GitHub says the repository has passed 4,300 stars. The license file says AGPL-3.0. Between those two facts sits most of the argument a founding AI engineer, Ritvik Puranik, made on LinkedIn this month about MikeOSS, an open-source legal AI that now does a striking amount of what Harvey charges for.

Puranik listed the feature set. Document review with verifiable citations, legal research, tabular review, MCP connectors, Word integration, matter-scoped context, local model support. Seven items. Most of Harvey’s own pitch.

Then he asked the question that follows from that list. If the code is public and anyone can run it, what exactly does a law firm pay Harvey for?

We already covered the other half of Harvey’s year, the in-house model it trained on China’s Kimi K3 to cut its per-query cost, in that analysis. This one isn’t about margin. It’s about who holds the risk once the software is free.

What you actually buy when you buy Harvey

Will Chen published MikeOSS on May 5, 2026, under the AGPL-3.0 license. Per Legal Cheek and Legal Futures, Chen trained and qualified as an associate at Latham & Watkins in Singapore and London, and read law at Oxford. He posted the repository saying he had rebuilt the core of what Harvey and Legora do in two weeks, and gave it away. It now lives at github.com/open-legal-products/mike.

Nobody in the thread argued the code doesn’t work. The answer that mattered came from Anil Kona, executive director and COO at Covasant, in the comments. Who does the work MikeOSS doesn’t automate? Negotiating data retention with the model provider. Reacting when a dependency gets compromised. Patching. Auditing. That work doesn’t vanish because the software is free. It moves from Harvey’s payroll to yours.

Harvey’s moat was never only the model. It’s how much of that list it takes off a customer’s desk.

What Harvey bills for, and what MikeOSS leaves with you

Harvey’s security page is specific about what it promises. It contractually requires zero data retention from its model providers, logically separates data between customers, offers residency in the EU and Switzerland, the US, or Australia on Azure, and holds SOC 2 Type II, ISO 27001, ISO 27701 and ISO 42001 certifications, backed by annual third-party penetration tests. Those commitments, and its incident-response terms, are written into the contract, not just the marketing page.

MikeOSS solves the other half of the trade. It runs on BYOK, bring your own key: connect your own Anthropic, OpenAI or Gemini account, or run a local model instead, and pay the inference provider directly with no middleman margin. It self-hosts on Docker. The site’s own line is blunt. “Legal AI you can truly own.”

Who owns this	MikeOSS	Harvey
Data retention with the model provider	You negotiate it	Harvey requires it by contract
Data residency	Wherever you host the server	EU, Switzerland, US or Australia
Certifications	None, it’s your own instance	SOC 2, ISO 27001, 27701, 42001
Patching and auditing dependencies	Your team	Harvey

AGPL-3.0 adds one more line to that table. Modify MikeOSS and offer it as a service to outside users, and the license requires you to publish those changes. For a firm running it in-house, that clause barely bites. For anyone planning to resell it, it does.

Firm size decides which risks you can carry

In the same comment thread, Kona wrote out the full list to price in before comparing. Hosting, contracting with the model provider, security reviews, dependency and version management, integrations, monitoring, audits, governance, support, incident response, compliance. Then he added the part that actually matters. That list weighs differently depending on the size of the firm.

A 20-lawyer firm almost never has someone whose full job is negotiating retention clauses with Anthropic or tracking a CVE in a Node dependency. Today it buys that inside Harvey’s subscription. A global firm with its own information-security team can absorb most of that list without hiring anyone new. And an enterprise legal department sitting inside a larger company usually already has the IT function that does exactly this work for everything else the company runs.

At that point, MikeOSS stops adding new risk. It just uses risk the company already pays to carry.

The same call outside legal

Swap Harvey for any vertical AI vendor, and MikeOSS for its open-source alternative, and the question doesn’t change. Open source with your own support, or a vendor with a contract. We mapped the general version of that trade, with Anthropic itself running on Salesforce instead of building its own CRM, in why AI companies buy SaaS instead of building it. And on why the risk a vendor retains, not the model it runs, is what you’re actually paying for, we argued that in the model is a commodity, governance is the moat.

What changes case by case is how much of Kona’s list your company can already absorb without a new hire. Inside the AI Maestro discovery, that’s one of the questions we answer before the Go/No-Go gate. Which operational risk already has an owner inside your operation, and which one you’d be buying for the first time the day you drop the vendor.

Don’t run that math against the license price. Run it against Kona’s list.

Let’s map which risk you can actually carry