CaseCite is a self-hosted open-source take on Harvey, with 69 HN points: Reveneau

Why it mattersLegal AI buyers now have an open option. A firm that will not send client documents to a vendor cloud can host the same workflows itself.

A single developer has published a self-hosted take on the kind of product Harvey built. The repository is at thomasunise/casecite on GitHub. Harvey raised $500 million for legal AI earlier this year, as the author's own Hacker News headline put it. The post reached 69 points on 28 October. The repository had 74 stars on 29 October, up from zero on 27 August when the first commit landed.

What CaseCite does

CaseCite's README describes nine tools behind one login, all reading from one document corpus. Research is a chat over a firm's documents, with semantic retrieval, BM25 rescoring and inline citations back to the source. Documents imports from Google Drive, OneDrive, Box, Dropbox, Filevine or Clio. Contracts analyses one contract or compares two to four against a baseline, with redline, draft and analysis exports.

Tabular review handles hundreds of contracts at once. The README says every value is tied to the exact passage that supports it. Workflows saves a review or drafting process as an ordered list of steps. It runs the list over any document set as one background job. Drafting works against up to four reference documents. The user approves a section plan before any section is written.

Case view opens a CourtListener opinion and chats with it. Judge intel reads CourtListener for per-judge analytics. Legal tools covers case lookup and citation check. The README calls citation check a scan of citing opinions for negative-treatment language. It states that citation check is a research aid and does not replace a Shepard's or KeyCite check.

What stays on the firm's servers

The README says documents, the vector index, chat history and analyses are stored only on the deployment. Using the AI features sends content to the providers the firm picks. Uploading sends document text to the embeddings provider, OpenAI by default, or Voyage or Cohere if configured. Asking a question sends the question and the relevant document text to the chat provider. That chat provider can be OpenAI, Anthropic or Gemini. Case-law tools send search queries and citations to CourtListener.

To keep document text off every outside server, the firm hosts both the chat and embeddings models itself. The README says to point both base URLs at compatible servers it runs. Setting only the chat server still sends document text to OpenAI for embeddings.

Licence and custody

CaseCite is Apache-2.0, with a contributor patent grant the README notes is not in MIT. The README says versions up to 1.3.0 were released under the Elastic License 2.0, which still applies to those versions. Commercial support and private installs come from the maintainer, Thomas Unise. Per-user API keys are stored server-side, AES-256-GCM-encrypted on the data volume.

The README includes an AI-review notice on every generated answer, analysis and draft. It says CaseCite is a research aid, not a substitute for a lawyer's judgment or for a citator.

A firm picking a legal AI product in 2027 now has one more option. It is open under Apache-2.0 and the firm can audit and run it on its own servers. Harvey's $500 million round funds a different model: a hosted product in Harvey's cloud. Some firms have clients whose documents are never allowed off the firm's network. Those firms now have an open answer to point at.

Source

thomasunise/casecite on GitHub, the repository README
Harvey raised 500M for legal AI, I built the same and open sourced it, Hacker News discussion

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.