0807 | AI Ops: CopilotKit Channels, Shieldstral, Superlog, and Rippling AI Spend Console

Mia: Welcome back to ProductHunt Daily. I'm Mia.

Milo: And I'm Milo. Today we've got a packed lineup of developer tools fresh off Product Hunt.

Mia: We're starting with the big one: CopilotKit is launching an open-source toolkit for building and self-hosting AI channels.

Milo: And later, Mistral AI's new safety guardrail model, plus a free AI bug-fixing agent and a console for tracking corporate AI spending.

Mia: We'll also cover tools that make your products visible to AI agents, an API for turning websites into clean markdown, an AI model gateway, and a clever macOS app that turns screen recordings into agent prompts.

Milo: Lots to get through, so let's jump in.

Mia: Let's start with CopilotKit Channels SDK. It's an open-source, self-hostable kit aimed at one very specific pain point: teams build good AI agents, but getting one to run inside Slack, Teams, Discord, or WhatsApp often means effectively rewriting the whole thing. CopilotKit's argument is that that should be a rewrite at all — the real conversations already happen in these chat tools, so your agent should just live there.

Milo: And the key idea is that the agent doesn't get rebuilt or given a separate front end. You tag it in a thread, and it answers with native, interactive UI — buttons, forms, even charts — instead of a wall of text. The agent keeps its own tools, its own model, its own business logic. CopilotKit is pitching this as turning your agent into a real “coworker” across chat platforms.

Mia: Under the hood, the architecture is split. Your agent and app logic stay running in your own infrastructure, while CopilotKit's intelligence layer manages the platform connection. That integration is available now for Slack and Microsoft Teams, with more channels on the way, including Discord and WhatsApp.

Milo: And it's meant to work with the agents developers already have. The list includes OpenAI Agents, Claude Agents, LangChain, Mastra, Google ADK — plus LangGraph, CrewAI, and Pydantic AI mentioned in the repository — and really anything that speaks the AG-UI protocol. Onboarding with a coding agent is compressed to a single prompt or one setup command.

Mia: The pitch does list some ambitious capabilities — streaming responses, generated UI, per-user learning, human-in-the-loop approvals, and what they call sophisticated auth. The developer-focused draw here is clear: if a team has already built an agent and wants it usable inside the tools where people already work, this is meant to remove the rewrite entirely.

Mia: Next up, Shieldstral from Mistral AI — a 3-billion-parameter, open-weight multimodal model that rethinks safety as something you define at runtime, rather than baking a fixed policy into the weights. It's released under the Apache 2.0 license and treats content moderation as a question-answering task: you give it evaluating context, a single yes-or-no question, and the content to judge — which can be a prompt, a response, both, or an image.

Milo: So instead of a discrete yes or no label, it reads out only the positive and negative signals and normalizes them into a calibrated continuous safety score. That means a team can set its own threshold — or rank results by confidence — rather than being stuck with one fixed verdict. The differentiator is policy adaptability: a kids' app, a cybersecurity tool, and a mental-health platform can look at the same content very differently, and deciding where to draw the line is the hard part.

Mia: So you can pass in a plain-language policy at inference time — something like “is this image safe for minors?” or “does this response promote physical violence?” — and the same single checkpoint adapts to it without retraining. That one formulation is meant to unify things like prompt classification, response moderation, refusal detection, and toxicity detection across both text and images.

Milo: Mistral's claim is that Shieldstral matches or beats open guard models up to seven times its size across text safety, refusal detection, policy adaptability, and multimodal benchmarks — and they call it a new state of the art on the multimodal side. The practical draw for a team is that safety policy no longer lives inside the model's weights; it becomes something you can tune per situation.

Mia: Then there's Superlog Responder — a free, open-source AI bug-fixing agent aimed at teams that already run Sentry or Datadog alerts into Slack. Co-founder Nicolò says feedback from the first launch drove this version, because users kept telling them they already have telemetry and alerts in their ops channel and don't want another dashboard to click through. Essentially — “I just want the bug fixed.”

Milo: So Responder plugs into that existing channel with one-click sync and no new telemetry to install. On each alert, the company says, it investigates with full context, filters out the noise, and for genuine issues replies in the thread with a root cause, the evidence, and a mergeable pull request. It can draw on Datadog, Sentry, Notion, the repo itself, and a read-only database.

Mia: The prompts, memory, tools, repo access, and escalation rules are all customizable, and there's a cloud option for teams that prefer not to deploy it themselves. They're onboarding early teams by hand this week, and commenters highlighted the free tier for testing.

