Product Hunt launches are chasing the same gap — AI tools that give you a demo, but not something you can actually ship. First up, LaraCopilot has relaunched as V2 and calls itself an agentic AI engineer. You describe what you want in plain language and it produces a full stack — interface, backend, database, authentication, APIs — as real, exportable code, with no lock-in. The key detail here is that it's built natively for Laravel, and one commenter said they can't believe the quality of the Laravel code it writes.

Milo: V2 adds plenty: prototyping and build modes, importing existing projects, prompts in multiple languages, a choice of frontend frameworks, one-click deployment to common hosting setups, and team workspaces. The maker reports a beta launched at a large Laravel conference a year ago, now thousands of developers and over two and a half million lines of code generated. Pricing starts with a free tier, then monthly plans ranging from about 29 up to 199 dollars, billed yearly, with enterprise custom pricing. One commenter did ask how it compares to the popular build tool Lovable — and the launch didn't supply a direct answer.

Mia: The other one is Assembly Studio, and it's aimed at a very different buyer — professional service firms: accountants, law firms, agencies, consultants. The cofounder says clients could prototype in Claude Code, but got stuck on authentication, permissions, and hosting. So Assembly bakes the client experience in. Published apps live in a branded portal on your own domain with a single client login, and authentication is handled by the platform — magic links or Google with multi-factor — never generated by AI. It ships more than twenty templates for service firms, and the maker says firms in the private beta built over two hundred apps, many of them replacing Airtable.

Mia: Next, a voice-AI platform that's betting on openness. Dograh launched on Product Hunt as the open source alternative to VAPI — a self-hostable platform for building calling agents that answer phones, book appointments, qualify leads, or run payment reminders. The maker's argument is that closed platforms make teams rent agents on their cloud at their price, and he claims big labs are pushing to restrict open source AI. Dograh's stated position is blunt: no single company should own voice AI.

Milo: On paper, it's a strong package — and it's worth flagging these are all the maker's claims. It's released under a permissive license with a one-command install and nothing locked behind a paywall. You bring your own model keys across more than thirty integrations, or run local models completely offline and air-gapped. Stated capabilities include a visual flow builder, real-time speech in over seventy languages with mid-call language switching, warm handoff to a human agent, call recordings with automatic quality checks, and a free API.

Mia: There's also a server component so coding tools like Claude Code and Cursor can create and deploy voice agents directly from the editor, plus on-premises deployment where the calls, recordings, transcripts, and model inference all stay inside your own data boundary. They claim compliance relevance for HIPAA, GDPR, SOC 2, and similar frameworks. And there's a feature they call double conversions — it mixes pre-recorded human audio with text-to-speech in the same voice, letting the model pick your recorded lines when they fit naturally, with claims of cost cuts up to three times.

Milo: The community comments are worth treating as experiences, not verified results, but at least one early adopter from a startup called Floworks says they tried several voice-agent platforms before settling on Dograh. The recurring theme from a list like this — whether it's Unsloth's local desktop, or Dograh's self-hosted stack — is the same: keep the AI, cut out the platform middleman.

Mia: And finally, today's featured launch is Ballet — an agentic workflow automation platform from the team behind Brainfish. The maker says it follows twelve months and hundreds of conversations with operations leaders, which surfaced three recurring patterns. First, ops staff acting as human middleware — gluing systems together, waiting on IT, or building automation flows that break the moment a process needs judgment. Second, teams building disconnected AI agents that don't integrate with each other. And third, a clear preference for graduated control. His quote is pointed: nobody wants full autonomy.

Milo: The audience is operations teams — RevOps, support ops, product-led growth — with multi-system workflows stuck behind an engineering queue. Instead of telling a tool how to do a job, you describe an outcome — lead routing, attribution, a renewal play — in plain English. Ballet writes it as reviewable, version-controlled code, runs it, and keeps it working, with a full audit log, one-click rollback, simulation mode, and autonomy levels you choose yourself.

Mia: That's the key design choice. Output is deterministic where accuracy matters, and agentic reasoning is reserved only for steps that actually benefit from flexibility. So engineers can read, approve, and replay the generated code. A few illustrated use cases: instant lead capture into your CRM with enrichment and dedupe before routing, detecting a surge in account activity and pushing next-best-actions to reps, and propagating order changes through internal billing without opening an engineering ticket.

Milo: Against tools like n8n, Zapier, and Claude-style agents, the maker claims faster builds in plain English, reach into your proprietary internal systems, consistent output, auditable code, and predictable cost — though those claims are theirs to prove. But the deeper point is the same one threaded through all four launches today: the shift from assistants that suggest to systems that execute — with humans still deciding exactly how much rope to give them. That's the tension Ballet is explicitly selling against, and it's worth watching who wins it.

Mia: So this one actually stood out because of how the developer framed it. RightCard is a Safari extension from a solo developer, pitched as “the honest credit card picker — no bank login.” The whole premise is that it tells you which card to use at any given store, without ever asking you to link a bank account or hand over credentials. It reads the offers page you're already signed into, so your login never enters the picture.

Milo: And it's not just a recommendation tool. Per the maker, it auto-adds bank offers you'd otherwise forget to activate — across Amex, Chase, Citi, and Wells Fargo — it tracks rotating five-percent categories, and it warns you before card credits and annual fees renew. No account, no ads, no tracking SDKs, and the recommendations are computed on-device, even offline.

Mia: Right, and the developer's selling point is that one-tap offer activation across those four banks is free here, where it's a paid tier elsewhere. They also claim “honest numbers” — meaning no invented “you saved this much” math, just the banks' own stated ceilings, like “up to this amount.” So it'll flag when a bonus won't actually post. For example, Walmart usually codes as a superstore rather than a grocery store, so grocery bonuses won't apply there.

Milo: That kind of specificity seems to be the appeal. The developer says the last two months of development shipped largely from user bug reports, sometimes on the same day. And community reaction was mostly questions — but a few preferences came through. One commenter said removing the bank login setup is “where I stop trying apps like this,” and another admitted checking rewards feels like too much work, so the habit-fixing angle resonates.

Mia: The open questions are telling though. Would people with just a few cards actually find this useful, or is it mainly for reward enthusiasts? And is the biggest challenge really the tech — or getting people to change their habits at the register? Either way, it's a nice counterpoint to the login-everything trend in this space.

Mia: That wraps up today's Product Hunt briefings. Before we go, a quick rundown of the launches we covered.

Milo: We looked at BearDrive, an open-source shared folder, and Chat Agent by Trigger.dev, which keeps AI chat running after you step away. Plus Grok Bot's AI teammates and Unsloth Desktop for running models locally.

Mia: Also on the list — two AI app builders, Dograh as a self-hostable voice-AI platform, Ballet for agentic workflow automation, and RightCard, the credit card picker that skips the bank login.

Milo: A busy cycle of launches tackling everything from AI coding spend to voice agents and card recommendations. Thanks for tuning in — see you next briefing.