Launched this week
OpenOwl
Automate what APIs can't in one prompt done locally
343 followers
Automate what APIs can't in one prompt done locally
343 followers
OpenOwl is a desktop automation agent for macOS. It gives AI assistants (Claude, Codex, or any MCP-compatible AI) the ability to see your screen, click buttons, type into fields, and navigate across any app or browser. You describe a task in plain English. OpenOwl does the rest. It automates the tasks that APIs can't touch LinkedIn prospecting, Shopify admin updates, legacy CRM data entry, form filling, competitive research, and anything that normally requires you to sit there clicking for hours
OpenOwl
Hey Product Hunt! Mihir here, maker of OpenOwl.
I built this because of a frustration I couldn't shake:
I'd ask Claude to help me with a task, and it would give me perfect step-by-step instructions... that I then had to spend 45 minutes clicking through myself.
LinkedIn prospecting? "Go to this profile, click Connect, type this message." Great advice. But I still had to do all the clicking.
Updating 200 product prices in Shopify? Claude knew exactly what to change. But there's no API for the admin panel. So I sat there. Clicking. For two hours.
AI can think for you. But it couldn't act for you. That's the gap.
OpenOwl is an MCP server that gives your AI assistant actual eyes and hands on your screen. It sees your screen, moves the cursor, clicks buttons, types into fields, and navigates between apps — not through APIs, but through the actual UI, like a human would.
You describe what you need in plain English. OpenOwl does the rest.
A few things worth knowing:
Works with Claude, Codex, and any MCP-compatible AI
Runs 100% locally on your Mac — screenshots and data never leave your machine
Install with one command: npm i openowl
Free tier included — 50 tool calls/day, no credit card needed
macOS only right now (Apple Silicon + Intel)
This started as a weekend project to scratch my own itch. Now it's something I use every single day — and genuinely can't go back to doing these tasks manually.
What repetitive screen task would you hand off to your AI if it could actually click for you? Would love to hear what you'd use this for.
@mihir_kanzariya he Shopify admin example landed for me. There’s a lot of work sitting in that gap where the model already knows what to do and a person still has to sit there clicking through the UI. That’s a very real kind of wasted time. How often are people using OpenOwl for browser tasks like prospecting, and how often for messy back-office stuff like Shopify or CRM work?
OpenOwl
@artem_kosilov That's the gap right there. AI knows what to do. You're still the one clicking buttons.
From what I'm hearing, it breaks down two ways. Browser stuff like prospecting and outreach happens every day. Quick wins, high volume. Back-office work in Shopify, CRMs, random internal tools happens less often but eats way more time per session.
Honestly the back-office side is more interesting to me. That's where automation usually dies. No APIs, messy UIs, edge cases everywhere. Nobody builds integrations for your company's janky admin panel. OpenOwl works there because it doesn't care. It just clicks through the UI like you would.
What's the workflow eating most of your time right now? Walk me through it and I'll show you how I'd set it up.
@mihir_kanzariya Happy launch day. Very cool project. I wonder if you got affected by the latest Claude blocking third party app restriction? Hope not.
OpenOwl
@syaman It does not block for me, my macmini run 24/7 on autopilot mode, if it does it just need simple back and forth to use explicitly use openowl in anycase.
@mihir_kanzariya is this not just claude co work....?
OpenOwl
@stevenlawx Fair question. Claude computer use gives Claude a virtual screen inside Anthropic's cloud. It's a sandboxed environment, not your actual machine.
OpenOwl runs locally on your Mac. It controls your real screen, your real apps, your real browser sessions. Logged into Shopify? It can edit your products. Have a CRM open? It can fill in fields. It works with whatever's already on your desktop, no setup, no sandboxing.
The difference matters because most of the tedious work people want to automate lives in apps that only run locally or need your logged-in session. Cloud-based computer use can't touch that.
I run mine on a Mac Mini that stays on 24/7. That way it handles tasks in the background without taking over my main machine. Recommend the same if you plan to use it regularly.
This is cool - can you tell me a bit about where it's more powerful than Claude Cowork or Perplexity's "browse for me" functions, or any major LLM with Playwright or similar installed? Again, really cool either way, definitely going to give this a swing! congrats on the launch :D
OpenOwl
@grey_seymour Thanks! The main difference is scope. Cowork and Perplexity browse-for-me operate inside a browser. Playwright is browser-only too.
OpenOwl controls your entire desktop. It can switch between Slack, Excel, your CRM, a browser tab, and a terminal in the same workflow. It also uses your logged-in sessions, so there's no re-authenticating or dealing with OAuth flows.
If your task lives entirely in a browser, those tools work fine. If it touches multiple apps or needs your real desktop, that's where OpenOwl comes in. Hope you enjoy it!
Congrats on the launch! The local-first approach is smart, and browser automation that doesn't phone home is a real unlock for privacy-sensitive workflows. How does it handle sites with aggressive bot detection?
OpenOwl
@kavin_jeya Thanks! Yeah the local-first piece was non-negotiable. Screenshots and screen data stay on your machine, nothing gets sent anywhere except your normal Claude conversation.
On bot detection, because OpenOwl controls the actual screen through OS-level inputs, it looks identical to a human using a mouse and keyboard. There's no headless browser or injected scripts for sites to detect. To the site, it's just you clicking around.
most workflow automation tools nail the happy path. partial failures are where they fall apart - one step errors, the rest queue up, nobody notices until the downstream data is wrong.
OpenOwl
@mykola_kondratiuk
Yeah that's a real problem. Most automation just stops or silently keeps going with bad data.
The way OpenOwl handles this is different because it's not a scripted workflow. Claude is looking at the screen after every action. If something errors out, a popup appears, or a page doesn't load right, it sees that and reacts. Same way you would if you were doing it manually and something went wrong.
It's not bulletproof, but it's closer to how a person handles partial failures than a traditional automation chain where step 3 doesn't know step 2 broke.
Glad the silent-failure angle landed. That mode - silently continuing with bad data - erodes trust faster than a hard crash. At least errors are visible.
OpenOwl
@mykola_kondratiuk Exactly — silent failures kill trust way faster than crashes.
That’s what we’re trying to fix with OpenOwl: instead of blindly executing, it verifies each step and flags when something looks off.
Goal is simple — automation you can trust, not just automation that “ran.”
Yeah - most teams figure out they needed observability after something already went wrong, not before. Built-in is the right call.
The LinkedIn prospecting use case is the one I'd stress-test first. LinkedIn actively fights automation — rate limits, CAPTCHAs, shadow restrictions on accounts that move too fast. Since it's running locally and watching the screen like a human would, does it pace itself to mimic natural browsing behavior, or does that throttling fall on the user to configure?
OpenOwl
@sounak_bhattacharya Good question. The pacing is on you to configure. Claude is doing the clicking but you tell it how fast to go. I'd always tell it to add delays between actions and keep volumes low.
The advantage is that it's using your real browser, your real IP, your real mouse movements. There's no headless browser fingerprint to flag. But LinkedIn will still notice if you connect with 200 people in an hour, regardless of how it's done. So yeah, keep it human-speed and you're fine.
Cool, Mihir! Currently dealing with that problem cause Calude perfectly set a plan for me but then I want to execute it, which should be made for an agent. Happy to see you helping on this!!
OpenOwl
@german_merlo1 Sure, let me know your use case and I can help you set it up. I've built some templates here: https://github.com/mihir-kanzariya/openowl-templates.
If you have a different use case, I can craft a solution for you and even jump on a call to help you get it running.
OpenOwl
@lovecraft_ack we are working on linux and windows.