On May 5, 2026, a former Latham & Watkins associate named Will Chen pushed an AGPL v3 repository to GitHub and wrote a single sentence on LinkedIn: he had spent two weeks rebuilding the core of Harvey and Legora, and he was giving the code away. By the end of the week the repo at github.com/willchen96/mike had crossed 2,200 stars and 600 forks. The project is called Mike OSS, the live demo lives at mikeoss.com, and U.S. legal-tech buyers should care about it for reasons that have nothing to do with whether they install it.
The headline numbers tell most of the story. Harvey just raised at an $11 billion valuation. Stockholm-based Legora just closed a Series D at $5.6 billion. Between them, the two companies are valued at more than the combined market cap of three large AmLaw 100 firms. Chen’s argument, in source code rather than English, is that the product underneath those valuations can be rebuilt by one unemployed lawyer in fourteen days.
That argument has already moved through Artificial Lawyer, Legal IT Insider, and a 1,300-comment Hacker News thread. What follows is the version aimed at managing partners, general counsel, and legal operations leaders inside U.S. firms who need a real read on what Mike OSS is, what it is not, and what it should change about the next renewal conversation.
What Mike OSS actually is
Mike OSS is a TypeScript web application that replicates the core lawyer-facing workflows of Harvey and Legora: a chat assistant that drafts and reviews documents, an encrypted “vault” for matter files, a tabular review mode for mass document analysis, and saved workflows that bundle prompts, column sets, and output templates into reusable presets. The name is a direct reference to Mike Ross from Suits; the “OSS” suffix marks it as open source software.
The license is the part that matters legally. Mike ships under the GNU Affero General Public License v3. A firm can read, modify, and deploy the code internally without publishing anything. The moment a firm exposes a modified version as a service to anyone outside the firm, AGPL v3 requires it to release the modifications. The clause is designed specifically to prevent a Harvey or a Legora from forking Mike, closing it, and reselling it under a proprietary banner. For a U.S. firm running an internal install for its own attorneys, the obligation never triggers.
The “bring your own API key” model, explained for U.S. buyers
Mike does not bundle a model. The software does inference by calling either Anthropic’s Claude API or Google’s Gemini API, and the firm must supply its own key. The provider bills the firm directly. There is no subscription to Mike OSS and no margin on tokens; whatever the model costs is what the firm pays.
For a U.S. firm, this changes the budgeting conversation from “what is the per-seat license fee” to “what will inference cost across our matter mix.” A boutique twenty-attorney litigation shop doing heavy document review on Claude Sonnet can land between $400 and $1,500 a month in token spend during an active matter, sometimes higher in document-heavy litigation. That is meaningfully less than the enterprise contracts the larger legal-AI vendors are quoting to small and mid-sized firms, which typically start in the five figures annually and rise from there.
The non-obvious benefit, for a firm running its own privilege analysis, is that the data path becomes legible. Every prompt Mike sends to Anthropic or Google is visible on the provider dashboard the firm controls. The firm sets retention, no-training, and zero-data-retention posture directly with the model provider under a contract the firm signs. There is no intermediate SaaS layer caching, logging, or sub-processing the data before it reaches the model. We walk through why that matters for attorney-client privilege and AI vendor due diligence in detail, and Mike OSS happens to land cleanly on the right side of that analysis when it is self-hosted.
The Harvey and Legora features Mike actually replicates
Will Chen’s claim is feature parity on the four modules that drive the majority of attorney usage at firms that have deployed Harvey or Legora. Spending an hour in the live demo, the parity is real for daily drafting and review work, partial for matter-wide workflows, and absent for anything that touches proprietary case-law corpora.
1. Conversational assistant
The chat interface is the workhorse. Mike can draft a memo from a fact pattern, summarize a deposition transcript, redline a contract clause, compare two versions of an NDA, and produce a first-pass research memo grounded in pasted authorities. Chen claims Mike actually beats Harvey on tracked-changes review across multiple documents in parallel, which has been a recurring complaint in user reports of the commercial product. On a small-N test with five real M&A redlines the rendering was clean and the diff logic legible. It is not a magic act; it is competent.