Milo: The strongest community evidence, though, is limited. One commenter said they used it in beta for a few days and found it awesome, and another congratulated the team on a strong plus-90 percent pull-request merge rate — a figure the maker hasn't verified. The thread mostly raised open questions: what actually changed since the previous launch, how the team tests and iterates on prompts and models, whether noisy triaged alerts still end up looking like standard Sentry or Datadog alerts, and how it stops a technically correct fix from ignoring a team's architecture or coding standards.

Mia: And finally, Rippling has launched its AI Spend Console, aimed at finance and engineering leaders who need to track AI spending and tie it to business outcomes. Product lead Kevin points out that every company is adopting AI fast, but few have the infrastructure to manage AI spend as its own expense category. Today, finance teams manually pull vendor billing data and run ad-hoc analysis just to get a point-in-time view — with no visibility into which teams, departments, or models are driving the increase, or whether that spend actually improves outcomes.

Milo: Rippling says it hit this same problem internally, and that shaped the product. The console breaks costs down by vendor, model, or individual employee, and maps that spend onto employee attributes — department, team, role — alongside GitHub output data like pull request volume and the number of code revisions. So unlike a simple usage dashboard, the point is connecting spend to business metrics like performance ratings or pull request volume to flag inefficient use.

Mia: According to the maker, Rippling AI generates personalized dashboards from the connected data, supports natural-language follow-up questions, and can even govern approved LLM use by enforcing policies on token spend and model access based on organizational data. It's free to start, with no Rippling subscription required.

Milo: Community commenters focused on a few standout features: the employee-level breakdown, to check whether some engineers get more value from AI tools than others; the model-level visibility, since some models cost more for marginal quality gains; and the ability to hand a CFO concrete evidence that AI tools are worth their license fees. One commenter from a small engineering org said the console pointed them at the exact problem of tying AI spend to outcomes.

Mia: Kicking things off with UCP Radar, a Product Hunt launch aimed squarely at ecommerce merchants. Its tagline says it all: “Make Your Products Visible to AI Agents.” The maker behind it has 18 years in digital marketing, most of it running paid ads and Google Shopping, and the core argument is that getting a feed accepted by Google is a very low bar. Titles get written for humans browsing category pages rather than for how Google actually matches queries, descriptions are whatever the CMS happened to spit out, and the meaningful attributes — material, age group, and the newer AI-facing fields like product highlights, details, and Q&A — sit empty.

Milo: And there's a concrete pressure behind this. The maker says clients recently started asking why ChatGPT recommended a competitor instead of them. And unlike Google Shopping, there's nothing to bid up, no way to pay to climb those rankings in AI recommendations. So the tool connects to Google Merchant Center in one click or in about 30 seconds, scores each product against 50-plus Google Merchant Center rules and 35-plus Universal Commerce Protocol rules, and then separately scores how readable that product is to an AI agent.

Mia: Then it rewrites weak titles, fills in missing fields, and publishes a supplemental feed that Google, Perplexity, and ChatGPT pick up on their own. It outputs a Google Shopping XML supplemental feed for Merchant Center, a ChatGPT Shopping JSONL feed, and server-rendered structured data for store pages, which auto-regenerates whenever products change. Prices and stock still come from the merchant's store, and the maker highlights a “Brand Protector” feature to keep the rewrites from drifting off-brand.

Mia: Next up, Brandfetch MCP — a connector for developers building AI-agent workflows that need real brand assets. The pitch is that AI agents typically redraw logos, invent hex codes, pull outdated assets, and make up a brand voice. This connector is meant to fix that by supplying logos, colors, fonts, company details, and brand context for more than 50 million brands.

Milo: It's available as a directory connector in Claude, or from any MCP-compatible client like Cursor, VS Code, or Codex. There are five tools on offer. One resolves a company from a name, a domain, or a fuzzy query. Another returns the actual logos, colors, fonts, company details, and social links. A third surfaces the brand's voice, positioning, audience, and products. There's a tool that resolves merchants from raw transaction descriptors, and one that builds production-ready CDN logo URLs.

Mia: Suggested uses include keeping customer-facing documents on brand, building pitch decks and mockups with the correct customer logos, and running enrichment flows with firmographics. The maker's example walks through building a sales deck comparing three companies using each one's actual logos, colors, and positioning. And the community reinforced the problem — one commenter said they genuinely hit the invented hex code issue while building landing pages with Claude, noting that Claude would confidently return colors nowhere near the real brand palette.

