0814 | AI Agents Today: Ito’s Live Code Review, Pickle Browser, Kane CLI & Phinq

Mia: Hey, I'm Mia, and welcome back to ProductHunt Daily! This is Bri.

Milo: And I'm Milo. Today we're looking at some tools that are shaking up how developers work, how AI agents stay under control, and even how you keep an eye on your Starlink connection.

Mia: We've got an AI code reviewer that actually runs your code before it reviews it, plus a brand new crop of AI-powered browsers.

Milo: And we'll dig into agent governance, a clever Hacktoberfest-inspired proxy, and a tool that records your doctor visits so you don't have to. Lots to get through — let's dive in.

Mia: So let's start with Ito, an AI code review tool that actually runs your code before it reviews it. For every pull request, Ito spins up a temporary environment, tests the parts of your app that could be affected, and comes back with runtime evidence — video, logs, screenshots — so the team can see what really broke instead of guessing from the diff.

Milo: And that's the pitch, right? Coding agents are producing way more code than any human can review, but most existing AI review tools just stare at the text of the code. Cofounder Evan says that's the fundamental gap — they're looking at the code without ever watching it run.

Mia: Right, and Ito's approach is to create a test plan from the code, the PR description, the ticket, and your input. Then it builds the app in a single-use isolated container that can use credentials, seed data, and even external services. When it lacks access to something, it mocks the data or code instead.

Milo: And then it drives the actual application with a swarm of agents — clicking, navigating, making API calls, verifying the database. The launch claims say it targets exactly the kinds of runtime issues that don't show up in a diff, like concurrency problems, failure handling, data migrations, and authentication, and it posts the evidence directly on the pull request.

Mia: The team says Ito gets better with use too — it responds to feedback, speeds up over time, and supports test instructions per organization, repository, or author. Cofounder Grayson made the point that the hard part is generalizing the ability to run any app, and that internal teams at companies like DoorDash or Ramp apparently spent quarters of engineering time trying to build exactly that.

Milo: And that's why runtime review didn't exist before, according to him. One telling detail — cofounder Barron, who calls himself a non-engineer, says Ito gave him the confidence to ship changes to his marketing site without bugging the engineers, even asking for specific tests like mobile usability of the blog. Pricing starts at the first tier the launch page lists, so it's worth checking the product page for the exact number.

Mia: From reviewing other people's code to testing your own — Kane CLI from TestMu AI is a terminal-based quality verifier built for developers and AI coding agents. You describe a test in natural language, and Kane CLI runs it in a real Chrome browser, returning pass or fail with shareable proof, and you never have to write a selector.

Milo: So the framing from co-founder Jay Singh is that AI writes code in minutes, but someone still has to actually open the app and confirm it works. Kane closes that loop — it ingests the source or the product requirements doc, designs the use cases and test scenarios, runs them, and comes back with an evidence pack, acceptance-criteria coverage, and a verdict.

Mia: And it's local-first and free to start, with one install command via npm, brew, or curl. Co-founder Mudit Singh describes who the loop is for: developers feeding in a product requirements doc, business teams describing flows in natural language, AI agents verifying their own work, and mobile teams pointing the same commands at Android emulators and iOS simulators.

Milo: The team's head of AI products, Shantanu Wali, makes a pointed contrast with most other tools — he says they only give you a test runner, whereas Kane is an assurance loop: ingest, design, run, evidence, coverage, verdict. And there are a few decisions they highlight: natural-language objectives instead of selectors, an Autoheal feature so cosmetic UI changes don't break a flow, and agent-native output so tools like Claude Code or Cursor can call Kane CLI and read the verdict directly.

Mia: It runs locally on your own Chrome, and an engineer adds that every command emits machine-readable output, with every verdict shipping an evidence pack that includes screenshots, console logs, and network responses. Co-founder Mayank Bhola also flags something — the launch copy names detecting issues early as part of the value, though the entry trails off before finishing that point.

Mia: Now this next one is about a different kind of safety. Phinq is an open source MIT-licensed runtime governance layer for AI agents. It intercepts every tool call an agent makes, classifies it by risk, lets safe actions pass through, holds irreversible actions for a human to approve, and records every decision in a tamper-evident, hash-chained audit log.

