In short: open source legal software in October 2026 runs from data (CourtListener, 10 million+ opinions) to full applications. Mike OS, an AGPL-3.0 Harvey-style platform first published on GitHub on 29 April 2026, had 4,363 GitHub stars on 20 October 2026 and ships both a self-hosted build and a hosted app. The private work product of firms remains the gap.
Key facts
- Mike OS (open-legal-products/mike): AGPL-3.0, repository created 29 April 2026, 4,363 stars, latest release v0.4.0 (GitHub, read 20 Oct 2026). Hosted version at app.mikeoss.com, self-hosting via Docker Compose (mikeoss.com, read 20 Oct 2026).
- Free Law Project's CourtListener holds over 10 million court opinions, and its RECAP Archive holds tens of millions of PACER documents (free.law, read 20 Oct 2026). The CourtListener code is AGPL-3.0 with 1,047 stars (GitHub, 20 Oct 2026).
- Harvard's Caselaw Access Project covers over 6.7 million US cases across 360 years of published decisions (Harvard Library Innovation Lab, read 20 Oct 2026).
- LegalBench has 162 legal reasoning tasks from 40 contributors (HazyResearch/legalbench README, read 20 Oct 2026).
- CUAD, from the Atticus Project, labels 13,000+ clauses in 510 contracts across 41 categories, under CC BY 4.0 (Hugging Face dataset card, read 20 Oct 2026).
- Lawve's awesome-legal-skills registry (formerly Lawvable) lists 276 agent skills and has 878 stars (GitHub, read 20 Oct 2026).
- OCCRP ended maintenance of open source Aleph on 31 December 2025 and moved its effort to Aleph Pro (alephdata/aleph README, read 20 Oct 2026).
What is Mike OS, and where does it stand in October 2026?
Mike (also written MikeOSS or Mike OS) is an open source legal AI platform built by Will Chen, a former Latham & Watkins associate. It reached the top of Hacker News within hours of its first commit in late April 2026, and the pitch was simple: the core features of Harvey and Legora, self-hostable, with your own model API key and no per-seat licence.
Six months on, the project has moved from a weekend repository to a maintained product. The code now lives under the open-legal-products organisation on GitHub (the old willchen96/mike URL redirects), carries an AGPL-3.0 licence, and has shipped four tagged releases up to v0.4.0. The current README lists chat over documents and matters, document review with suggested edits, reusable assistant and tabular-review workflows, a document library, citation checks and US case-law research through CourtListener, a Microsoft Word task-pane add-in, and local models through Ollama. The stack is Next.js, Express, Supabase Auth and Postgres, with R2-compatible object storage.
There are two ways to run it. The hosted app at app.mikeoss.com is free to sign up for, and you pay your model provider directly. Self-hosting uses the bundled Docker Compose stack. Early coverage, including LawSites and Legal Technology Monitor, warned against putting client documents into the hosted demo; for confidential matters, the self-hosted route is the one to evaluate.
Mike OS is also one of the entrants in our 50-point task benchmark, where it placed ahead of several commercial tools. The live scores are in our benchmark write-up. Disclosure: HAQQ built and ran that benchmark, and HAQQ's own product is scored in it.
The AGPL choice matters for buyers. A firm can run Mike internally without publishing anything, but a firm that modifies it and offers it to clients over a network has to release its changes. Procurement teams that only approve MIT or Apache 2.0 code will want legal review before adopting it.
Which open source legal projects matter, and under what licence?
This table covers the projects we see most often in legal AI stacks. Stars and licences were read from GitHub (or Hugging Face for models and datasets) on 20 October 2026. "Not declared" means GitHub detects no licence at the repository level, so check individual files before reuse.
