I built a free, open-source alternative to Claude Cowork and Manus Desktop
I got tired of paying $20–$200/month for AI agents that send all my data to the cloud. So I built EverFern — a local-first desktop AI agent that uses your computer the way you would. Free forever, MIT licensed, nothing leaves your machine.
The problem
Claude Cowork is $20+/month. Manus Desktop is $200+/month. Both are closed source and cloud-processed.
I wanted something that could click buttons, navigate apps, fill forms, and run workflows in plain English — but without a subscription and without my data going anywhere.
I couldn't find it. So I built it.
What EverFern does
You type something like:
"Open Spotify and play my liked songs"
"Summarize all the PDFs in my Downloads folder into one doc"
"Organize my pictures so that pictures of my face in the images should be a in folder called Codenrust"
"Research the top 5 AI coding tools and make a comparison spreadsheet"
EverFern breaks it down, shows you its thinking in real time, and executes — pausing before anything destructive.
It has:
- 🖥️ Full computer use (sees your screen, moves the mouse, clicks, types)
- 🌐 Navis — a built-in browser agent for navigating websites and filling forms
- 📄 Document processing (PDF, Word, Excel, PowerPoint, CSV)
- 🤝 Peer Agent Debate — multiple agents challenge each other's plan before executing
- 🔒 Linux VM sandbox so shell commands can't break your system
- 🔌 10+ AI providers — OpenAI, Anthropic, DeepSeek, Gemini, Groq, Mistral, Ollama, LMStudio, and more
Why free and open source?
Because this kind of tool shouldn't require a subscription. Your API key, your machine, your data.
All history and keys are stored in ~/.everfern/store — never synced anywhere.
Where it's at
We just shipped v0.1.6 Beta. Windows installer is live. MacOs installer is also available.
20 GitHub stars so far, 3 forks, small but growing community on Discord.
GitHub: github.com/Everfern-AI/Everfern
Website: everfern.app
Would love your feedback
- What workflows would you actually use this for?
- What's missing before you'd try it?
- Anyone else building in this space?
Drop a comment, and star the repo or come say hi on Discord. This is early, but it works, and I'm shipping fast.
The local-first angle is genuinely underserved. Most agent tools treat cloud processing as a given and privacy as a premium feature you pay extra for. Flipping that default makes sense.
The Peer Agent Debate feature is the most interesting part to me — having agents challenge each other's plan before executing is a real solution to the confidence problem most single-agent systems have. They just execute without questioning whether the plan is actually correct.
Two things I'd want before trying it on real workflows: a clear undo/rollback for anything it touches, and some kind of dry-run mode where I can see exactly what it plans to do before it does it. The "pausing before destructive actions" is a good start but I'd want that granularity on everything, not just destructive steps.
Navis is the part that got me. I run scraping infra for work and it always breaks on sites that block datacenter IPs. An agent driving my own logged-in browser sidesteps all of that. To the site it's just me, clicking around.
One question decides it: when Navis hits a login, does it reuse my existing browser session or handle auth itself? If it's the former, you've quietly built the thing half of us have been duct-taping for years. Starred either way.
This is impressive -the fact that it runs locally is a big deal for anyone privacy-conscious. Most people don't realise how much context they're handing over to cloud-based tools just by using them daily. How does it handle tasks that require real-time data or web access? That's usually where local models hit a wall. Would love to know how you've approached that.
Yep, for real time data, we use the users browser and controls it and fetches the data, so basically no interaction between any cloud provider
The peer agent debate feature is what caught my attention. Multiple agents challenging each other's plan before executing — that's the same pattern I've been using for spec review.
One thing I'd add: give each agent a specific role (security, UX, performance) instead of having them debate generally. Role-based debate produces sharper pushback than generic 'challenge this plan'.
Good shit. Staring the repo.
Hey there, Thank you for the feedback, i will work on this
This is awesome, preetham! Love the local-first approach and the Linux VM sandbox feature for safety. I'm actually building an AI project focused on local system/cybersecurity protection right now, so this definitely hits home. Leaving a star on the repo!
Hey thank you soo much, if you wan can you help us by joining the discord server too! https://discord.gg/mtffm68Ge
What stood out to me is that you're changing the trust model more than the pricing model.
A lot of people assume the biggest tradeoff is subscription vs. free. It feels like the bigger question is whether users are comfortable giving an AI full control of their computer in the first place.
Running locally makes that decision feel very different.
You've reframed this better than I did in the original post.
The pricing argument gets clicks. But what you're describing is the actual product decision.
When an AI agent runs in the cloud, you're trusting:
When it runs locally, the trust chain collapses to one question: do you trust yourself?
That's a fundamentally different relationship with the tool. And I think most people haven't consciously made that choice — they've just defaulted to whatever was easiest to sign up for.
The Linux VM sandbox in EverFern exists precisely because of this. Even locally, full computer control is a big ask. So we isolate shell execution — the agent can't reach outside the sandbox unless you explicitly allow it. You stay in the loop.
I'd actually argue this is the thing that should be in the headline, not the price. "AI agent you actually control" is a more durable value prop than "cheaper than Manus." Price can be matched overnight. Trust architecture is a design philosophy.
Might rewrite the title after reading your comment.
I think this actually connects with the note I sent you.
Reading this reply made me even more convinced the underlying decision is a bit different from the one most people would take away from the thread.
It does connect — and I think you're both pointing at the same thing from different angles.
Your note was about capability vs. infrastructure ownership. This thread landed on trust architecture. But underneath both is the same question:
Who is actually in control when the agent runs?
With cloud tools, the honest answer is: not you. You're a user of someone else's system. The agent acts on your behalf but within their constraints, their logging, their uptime, their terms of service that can change next quarter.
With EverFern locally, the answer is: you, completely. Which sounds better until you realize that also means you're responsible when something breaks.
That's not a bug. That's the product. EverFern is for people who want that responsibility because they understand what comes with it — portability, privacy, no vendor lock-in, no surprise deprecations.
The people who don't want that responsibility aren't wrong. They're just not the customer.
I think the reason this thread has gotten interesting is that most AI agent marketing pretends this tradeoff doesn't exist. Everyone promises "powerful AND simple AND private AND reliable."
We're just being honest that you pick two.