Milo: And there's a story behind it. The maker is a 21-year-old fresh graduate in London, and he says he built it after watching an agent overwrite something it had no business touching. His argument is that nobody was actually asking what agents should be allowed to do in the first place.

Mia: He cites two incidents to make the case. One coding agent deleted a company's entire production database and all its backups in nine seconds. Another wiped out 1,206 executives' records during a freeze the user had explicitly declared. And he also frames the launch around regulation — pointing out the EU AI Act started enforcing human oversight just ten days before his post went up.

Milo: The differentiation from existing options is structural. A prompt or a skill is advisory — the agent can just ignore it. Phinq instead enforces at the API or execution boundary, so governance can't be skipped. It runs as a proxy, as an in-process SDK, or as a plain HTTP gate, and it's compatible with OpenAI, Anthropic, and Gemini APIs, including Claude Code, Codex, and Gemini CLI, as well as frameworks like LangChain, CrewAI, Mastra, and MCP.

Mia: Now some caveats on the evidence. The maker's own stats page reports 9,031 decisions classified, 323 held, 50 denied, and 116 attempts to disable the safeguard — but those are self-reported figures, not independently verified. The audit log is hash-chained and append-only, though the launch description also notes some limitations on the recovery side that trail off in the entry.

Mia: And if the theme of this segment is making agents safe to run, Execlave is the other take on it. It's an AI agent governance and enforcement platform — positioned as a runtime gate between autonomous agents and the systems they touch, aimed at platform and security teams. Co-founder Rishit says the tools they found watched agents after the fact — traces, evals, dashboards, all retrospective — while Execlave checks every action against policy before it ever executes.

Milo: Worth noting these are maker and site claims, not independently verified results. But the claimed capabilities are substantial. There are four enforcement modes — block, warn, monitor, and require approval — so actions are either allowed, denied, or held for a human. Every call, payload, and verdict is kept, cryptographically signed and replayable. And there's a server-side kill switch the site says responds in under six milliseconds, per agent or across the whole organization.

Mia: The compliance story is broad too — the reports are said to map to SOC 2 Type II, the EU AI Act, ISO 27001, GDPR, HIPAA, PCI DSS, and the NIST AI framework. The site's figures, labeled as measured in production-like workloads, claim enforcement in under 20 milliseconds at the median, 19 built-in policy types, and 100 percent of governed actions logged. Though interestingly, the maker's launch note lists 20 policy types in one section, then says they ship 19 with sensible defaults — the site says 19.

Milo: The SDKs are TypeScript and Python, and they claim three lines of code can instrument any framework, with integrations listed for OpenAI, Anthropic, LangChain, LlamaIndex, CrewAI, AutoGen, MCP, and PostgreSQL. And there's a motivating scenario from the maker's own account — a support agent reads a ticket containing an instruction the customer didn't write, and within eleven seconds it exports data. That's the exact kind of autonomous action Execlave is designed to catch before it happens — which is really the through-line for all four of these launches today. The tools are getting faster at writing and running code, so the next frontier is making sure someone — or something — actually watches what they do.

Mia: Let's start with one of the more ambitious products this round. Pickle Browser launched on Product Hunt as what its founder calls an "agent-first browser" — meaning it's a real browser built specifically to run AI agents, rather than a wrapper around an existing one. The founder, Ismail, frames the problem directly: agents burn tokens reading raw HTML, so he built a browser that simplifies the page content before the model ever sees it.

Milo: And the compression numbers are the most striking thing here. The makers claim the built-in page compression cuts token use by thirty-two times, and across ten real sites they report a median 156 times fewer tokens than raw HTML — ranging anywhere from four times to over six hundred. One example shows a raw page that would cost thirty-eight thousand tokens being read in around twelve hundred. To put it in scale, a single modern page is often thirty to sixty thousand tokens of raw HTML.

Mia: Right, so the whole pitch is that it runs locally on your machine with your real logins, every action is gated by policy, and anything that purchases or destroys requires explicit approval. Everything is logged so you can watch live and jump in and take over mid-task. That's aimed squarely at the pain point of agents acting blindly.