| Project | Licence | Stars (20 Oct 2026) | Use |
|---|---|---|---|
| Mike OS | AGPL-3.0 | 4,363 | Self-hosted legal AI: chat, review, tabular review, Word add-in |
| Aleph (OCCRP) | MIT | 2,449 | Investigative document search; maintenance ended Dec 2025 |
| OpenContracts | MIT | 1,518 | Document annotation and contract analytics |
| CourtListener | AGPL-3.0 | 1,047 | US case law search, RECAP archive, APIs |
| Docassemble | MIT | 992 | Guided interviews and document assembly |
| awesome-legal-skills (Lawve) | Custom (see repo) | 878 | Registry of 276 legal agent skills |
| LexNLP | AGPL-3.0 | 794 | Extraction of dates, amounts, citations, parties |
| Blackstone | Apache-2.0 | 700 | spaCy pipeline for UK and Commonwealth legal text |
| LegalBench | Not declared (per-task licences) | 648 | 162-task legal reasoning benchmark |
| Juriscraper | BSD-2-Clause | 647 | Scrapers for US court websites |
| eyecite | BSD-2-Clause | 291 | US citation extraction |
| LexGLUE | Not declared | 273 | Older legal NLP benchmark suite |
| ContraxSuite | AGPL-3.0 | 190 | Contract analytics platform; last push Feb 2023 |
| OLAW (Harvard LIL) | MIT | 166 | Open legal AI workbench for RAG research |
| Document Assembly Line (Suffolk LIT Lab) | MIT | 70 | Court-form automation on Docassemble |
| lawskills-hub (Harvard LIL) | Not declared | 45 | Agent skills for legal education |
| Open Decision | MIT | 30 | Decision-tree builder for legal self-help |
| CUAD (dataset) | CC BY 4.0 | n/a | 510 labelled contracts, 41 clause types |
| Pile of Law (dataset) | CC BY-NC-SA 4.0 | n/a | 256 GB legal text corpus for training |
| Legal-BERT (model) | CC BY-SA 4.0 | n/a | BERT pretrained on legislation, cases, contracts |
| Saul-7B-Instruct (model) | MIT | n/a | Open-weights LLM continued-pretrained on law |
Why did legal get open source later than other industries?
Healthcare has OpenMRS and finance has Hyperledger. Legal had very little for a long time, and the reasons were commercial more than technical.
The billable hour is the first. Automating a 10-hour task down to one hour removes nine billable hours, so few partners had a reason to fund shared tooling. Brief banks and templates are the second: firms treat them as trade secrets, and you cannot publish a litigation playbook you may use against the firm down the street. Licensing caution is the third. Compliance teams that see "GPL" often stop reading.
The fourth is data. Thomson Reuters and LexisNexis built their position on citators (KeyCite, Shepard's) and editorial taxonomies that took decades to assemble. Even with good open source software, the data layer underneath stayed closed. That is the part that has been opening, slowly, through public court archives and government publishing.
What does the open legal data layer include?
Free Law Project runs CourtListener, RECAP (a browser extension that moves federal filings out of PACER into a free archive), eyecite (the standard US citation parser) and Juriscraper (scrapers for hundreds of US courts). Many legal AI products use this data, and few credit it.
Harvard's Caselaw Access Project digitised the official US reporters: over 6.7 million published cases from 1658 onwards. It is the starting point for anything that needs American precedent at scale.
Pile of Law is a 256 GB corpus of legal and administrative text on Hugging Face. Its CC BY-NC-SA licence rules out commercial training without separate permission, which many users overlook.
Find Case Law from the UK National Archives publishes judgments as LegalDocML XML with Atom feeds, the cleanest public schema we know of. EUR-Lex exposes EU legislation and CJEU case law with a SPARQL endpoint. OpenLegalData normalises German court decisions from fragmented official portals. InLegalBERT from IIT Kharagpur covers Indian Supreme Court and High Court judgments.
Brazil shows the usual pattern outside these jurisdictions: the CNJ DataJud API is public, and the community wrappers that make it usable are maintained by a few volunteers. For the Middle East and North Africa, my mena-legal-atlas maps the gazettes, consolidated legislation portals, apex courts and regulators of 25 jurisdictions; most still publish statutes as scanned PDFs.
- Free Law Project
- Caselaw Access Project (Harvard)
- Pile of Law (Hugging Face)
- Find Case Law (UK)
- EUR-Lex
- OpenLegalData
- InLegalBERT
- MENA scope of the HAQQ benchmark
Which open NLP libraries, models and benchmarks do legal teams use?
Blackstone is a spaCy pipeline for UK and Commonwealth legal text, one of the few non-US legal entity recognisers. LexNLP extracts entities, citations, durations and amounts; it predates LLMs and still sits inside a number of commercial tools. Note its AGPL-3.0 licence.