2. Projects vault
The vault is an encrypted matter file store. Attorneys tag documents by client and matter, share them through internal links, and Mike maintains an index the chat assistant can retrieve from. The backing storage is S3-compatible — Chen recommends Cloudflare R2 because R2 has no egress fees and runs about $1/month per 100 GB. Metadata sits in Postgres through Supabase. When self-hosted, the documents live wherever the firm tells them to live; nothing leaves the firm’s perimeter except the specific prompt content the firm chooses to send to the model.
3. Tabular review
The tabular review mode is the feature that most impressed the early reviewers, and for good reason. The user uploads a stack of contracts, defines column headers as questions (“What is the governing law?” “Is there a non-compete clause and what is its term?” “Does the indemnity carve out gross negligence?”), and Mike fills the grid by running each cell as an independent query. The result exports to Excel and is, on its face, a working due-diligence tool — the exact workflow Legora has built much of its enterprise narrative around.
Two caveats for U.S. buyers. First, accuracy on tabular review depends on careful prompt engineering and column-level validation. Drop two hundred franchise agreements in and ask ambiguous questions, and you get ambiguous answers. The same is true of Harvey and Legora; vendors who claim otherwise are selling. Second, this is the exact workflow where evals are the real competitive moat. An open-source clone gets you to a working table view; getting to a table view your partners trust on real M&A diligence is a separate, longer engagement.
4. Workflows and shared presets
Workflows are reusable bundles: a risk-analysis prompt, a column set for tabular review, an output template. A senior associate builds the workflow once; a junior attorney runs the same playbook with one click. This is the feature most likely to disappear into firm-specific customizations the moment a firm seriously adopts Mike, because workflows only earn their keep when they encode the firm’s actual review standards. Which is exactly why generic legal-AI platforms struggle to deliver durable value — and why a custom build wins for the workflows that define a firm’s practice.
The U.S. tech stack and what it takes to install
The stack is intentionally boring. The frontend is a Next.js application in TypeScript. The backend is an Express service. Authentication and the relational store run on Supabase Postgres. Document storage is any S3-compatible bucket. Office-to-PDF conversion uses LibreOffice invoked from the command line, which means the backend host needs the LibreOffice binary installed. Everything else is standard Node 20+ tooling.
Deploying for evaluation is a 30-minute exercise for any firm that has a developer on retainer. Chen has also published a Railway one-click template that runs about $5–$10/month for the runtime — fine for a private demo, but not the deployment a firm should use for actual client matter. For a production install inside a U.S. firm, the right surface is the firm’s own cloud account (AWS, Azure, or GCP) under the firm’s own VPC, with the model provider relationship contracted directly between the firm and Anthropic or Google. We outline what that buildout costs end-to-end in the real cost of custom AI for a 20-attorney law firm.
Where Mike OSS falls short for U.S. firms
The early enthusiasm has run into pointed criticism on Hacker News, in private legal-tech Slack channels, and from inside firms that ran Mike against real workloads in the first week. The criticism is fair, and it lands on three specific gaps that any U.S. firm should treat as gating questions before adoption.
No integrated case-law database
This is the biggest gap and the hardest one to close. Mike replicates drafting and review, but it does not connect to Westlaw, Lexis, Bloomberg Law, or Fastcase. For litigation work — brief writing, motion practice, anything that requires finding controlling authority in a jurisdiction the firm does not already have memorized — Mike cannot replace a research subscription. The case-law incumbents hold near-exclusive contracts with U.S. courts and publishers; no open-source project is going to negotiate those rights in the short term.
For due diligence, contract review, transactional drafting, intake triage, and any internal-document workflow, the case-law gap does not matter. For litigation citation work, it is dispositive. A firm using Mike for the wrong workflow risks missing a controlling case in a way that has serious professional-responsibility consequences.