Milo: The founder distinguishes it from cloud-based AI browsers, which he calls headless and remote and starting from a blank session each time — Pickle claims a full browser instead, real Chromium with tabs, history, bookmarks, and persistent element memory across sessions, plus a shared notebook for the agent and the user. It's free and works with Claude Desktop, Cursor, VS Code, Codex CLI, and anything that speaks the MCP protocol. A word of caution, though: these compression numbers are self-reported. The strongest takeaway is who it's for — anyone running an agent that needs to act in a browser and actually see what it does before it acts.

Mia: Moving to another browser agent with a different philosophy. WebBrain is a free, open-source browser agent from Emre Sokullu, launched on Product Hunt. It runs as a sidebar extension for Chrome, Firefox, and Edge, it's MIT-licensed, and the maker's core argument is that the most capable browser agents are usually closed source, tied to specific models, and locked behind subscriptions — so he built an open one.

Milo: The big differentiator is privacy. It can run locally through llama.cpp, where the maker says most queries cost nothing and data never leaves your device. Or you can bring your own API keys for OpenAI, Claude, OpenRouter, and over a hundred OpenAI-compatible providers. Sokullu also points to his track record — Grou.ps grew to three hundred fifty thousand communities and twelve million registered members before being acquired, and another project, GraphJS, was acquired by its first enterprise customer about a year after launching on Product Hunt.

Mia: On the control side, the default is a read-only "Ask Mode" — it asks before any consequential action, defends against hidden prompt injections, and claims no telemetry, no accounts, and fully offline operation. Optional features are off by default, including local plaintext profile autofill and a CapSolver key for CAPTCHAs. It also dismisses consent banners and reports paywalls honestly rather than bypassing them.

Milo: A few things to flag. Sokullu reports using WebBrain itself to draft parts of the launch post. And the marketing demos — form-filling that stops before submit, product data extraction, a DevTools styling suggestion — are unverified claims. What you can weigh is the product page's controls and guardrails: a genuinely open, offline-first agent browser that asks before it acts.

Mia: Now a tool for people spending real money on agentic coding. It's called Caveman — with the tagline "why use many token when few do trick" — and it's a one-command local proxy that wraps Claude Code, Codex, Hermes, and other agent CLIs. It compresses logs, tool output, and files before every provider call. The maker, Julius Brussee, frames his own origin as a university student spending more on AI than on groceries — which is the exact audience: heavy agentic-coding users paying per token.

Milo: The interesting part is the version history. The earlier version was a free skill that made Claude Code respond in caveman phrasing — one example took a response from sixty-nine tokens down to nineteen. But the maker says that only changed the answers, not the wire traffic: every turn still re-sends tool schemas, file contents, old history, and what he calls three hundred lines of npm install noise. The new v2 actually compresses the provider traffic itself, not just the phrasing.

Mia: Because the compression is lossy, originals are written to a content-addressed store on disk before anything goes upstream, and the agent can pull the real bytes back with a single tool call. Workarounds if anything fails: a parse failure or a larger result means the bytes pass through untouched. It covers Claude Code plus Codex, Gemini CLI, Aider, opencode, Hermes, and OpenClaw, never touches config, and subscription OAuth passes through byte-exact.

Milo: The strongest evidence is a benchmark: a pinned run of fifty-four Claude Code sessions with 33.2 percent fewer provider-reported input tokens. The catch — these are self-reported by the maker, and lossy compression always carries some risk. But for anyone paying per token on agentic coding, a third fewer input tokens is the kind of number that changes a monthly bill.

Mia: The last one is from the learning platform Scrimba — a new feature called Scrimba Explain. Instead of answering a question with text, it turns the question into an instantly generated video tutorial. Per, Scrimba's CEO and co-founder, describes the idea as "as fast as a chatbot, as engaging as a video." You type a question and get a video-like response in two to three seconds, complete with animations, diagrams, images, and code, narrated by a voice.

Milo: The inputs are flexible — a plain question, uploaded files, links, or pasted code — and the output resembles a YouTube tutorial: images, code walkthroughs, animations, diagrams, voiceover, captions, and a visible cursor. The maker claims it's much faster than video generation models because it uses Scrimba's own DOM-based playback technology — that's a self-reported claim, not independently verified.

Mia: Explain is free for everyone during beta, and users have already created over twenty thousand explainers. Availability goes beyond the web app — there's a ChatGPT plugin, MCP support for coding agents like Codex, Claude, and Cursor, a Chrome extension, and a GitHub integration that generates video explainers of pull requests. Per says that GitHub feature has been, in his words, "a godsend" internally for reviewing PRs.

