One Big Law associate cloned Harvey AI in 2 weeks

Ajouté : 2026-05-26

[00:00:00]A big law associate looked at the 11 billion dollar legal AI category, sat down for two weeks, and shipped the entire thing as open source. The repo went viral, and partners at most of the top US law firms are about to read a very awkward renewal contract. His name is Will Chen. He spent three years as an associate at Laam and Watkins in London after qualifying from Oxford with a double first. On evenings and weekends, he wrote code. Then in late April, he uploaded a project called Mike under the AGPL3 license. Mike is a chat interface for legal documents. It reads contracts, drafts new ones, runs spreadsheet style review across hundreds of files at once, and ships preset workflows for the kind of work juniors burn nights on. The whole thing wraps the three frontier model providers behind a simple next.js front end and an express backend. Within days of launch, it had crossed a thousand stars. The growth chart has not slowed down, and the founder admits the response is bigger than he expected.

[00:01:05]Enjoying this so far? Subscribe to the YouTube channel and grab the free daily newsletter at indiehacker.news.

[00:01:13]New deep dive every weekday. Here is the landing page he put up at micos.com.

[00:01:19]Notice what the tagline says. The open-source alternative to Harvey and Lora. All the features without an enterprise contract. That second line is the whole pitch. The site is calm. Monae paintings as wallpaper. No enterprise sales copy. Scroll down and you hit the feature grid. Four cards. Assistant. The chat interface that reads documents and sites verbatim. Projects where you upload files and share them with your team. Tabular review. the bulk extraction engine that runs hundreds of documents in parallel and cites every cell back to a page. Workflows where you save proven prompts as reusable templates. The bottom of the page says your firm's AI on your own terms. Start in the cloud or clone the repo. That is the entire enterprise category listed as four free features on a static page. To understand why this lands so hard, you need to know what it is replacing.

[00:02:20]Harvey is the category leader. It runs inside more than 60% of the Amlaw 100, the largest firms in the United States.

[00:02:29]Legora is the European challenger out of Stockholm, last valued at 5.5 billion with over 1,000 firms on it, including linkers. Both sell enterprise contracts.

[00:02:40]The pricing is not on a website.

[00:02:42]Industry estimates put a midsized firm contract at over $200,000 a year with per seed fees on top. The actual moat was never the model. Both products wrapped the same API keys you can buy yourself from a model provider. The moat was being the polished interface and the trusted vendor. Will Chen is arguing that the polished interface is now a twoe project. The numbers on the launch are the part that gets a lot of attention. 72 hours after the repo went public, it had crossed a,000 stars and 300 forks. As of today, 20 days in, 3.1,000 stars, 912 forks, and a watch list filling up by the day, a GPL3 license, single contributor.

[00:03:29]Okay, so what is actually in the product? The assistant is a chat interface that reads contracts, spas, leases, anything you upload and writes verbatim citations back to the page and quote you asked about. No invented case law. The kind of thing every lawyer is paranoid about. Projects are mattercoped workspaces. You drop credit agreements, spas, diligence packs into a project, and the assistant carries full context across every conversation and every document inside it. That alone is what most associates spend their first two years learning how to keep track of in their head. Tabular review is the unlock. You define column questions like what is the governing law? Who is the indemnifying party? What is the termination notice period? And the system spawns parallel queries across hundreds of documents at once. Every cell is verifiably cited back to a page and a quote. No hallucinated answers, no dead links. The bulk review is the workflow that justifies the entire enterprise category and Mike does it in the open. The other half is workflows.

[00:04:40]Save a proven prompt as a reusable template and your juniors can run it in one click. Mike ships with a stack of presets out of the box that look like a snapshot of what gets built at a corporate practice. Change of control review, credit agreement review, eiscocovery review, supply agreement review, SPA review, NDA review, commercial lease review, limited partnership agreement review, shareholder agreement review, employment agreement review. 10 presets. Each one is a saved prompt the firm can edit, fork, hide, or share internally. The bring your own key model means you pay the model provider directly. No markup on tokens, no per seed fee, no annual minimum. If a junior associate runs 10,000 documents through bulk review for a diligence pack, the bill goes to your own account. The platform itself is free to host or self-eploy.

[00:05:38]Word and Outlook plugins are not in the build yet, and the founder is open about that gap. The stack is the boring part, and that is exactly the point. Next.js JS on the front end in Typescript, Express on the back end, Superbase for O and Postgress, NES3 compatible bucket for file storage with Cloudflare R2 as the recommended default. Libre Office for converting Word documents to PDF so the model can read them. Five database tables give or take. The model providers are interchangeable. Claude, Gemini or OpenAI configured per user or as a back-end default. Resend handles transactional email. The whole thing is 65% TypeScript AGpl3 licensed and small enough that you can read the entire schema.SQL file in one sitting. There are no proprietary embeddings, no custom legal training run, no fine-tuned legal model.

[00:06:39]The author's argument is exactly that.

[00:06:41]The fancy parts are not what makes Harvey work. The fancy parts are what makes Harvey expensive.

[00:06:48]Installing it is two commands. Clone the repo. Install dependencies for the backend and front end. Drop your Superbase URL, your R2 credentials, and at least one model API key into an AMV file. Run the backend on port 3001. Run the front end on 3000. And you have a self-hosted legal AI platform on your laptop. That is the entire installation, maybe 10 minutes if you already have a database project provisioned. 20 days into the project, the active commits are the kind of work you expect when a vibecoded launch meets actual users. The last merged pull request number 148, ships document UI improvements, tabular review caching, and storage caching. The pull request Q has 31 open. The issue tracker has 38 open. The founder is reviewing pull requests himself. The contributing guide and a safe local testing guide both landed in the docs folder within the first two weeks. The project has zero packages published, zero releases tagged, and the Redmi is 124 lines. The code base is 625 kilobytes. This is what a serious side project that hit escape velocity looks like in week three. Here is what you are actually watching. This is not going to displace anyone at Kirkland and Ellis.

[00:08:13]Big firms pay the incumbents for Westlaw integration, audit logs, and a vendor to sue when something goes wrong. None of that ships in this box, and the author says so himself. Where it lands is the middle of the market. The solo consultant who walked away from a quote last quarter. The five lawyer boutique that could never justify a per seat fee.

[00:08:34]The in-house team at a seriesB startup whose general counsel was told to figure something out by Q3. Those people just got handed an option they did not have a month ago. The harder question is what happens inside the incumbent companies.

[00:08:48]Their renewal meetings sound different starting today. The conversation shifts from is this magic to what exactly does the contract include that I cannot get from a clone. That is a much worse question for them to answer. So watch the next funding round at one of those companies. If it happens, it tells you the mode held. If it does not, it tells you something else entirely.