Legal-BERT from AUEB was pretrained on EU legislation, ECHR cases and US contracts and has over a million downloads. Saul-7B-Instruct from Equall showed that continued pretraining of an open model on legal text works. CUAD from the Atticus Project is the dataset most contract-AI products train or evaluate on.
For measurement, LegalBench (Stanford and Hazy Research) is the benchmark frontier labs now report. LexGLUE is the older suite (ECtHR, SCOTUS, EUR-Lex, LEDGAR, CaseHOLD, UNFAIR-ToS). Legal Benchmarks AI is practitioner-facing and vendor-independent, with published methods for its contract drafting study.
- Blackstone (ICLR&D)
- LexNLP
- Legal-BERT
- Saul-7B-Instruct
- CUAD (Atticus)
- LegalBench (Stanford)
- LexGLUE
- Legal Benchmarks AI
What are legal agent skills, and where do you find them?
A skill is a folder with a SKILL.md file and supporting material that teaches an AI agent a procedure: an NDA triage, a GDPR breach assessment, an EU AI Act classification. Claude, Codex, Gemini CLI and other agents read the same format, so one skill written by a practitioner can run in several tools.
awesome-legal-skills is the largest registry we know of: 276 skills from practising lawyers, in-house counsel and academics, synced from lawve.ai (the project was called Lawvable when we first wrote about it). Disclosure: I have contributed nine skills to it. lawskills-hub from Harvard's Library Innovation Lab collects skills for legal education.
anthropics/skills is where the SKILL.md format started. Anthropic's knowledge-work-plugins repository (Apache-2.0) includes the legal plugin released with Claude Cowork in February 2026, which covers contract review, NDA triage and compliance workflows; Reuters reported Thomson Reuters shares fell nearly 18% in the days after.
Beyond skills, ContraxSuite remains the most complete open contract analytics pipeline, though it has not been updated since February 2023. OLAW is Harvard LIL's workbench for retrieval experiments against legal APIs such as CourtListener. LangChain and LlamaIndex are general frameworks, and most legal RAG demos are built on them. Academic legal LLMs such as ChatLaw and DISC-LawLLM (Fudan University) are worth a look for Chinese-language work.
- Mike (mikeoss.com)
- Mike on GitHub
- awesome-legal-skills (Lawve)
- lawskills-hub (Harvard LIL)
- anthropics/skills
- anthropics/knowledge-work-plugins
- ContraxSuite
- OLAW (Harvard LIL)
Does HAQQ publish open source code?
HAQQ's product is commercial and closed source, and the HAQQ-Inc organisation on GitHub has no public repositories as of 20 October 2026. The open source legal projects connected to us are ones I publish, as HAQQ's head of growth, on my personal GitHub account. The ones most relevant to legal teams, with stars read on 20 October 2026:
| Repository | Licence | Stars (20 Oct 2026) | What it does |
|---|---|---|---|
| master-claude-for-legal | MIT (per README) | 72 | Skill pack and templates built from Anthropic's Claude for Legal Teams webinar |
| nomos | Apache-2.0 (per README) | 23 | Experimental language for legal rules with jurisdiction and validity dates |
| mini-claude-for-legal | MIT | 14 | 982 legal skills in SKILL.md format, MENA-first |
| legal-ai-model-router | AGPL-3.0 | 6 | Routes legal tasks to models using public benchmark results |
| mena-legal-atlas | Not declared | 4 | Map of legal data sources in 25 MENA jurisdictions |
| hallucination-shield | MIT | 3 | Citation checker for legal briefs |
| cite-guard | Not declared | 1 | Citation hallucination risk score for AI-drafted filings |
Two corrections to earlier versions of this post. The LegalMD repository I linked in May is no longer public. And the repository github.com/sboghossian/haqq-ai is an old copy of our website code from before it moved to the HAQQ-Inc organisation; it is not maintained and is not HAQQ's product.
Our working rule is that parsers, schemas, scrapers, evaluation harnesses and skill registries are not where a legal AI company wins, so they can be public. Jurisdiction depth, firm-specific configuration and the product a lawyer uses every day stay commercial.
- master-claude-for-legal
- nomos
- mini-claude-for-legal
- legal-ai-model-router
- mena-legal-atlas
- hallucination-shield
Vyzkoušejte HAQQ AI zdarma
Zažijte sepisování a rešerše s podporou AI
Which open source tools handle document automation and access to justice?