The repository is a prototype, not a product
At release, Mike had two main commits, eighteen open issues, fifteen pending pull requests, no documented multi-tenant story, no SSO/SAML integration, no enterprise audit log, and no SLA. Chen acknowledges this directly in his Artificial Lawyer interview, where he describes Mike as “production for one-person firms, work in progress for enterprise.” That is honest. It also means an AmLaw 100 firm cannot adopt Mike OSS as a Harvey replacement next quarter; the wrapping layer that turns a working app into an enterprise product is not there yet, and may take a year of community contributions to mature.
Privilege and ethics still live with the firm
Self-hosting Mike OSS narrows the privilege surface but does not eliminate it. The moment a prompt leaves the firm and hits the Anthropic or Google API, the firm is sending privileged communications to a third-party service. That is a question ABA Formal Opinion 512 addresses directly, along with the related guidance from California, Florida, New York, New Jersey, and Texas state bar ethics committees. The duty of competence (Model Rule 1.1), confidentiality (Model Rule 1.6), and supervision (Model Rule 5.3) all attach. Adopting Mike does not change any of those duties; it simply moves the diligence work from the SaaS vendor to the model provider.
The good news for U.S. firms is that both Anthropic and Google publish detailed enterprise data-processing terms. Anthropic offers zero-retention API processing on paid commercial tiers; Google offers comparable terms under its Cloud data processing addendum. Any firm running Mike in production should sign the addendum, document the data path, and walk the firm’s AI-use policy through the same review framework we describe in our law-firm AI-use policy framework.
Mike OSS vs Harvey vs Legora: the comparison U.S. buyers should run
Comparing a one-person open-source project to two nine-and-ten-figure private companies feels asymmetric. It is also exactly the comparison Chen wants you to run, because the point of Mike OSS is to make that comparison legible. The relevant axes for a U.S. legal-tech buyer:
- Software license cost. Mike: $0. Harvey and Legora: enterprise contracts, with market reports putting mid-firm deployments at $200K–$600K annually depending on seat count and module scope.
- Model cost. Mike: pass-through, paid directly to Anthropic or Google. Harvey and Legora: inference is bundled inside the per-seat fee, which is the single biggest reason their gross margins look like SaaS while the underlying cost of goods looks like an AI startup.
- Deployment model. Mike: self-hosted in the firm’s cloud (or on-prem, if the firm still runs infrastructure). Harvey and Legora: multi-tenant SaaS, with some enterprise customers offered single-tenant deployments at a premium.
- Case-law integration. Mike: none. Harvey and Legora: integrations with Westlaw, Lexis, and selected regional databases, depending on the customer’s existing subscriptions.
- Customization. Mike: full source-code access; firms can fork, modify, and integrate at any layer. Harvey and Legora: configurable within the bounds the vendor allows; no source access, no architectural changes.
- Maturity. Mike: prototype, two months old at release. Harvey: production since 2022. Legora: production since 2023. The gap on tooling, SSO, audit logging, and enterprise admin surfaces is real and not trivial to close.
- License. Mike: AGPL v3 (firm-favorable for internal use). Harvey and Legora: proprietary, with the full set of vendor-lock-in clauses the buyer should expect.
The honest summary is that Mike OSS is not a drop-in replacement for Harvey or Legora at a 200-attorney firm with a complex matter-management environment. It is a fully credible replacement for the Harvey-style workflow at a 5- to 25-attorney boutique that has been quoted enterprise pricing and walked away. And, more importantly, it is a working demonstration that the generic-platform thesis underneath the entire legal-AI category has gotten thin enough to clone in two weeks.
Why “changes the negotiation” is the right phrase
Caroline Hill of Legal IT Insider captured the second-order effect with one line: Mike OSS “changes the negotiation.” The argument every legal-AI vendor has used since 2023 — “our product is worth the price because rebuilding it would cost millions of dollars and years” — loses force when a former associate distributes a working clone for free. Adoption is not the point. The existence of a credible open-source baseline reprices the commercial product, even at firms that never install the baseline.