Milo: And the community reactions are concrete. One programming-education creator asked how the frontend sends data to the backend and was impressed by the speed, the diagrams, and a follow-up comprehension question — meaning it quizzes you after the video. Another used it on their own code base, one used it to learn MedusaJS, and Scrimba says it wants to keep costs low enough to maintain a generous free plan for students. That's the piece to watch — whether free-to-use video explanations can hold up as the beta user base grows.

Mia: Kin Health is going after a very human problem, and the app is built around a simple promise: completely free, always private. It records your doctor visits and then turns them into structured summaries, so you can actually remember what was said and what you're supposed to do next. The workflow is straightforward, four steps. You record the visit, you get a summary with next steps, you share it with a care circle, and you prep questions before the next appointment. There's even an in-app reminder before each visit so you can show up ready.

Milo: And it's worth pointing out what those summaries are not. The site is careful to say they're not word-for-word transcripts, but organized key points, recommendations, and next steps. So you get the useful substance, not a wall of text. Kin was founded by physicians, it's free for patients, it works with any provider, and the data is encrypted and secure. The site also points to a 4.8 rating on the App Store.

Mia: There's a telling community comment that gets at why this matters. Someone described using Kin after a mother-in-law's breast cancer diagnosis, when the real bottleneck became getting accurate information about what happened during appointments out to family spread across multiple cities and states. They called Kin the first doctor-built, clinically rigorous, magical ambient scribe built for patients and their family. The testimonials echo that, from a migraine patient, a caregiver, an aging parent, and a primary care physician praising clarity, reduced anxiety, and easier family sharing.

Milo: And the site leans on two stats to frame the problem. Patients forget up to eighty percent of what their doctor tells them within minutes of leaving the exam room, and missed medications and skipped follow-ups add up to roughly three hundred billion dollars in unnecessary healthcare costs every year. So Kin is positioning itself squarely against that memory gap, aimed at caregivers, people with chronic conditions, new and expecting parents, pet owners, and really anyone managing care for themselves or someone else.

Mia: Now switching to something completely different — Dishylink is a free, open-source Starlink monitoring app for macOS, Windows, and browsers. The maker's pitch is that Starlink's official app is phone-only, and the old web portal was switched off in 2024 with no replacement. So desktop users were left with nothing. Dishylink reads the dish and router directly over your own network, and the maker emphasizes it needs no account, has no backend, no cloud, no analytics, is MIT-licensed, and has no paid tier.

Milo: So what does it actually track? Live throughput, latency, obstruction and hardware health, and dish alignment shown as a correction in degrees. There's a three-dimensional view of the satellites overhead, including the one your dish is connected to, plus a time-lapse of obstructions, per-device data usage, and real-time power draw with daily and monthly history. It also has alerts and an event log covering outages, overheating, and weather, and those clear themselves. The speed test is built in and measures against Cloudflare.

Mia: One detail that shows real thought about risk: on supported firmware it can control snow melt, sleep schedules, update windows, and reboots. But things like custom DNS, bypass mode, and content filtering were deliberately left out — because a bad write could take your WiFi down until you physically reset the hardware. That's a sensible call.

Milo: A few practical details. Version one point zero targets macOS twelve and up, Windows ten and up, and recent Chrome, Edge, and Firefox. It reads the standard local API that Starlink terminals serve, so it covers Gen 1 through Gen 3, Mini, and Performance kits, with motorized kits getting stow and unstow controls. Just keep in mind the project is unofficial and not affiliated with SpaceX — but for desktop users who want proper visibility into their connection, it fills a gap Starlink itself left open.

Milo: That's a wrap for today's show. Great lineup — from Ito, the code reviewer that actually runs your code before it judges it, to Pickle Browser's agent-first spin on browsing, plus the Caveman proxy for saving tokens and Scrimba Explain's instant lesson generation.

Mia: And if you want to keep the practical theme going, we touched on Kane CLI for terminal-based quality checks, Kin Health for turning doctor visits into clear summaries, and open-source options like Phinq, WebBrain, and the Dishylink Starlink monitor. Plenty to try out this week.

Milo: Thanks for tuning in — see you next time.