Docassemble (MIT) powers court self-help interviews in several US states; if you have filled in a free court form online, it may have been underneath. Suffolk LIT Lab's Document Assembly Line builds on it to automate court forms across states, and was updated as recently as 19 October 2026. Open Decision (MIT, Berlin) is a web decision-tree builder for legal self-help, though its repository has been quiet since February 2024.
OpenContracts (MIT) handles annotation and contract analytics, and the wider opensource.legal group also publishes Docxodus (DOCX redlining in WebAssembly) and Python-Redlines. Aleph did OCR, entity extraction and cross-document search for investigators, but OCCRP stopped maintaining the open version at the end of 2025, so treat it as a frozen codebase. The SALI Alliance LMSS taxonomy is an open standard for matter and service classification that many legal data projects adopt.
If you are shipping code in this space, read GitHub's open source guide to the legal side first, so the differences between MIT, Apache 2.0 and AGPL are settled before your first release.
- Docassemble
- Suffolk LIT Document Assembly Line
- Open Decision
- OpenContracts
- opensource.legal
- Aleph (OCCRP)
- SALI Alliance
- GitHub Open Source Guide: Legal
What can open source legal AI not do?
The limit is data. The work clients pay for (memos, deal documents, discovery productions, settlement letters, a senior partner's redlines) is privileged, confidential or contractually locked inside firms. It cannot be released, benchmarked publicly or used to train a model another firm will use.
What is public (published opinions, statutes, EDGAR filings, CUAD's 510 contracts) is a narrow slice: appellate, public-company, mostly English and mostly American. That explains why LegalBench feels narrow and why contract models overfit CUAD. Open source can ship excellent plumbing and good public-data models. It cannot close the loop on the work product that defines practice.
The open frontier we find most promising is federated evaluation, synthetic data, privacy-preserving fine-tuning and skill registries, where firms keep data private and share behaviour. The corollary is that public legal data matters more each year. Every government that publishes statutes and judgments in machine-readable form widens the ground where open tools compete on equal terms. The UK did it with LegalDocML and the EU with EUR-Lex. Most of the jurisdictions HAQQ serves have not.
Is open source legal AI safe for privileged work?
The software licence does not decide privilege; where the data goes does. In United States v. Heppner (S.D.N.Y., No. 25-cr-00503, February 2026), Judge Jed Rakoff held that documents a defendant generated with the consumer version of Claude were protected by neither attorney-client privilege nor work product, partly because the provider's terms allowed it to collect inputs and outputs (Orrick summary). We cover the ruling in AI conversations are not privileged.
A self-hosted deployment on firm infrastructure, with firm-controlled keys and a model provider under zero-retention terms, puts a firm in a different position from a consumer chatbot, although commentators disagree on how far Heppner's reasoning reaches. Open source makes self-hosting affordable without an enterprise contract. It also moves the security work (patching, access control, logging) onto the firm.
Common objections
It's just a wrapper around an LLM.
So are most AI products in any vertical, commercial ones included. The product lives in the wrapper: citation checking, document handling, workflow, deployment and security. Judge the wrapper on those.
Copyleft will kill enterprise adoption.
It slows it. Mike's AGPL licence will send it to legal review at many firms, and permissive projects (Docassemble, OpenContracts, eyecite) clear procurement faster. Several of my own repositories are AGPL too, so I live with this trade-off.
Big firms will never use open source for legal work.
They already run Linux, Postgres and Chromium. The open question is the legal-specific layer, and firms adopt it when they can self-host, audit and lock it down.
What should you watch next?
- Licensing: whether Mike stays AGPL or adds a commercial licence for firms that want to modify it privately and offer it to clients.
- Self-hosting: the Heppner ruling pushes firms toward private deployments, and projects with clean Docker images and documented security setups will benefit.
- Public legal data outside the US, UK and EU: court data in Brazil, India, Indonesia, Nigeria and the Gulf is where the next decade of open legal AI will be decided.
- Skills as the unit of legal knowledge: registries grew from about 40 skills to 276 in Lawve's case between May and October 2026.
- Weekend projects: more will go viral. Watch which ones practising lawyers, not only engineers, keep using after a month.
If you want to test an open tool against a commercial one, run both on the same ten matters from your own files, score the outputs blind, and check where the data goes in each. If you are building something in this space, open source or not, write to me at stephane@haqq.ai.