We have seen this dynamic before in other categories of enterprise software. Postgres reset the price of Oracle. Linux reset the price of Solaris. Kubernetes reset the price of every proprietary orchestration layer that preceded it. None of those open-source projects took 100% market share — many large enterprises still pay for Oracle. But all of them forced the commercial alternatives to justify their pricing in a way the market had not previously required. Mike OSS is doing the same thing to Harvey and Legora.
Will Chen’s thesis, in his Artificial Lawyer interview, is sharper: “Thin wrappers without unique added value will get replaced by the major model providers themselves. Thick wrappers with a real value proposition will survive.” That maps onto a framework we have been writing about under the heading of Software 3.0: the era when the model is the platform, the application is a workflow encoded as an agent, and the durable value sits at the workflow layer, not the chat-app layer.
What this means for U.S. law firms in the next twelve months
There are five concrete moves a U.S. firm should make in response to Mike OSS. None of them require adopting it.
1. Reopen any open Harvey or Legora negotiation. The pricing power on the vendor side has measurably shifted. Firms reporting back from the first week say discount offers have grown noticeably more generous since May 5. Procurement teams should be aware of the new ceiling.
2. Audit existing legal-AI subscriptions against actual usage. Many firms are paying per-seat fees on platforms that fewer than 30% of their attorneys actually log into in any given month. Mike OSS being free does not change the calculus on a tool nobody uses, but it does change the calculus on a tool the firm is paying $300K/year for and getting marginal adoption from.
3. Pilot self-hosted Mike OSS on a single non- privileged workflow. The lowest-risk way to learn what an in-perimeter legal AI feels like is to install Mike OSS on a firm-internal dataset that does not involve client matters: research memos, internal CLE materials, the firm’s own contract templates. The learning curve on deployment, IT support, and partner adoption is real, and the non-privileged pilot is where to absorb it before moving to matter data.
4. Use Mike OSS as the reference architecture for your custom build. Mike is, charitably, 60% of a real legal-AI product. The remaining 40% is the part that matters: firm-specific document types, partner-specific review standards, integrations with the firm’s document management and matter management systems, the evals that prove the system works on real matter files. That 40% is where a focused two-week AI sprint actually earns its keep — and Mike OSS gives every firm a running start on the first 60%.
5. Update the firm’s AI-use policy to reflect BYOK reality. The bring-your-own-key model is going to become normal in legal AI over the next two years, and firms whose policies were written assuming a single SaaS vendor relationship need to update them. The right starting point is our 2026 AI-use policy framework, which covers the model-provider direct-contract posture in detail.
The Brightline read: Mike OSS is a Trojan horse for ownership
We sit on a particular side of this argument and we will say it plainly. Brightline Labs ships custom AI modules for law and accounting firms, hands the source code to the firm, and does not run a SaaS platform. The economics that make Mike OSS interesting — pay-as-you-go inference, source-code ownership, no per-seat license, deployment inside the firm’s perimeter — are the same economics our entire business model is built on. Mike OSS does not compete with what we do; it accelerates the case for it.
Two specific things we’ve seen in conversations with partners over the past week:
First, Mike OSS makes the “build vs buy” conversation easier to have. Partners who would never have agreed to a custom build a year ago are now treating Mike as evidence that legal AI is buildable. The follow-up question becomes “okay, can someone build the version that actually fits our practice?” That is a conversation about scope and evals, not about whether the underlying category is real.
Second, Mike OSS makes the pricing math we already wrote about more concrete. When the SaaS price tag was $300K/year and the custom-build alternative was $80K, partners hesitated because $80K still felt like a lot of money. When the SaaS price tag is being justified against a free open-source baseline plus $1K/month in tokens, the same $80K custom build looks like the obvious move — because what you are buying is the part Mike OSS doesn’t solve: the firm-specific fit, the integrations, the evals, and the long-running care of a production-grade system.
Neither of those moves is hypothetical. Both have happened in the past ten days at firms we are now in active scoping conversations with.
FAQ for U.S. law firms evaluating Mike OSS
Is Mike OSS really free?