Mia: Moving on to the Website to Markdown API, a newly launched service from a maker named Johnny that converts any submitted URL into clean, LLM-ready Markdown in a single API call. This is pitched as replacing the usual scraping stack — rendering JavaScript pages like React and Next.js as if they were static HTML, stripping navigation, footers, cookie banners, and ads, and handling anti-bot measures with built-in proxy rotation, browser fingerprinting, and retries.

Milo: It also returns a CDN-hosted screenshot, and the same endpoint handles PDFs, Word documents, slide decks, e-books, images, audio, and video. There's a JSON mode if you'd rather have structured text chunks than Markdown. It's part of the broader Exabase platform, so the same key can tap into deep search, memory, and automation, though it works standalone. There's a free plan with no credit card required.

Mia: The stated problem is that building and maintaining a headless browser, writing cleanup logic, and retrying blocked sites works for the first ten sites and breaks on the eleventh. The claimed result is output that drops straight into an LLM context window, knowledge base, or RAG pipeline with no post-processing. Community reaction was mixed but substantive — one marketer said it seemed useful, especially for multiple formats, another asked about an MCP server or a hook for Claude or Codex, and a third just wanted concrete examples, a question the makers hadn't answered in the supplied discussion.

Mia: Finally, Token Harbor — a developer-facing gateway that bills itself as the easiest way to access frontier AI models through one OpenAI-compatible API. The pitch is that you configure once, switch models freely, and pay only for what you use, instead of juggling separate providers, APIs, accounts, and changing configurations. Launch copy names GPT, Claude, Gemini, DeepSeek, and Kimi, with Grok added in a community comment and Anthropic, OpenAI, Gemini, Zhipu, and Kimi listed on the site as tryable providers.

Milo: There's also a “Connect” feature pitched as making the switch simple — one command configures coding agents and tools to work with Token Harbor, pointing specifically at Claude Code, Cursor, or Codex. Pricing details are thin: there's a free trial and five dollars of free credit for new accounts, but no per-token rates supplied. The site lists four self-described promises that should be treated as claims rather than verified results — model integrity with no silent model swaps, no baseline data policy with retention set by the customer, full transparency about routing, and responsive support.

Mia: The company is Token Harbor PTE. LTD., and community evidence is limited to the maker's own launch comment. They say the product is still early and ask for feedback on which models people use most, the biggest pain point in switching providers, and where they should focus next — so this one's very much a work in progress and the real open question is whether a single gateway with that kind of transparency holds up as the front door for agent workflows.

Mia: A new free macOS app called Annotate is trying to fix a real pain point for developers: describing UI changes so an AI coding agent actually gets them right. It runs only on Apple Silicon and works fully offline. You press a keyboard shortcut to start recording your screen, use a couple of toggles to draw on or point at parts of the interface, speak what you want changed, and stop. Then you open that session inside Cursor, Claude Code, or Codex, or just copy the prompt — the maker says it works with basically any AI coding agent.

Milo: The interesting part is how it keeps token use down. Instead of dumping the whole video at the agent, the app sends keyframes of what was on screen plus your transcribed voice, and it aligns the transcript with what you were looking at. That all happens through a local connection, so nothing leaves your Mac — no cloud upload, no login required, and the frames and transcripts stay on your machine.

Mia: One commenter flagged exactly that local-only design as the big win, because removing the upload step lowers the friction of getting the thing set up. Another noted that sending keyframes and a transcript rather than the full video is what makes the approach practical in the first place.

Milo: There are still open questions, though. Someone asked whether the transcript stays tied to the right frames in order across a multi-step flow, and another person asked about Windows or Linux support. And to be clear — all of this is the maker's own self-reported claims, not measured results. There are no benchmarks or user tests in the evidence, just the description and the community's reactions so far.

Mia: That wraps up today's briefing. We kicked things off with CopilotKit's open-source Channels SDK for bringing AI conversations into your own apps, and Mistral's Shieldstral guardrail model for keeping text and image models safe at runtime.

Milo: Then we covered Superlog's free bug-fixing agent, Rippling's AI spend tracking, and UCP Radar for making ecommerce products visible to AI agents.

Mia: Plus connectors and tools for brand assets, website-to-Markdown conversion, frontier model access, and Annotate for turning screen recordings into agent-ready prompts on your Mac.

Milo: Thanks for listening, and we'll see you next time.