Is Legora Just a Wrapper? What It's Actually Built On

Is Legora just a wrapper? The short answer: it started as one, it's spending heavily to stop being one, and the label now obscures more than it reveals. The longer answer is worth five minutes, because it explains how most of the legal AI market actually works.

First, the jargon, because it does nobody any favours. A “wrapper” is a product that doesn't build its own artificial intelligence. It takes an underlying model built by someone else, Claude from Anthropic or the GPT models from OpenAI, and wraps a product around it: an interface lawyers recognise, security controls, document handling, legal workflows.

Legora began exactly that way. It launched in Stockholm in 2023 as Leya, and the early product was a secure legal layer over OpenAI's models, run through Microsoft's Azure OpenAI Service: a way for lawyers to use a large language model without pasting client documents into a public chatbot. That was the pitch, and it was a wrapper in the plainest sense of the word.

So what is Legora's underlying model today? There isn't one; there are several. The company has described itself as model-agnostic since its Leya days, when reporting had it running OpenAI and Meta models on Azure, and Anthropic's own customer page now says Legora uses Claude across its assistant tools, document review and workflows, including the agentic work: the multi-step tasks where the system opens documents, follows a sequence and finishes a job rather than answering a single prompt. If your search was “is Legora built on Claude”, the accurate answer is: substantially, but not exclusively, and by design.

The product layer on top is where the last two years have gone: tabular review that reads hundreds of documents into a structured grid, drafting inside Microsoft Word, integrations with document management systems like iManage, and workflows tuned to how legal teams actually pass work around.

None of that settles the wrapper question, because the model makers can build all of it too. This is the precarious position I described in AI won't differentiate your law firm: AI is sold by the unit, the model makers own the units, and they can sell them cheaply enough to make a reseller's economics impossible. Harvey, Legora's most prominent rival, now runs Claude alongside models from OpenAI, Google and Mistral, has a customer case study on Anthropic's own site, and lets its workflows be called from inside Claude. A product that only rents and resells someone else's intelligence has no floor beneath it.

Legora has read that danger more clearly than most, and its answer has been to buy things the model makers can't simply replicate. Its acquisition of Qura in April 2026, a Stockholm company founded in 2023 to build an AI-native legal research search engine, gives it structured legal sources a model can reason over rather than a search index bolted on the side. It bought the Canadian agentic legal AI start-up Walter in March 2026 and the London fact-intelligence start-up Wexler in July, for similar reasons. Proprietary legal data, wired into the product, is the classic escape route from wrapper status. Whether Legora has taken it in time is the open question.

So, is Legora a wrapper? By origin, yes. By trajectory, it's becoming something harder to dislodge: a legal layer that sits across several models rather than on top of one, with research assets of its own underneath. If you're assessing it as an investment or a competitor, the wrapper label tells you where the supplier risk lives, and Qura is the hedge against exactly that risk.

If you're a law firm deciding whether to buy it, though, I'd steer you to a different question altogether. Whatever Legora is built on, every firm that licenses it gets the same thing, and a shared tool cannot be an edge. At best it's the cost of staying level. I've written a practical guide to how UK firms should evaluate legal AI that starts from that premise.

The wrapper debate is good sport for legal tech commentary. Inside a firm, three sharper questions do more work: what does the tool do that the raw model doesn't, what does it cost against using Claude or ChatGPT directly, and what will you build around it that your competitors won't? Legora will answer the first two in a demo. Only you can answer the third. On the second, I've since written up what Legora costs a UK firm and what you get for it.

Update, September 2026

Two months on, the trajectory I described above has firmed up. On 14 September Legora announced a full ontology of law and an AI-native citator, built on the Qura and Wexler acquisitions. Artificial Lawyer reports the ontology is in limited beta now, with general availability due in the fourth quarter of 2026. That is the proprietary-data escape route from wrapper status being taken in plain sight. Arvid Winterfeldt, Qura's co-founder and chief executive and now Legora's head of legal research, told Artificial Lawyer: “We have catalogued more than 50 distinct ways AI fails at legal research.” That's a claim worth testing in due diligence, and a useful one to have in writing.

The model question moved too. When OpenAI released GPT-6 Astra in early September, Legora described its own test in Artificial Lawyer within days: an agent completing a financial-statement tie-out across 41 documents in a single run and finding all four planted errors. Harvey's view of the same model appeared in the same piece. So the honest answer to “is Legora built on Claude” is now: on more than one frontier model, chosen by task, which is the point of being a layer rather than a wrapper. And on the UK side, Edwin Coe, a £44 million London firm, announced this month that it has gone firm-wide across all 14 practice groups after a five-week pilot, the clearest sign yet that the product has moved beyond the magic circle.