The software is free under AGPL v3. The inference is not. Every prompt the system runs calls a paid Anthropic or Google API, billed directly to the firm’s account with the model provider. A small firm should expect inference costs in the low hundreds of dollars per month at light use, rising with document volume.
Can Mike OSS replace Harvey at an AmLaw 100 firm today?
No. At release, Mike is a prototype suitable for boutique and mid-sized deployments. Large firms have multi-tenant provisioning, SSO/SAML, enterprise audit logging, granular permissioning, and matter-management integration requirements that Mike OSS does not yet address. Those will likely come through community contributions or commercial forks over the next 12 to 18 months.
What about ABA Formal Opinion 512?
Formal Opinion 512 imposes the same baseline duties — Rules 1.1, 1.6, 5.3 — regardless of which AI tool a firm uses. Mike OSS does not exempt a firm from those duties; it changes which vendor relationships the firm has to diligence. Under a Mike OSS deployment, the diligence runs against the model provider (Anthropic or Google) and against the firm’s own infrastructure. Under a SaaS deployment, the diligence runs against the SaaS vendor and, transitively, against the model provider sitting behind them. Self-hosting collapses one layer of intermediary risk but does not remove the ethics obligation.
What does AGPL v3 mean if our firm modifies the code?
The firm can modify Mike OSS without restriction for internal use. The AGPL trigger is only pulled if the firm runs a modified version as a service exposed to people outside the firm — for example, a public client-facing portal built on a forked Mike codebase. In that scenario the firm must publish the modifications under AGPL v3. Pure internal use, including running the application for the firm’s own attorneys and staff, never triggers the publishing obligation.
Is the mikeoss.com hosted demo safe for confidential matter material?
No, and Chen says so on the site. The demo runs on infrastructure he controls, and prompts and uploads pass through his servers. Any firm wanting to evaluate Mike on real matter content should do so on a self-hosted install in the firm’s own cloud account, with a direct model- provider contract in place before the first prompt leaves the perimeter.
Does Mike OSS handle multi-jurisdiction U.S. work?
The underlying models (Claude and Gemini) handle multi- jurisdiction U.S. legal text well, including state-specific statutory and regulatory language. Mike itself is jurisdiction-agnostic. Where firms will hit edge cases is in the parts of legal work that depend on case-law citation accuracy — that is still a Westlaw and Lexis problem, and Mike does not pretend otherwise.
How does this compare to running Claude or Gemini directly?
Mike OSS is, in some sense, a structured way of running Claude or Gemini. The value it adds over raw API access is the matter file system (vault), the tabular review mode, the workflow presets, and the attorney-friendly UI. A firm that is comfortable with its attorneys using the model providers’ own web apps for narrow tasks may not need Mike OSS at all. A firm that wants attorney-friendly workflows over matter content, with audit logging and shared presets, gets meaningful value from the Mike layer.
The bottom line for U.S. legal buyers
Mike OSS is the most important piece of legal-tech news of 2026 so far, not because it will replace Harvey or Legora — it will not, in the near term — but because it has made the underlying argument legible. The legal-AI category has been priced on the premise that building a credible drafting and review platform is hard enough to justify $11 billion valuations and $300K-and-up enterprise contracts. A former Latham associate distributed evidence that the building part is not the hard part. The hard part is the firm-specific fit, the evals, the integrations, the case law, and the care.
Those are exactly the parts a generic platform cannot solve for any specific firm. They are also exactly the parts a custom build, scoped to one firm’s workflows, can solve in a few weeks of focused work. If Mike OSS pushes more U.S. firms to ask the right question — “what is the version of this that fits us?” instead of “which platform do we license?” — it will have earned its 2,200 stars and then some.
If you want to talk through what a Mike-OSS-style pilot might look like inside your firm, or what a custom build on top of the same underlying model providers would cost and how long it would take, we’re happy to run that conversation in 30 minutes. Bring the workflow you would most like to fix; we’ll work through the deployment, privilege, and budget questions in one call.
