A former Latham & Watkins lawyer, Will Chen, built an open-source legal AI platform called mike in two weeks of vibe coding with Claude. The repository was created April 29, licensed AGPL-3.0, and as of the October 6 GitHub snapshot had 4,305 stars and 1,330 forks, with commits as recent as the previous day. It positions itself against Harvey and Legora, the two leading commercial legal AI platforms.
Verified features and stack
Each of the seven core features checks out against the README:
- Chat with documents
- Review with suggested edits
- Reusable table-based review workflows
- Projects and document libraries
- Citation verification and US case law search via CourtListener
- Microsoft Word add-in (beta)
- Local models via Ollama
It's a real platform. But 113 Hacker News comments delivered the most important cut.
The asterisk on "Documents never leave your perimeter"
mike's marketing line is "Documents never leave your perimeter." An HN commenter dismantled it cleanly: this is only true in one case—if you don't send the documents to Anthropic or Google.
What you self-host is the platform shell. The model hop, by default, still goes through the Anthropic, Gemini, or OpenAI API—the README's quick start explicitly requires a vendor key. The shell sits in your firm's server room; the inference happens in Silicon Valley. The only true closed loop is running models locally with Ollama, at the cost of capability. The accurate phrasing is: *documents never leave your perimeter—except for the model-call hop.* That parenthetical isn't nitpicking; whether confidential workflows can run on this system depends entirely on it.
Two more cuts from the debate
The second cut came from a practicing lawyer: "What legal professionals actually pay for is the AI's access to case law databases, which is nearly impossible to replicate." mike's legal search uses CourtListener—a free US case law resource with decent federal coverage—but it lacks the century of editorial value Westlaw and Lexis accumulated: headnotes, KeyCite citation maps, judge annotations. The platform layer can be carved out in two weeks; the database layer cannot.
The third cut swung back at commercial vendors. Another commenter pushed in reverse: foundation model providers already offer zero-retention enterprise tiers and data residency configurations—how much of the old narrative that "law firms pay 95% for Harvey/Legora's data security buffer" still holds? If Anthropic promises no retention, what's the marginal value of another data-security shell on top? Together with the second cut, this splits legal AI's moat into three pieces: the platform layer (collapsed, replicable in two weeks), the model layer (owned by foundation providers), and the data layer (Westlaw's century of inventory). The middle layer's expensive "security premium" is being squeezed from both ends.
One caveat: a claimed list of "eleven hidden costs that don't disappear, just move from the vendor's payroll to the firm's" could not be traced to an original source. The cost types visible in HN comments (database subscriptions, maintenance patches, compliance audits, maturity risk) do support the direction, but the precise count of "eleven" is treated as the original source author's own summary.
The bigger picture
This is another domain data point in the deployment-collapse spectrum. The vertical SaaS playbook is repeating: when foundation model vendors turn data security into an API feature, and vibe coding compresses the platform layer to two weeks, the "security premium" collapses. What remains are data inventory and workflow lock-in—Harvey's real answer doesn't live in the inference layer.
The other angle is verification bandwidth: mike uses CourtListener for citation verification. AI-generated citations are assertions; a case law database lookup is enforcement. The legal industry's tolerance for hallucination is near zero, so this gate is a hard requirement.
Two observations from a cited practitioner persona hold up independently: AGPL-3.0's strong copyleft is a genuine barrier for commercial use—law firms aren't software companies, and modifying the code creates open-source obligations. And "start with non-confidential workflows" is exactly the boundary line drawn by the asterisk above.
A falsifiable prediction: within twelve months, vertical SaaS sales pitches will broadly shift from "security and compliance" to "data enrichment and workflow integration." If that doesn't happen, it means zero-retention enterprise tiers aren't actually valuable in real procurement.
The most level-headed HN comment is a good place to end: "Self-hosted legal AI is in principle a useful direction—it's just that the repo is very new, and the marketing site did a lot of heavy lifting." How many of those 4,305 stars turn into real case files overnight won't be known until the first due diligence report completed with mike appears.
---
Sources: GitHub API (open-legal-products/mike: 4,305 stars / 1,330 forks / created 2026-04-29 / pushed 2026-10-06 / AGPL-3.0, fetched early 10-07); full README (all seven features and deployment claims checked line by line); HN thread 47956739 (209 points, 113 comments, full API pull: perimeter debunk, database moat, zero-retention counter, maturity critiques quoted verbatim); Artificial Lawyer (5-04); legaltechnology.com (5-05, "vibe coded" phrasing); legalfutures.co.uk (5-13); davidkemp.ai (5-12, "Institutional Floor"). The "eleven hidden costs" list could not be traced to its original source and is treated as the original author's summary, flagged accordingly. The cited practitioner observations were independently verified before inclusion.