The product pitch: an open-source legal-agent platform forked from OpenCode where documents never leave infrastructure you control. Promises plain-English answers with citations to exact clauses, real tracked changes baked into .docx files, multi-agent contract review, a tabular review grid, and a choice of 75+ model providers down to fully local Ollama — with embeddings always computed locally. Points at start.sh --demo as the two-minute on-ramp and closes with a "not legal advice" disclaimer grounded in the citation discipline. First contact with the project, or when you need the product's own framing of its privacy posture.
doc.haus
Not a chatbot. A legal team that works on your machine.
Open-source multi-agent legal AI — your documents stay on your machine; the redlines land in Word.
doc.haus — the website, with video demos of everything below.
- Documents never leave your infrastructure. Each matter gets its own local index on your own disk; there is no doc.haus cloud, and the LLM models the agents run on can be local too.
- Word-native. Real tracked changes baked into the
.docxitself — open the result in Word and accept or reject each change, exactly as if a colleague had marked it up. - Every answer cites the clause it came from, quoted verbatim, so you can verify it before you rely on it.
Get started
You need Bun (curl -fsSL https://bun.sh/install | bash). Then:
git clone https://github.com/sure-scale/doc-haus.git && cd doc-haus
./start.sh --demo
The browser opens with a demo matter ready: a fictional letter of engagement, already ingested. On first run the app asks you to connect a model provider in Settings — an API key (Anthropic, OpenAI, …), a Google Cloud sign-in, or a fully local endpoint (Ollama, vLLM, LM Studio). See docs/providers.md for copy-paste configs.
Try asking: What is the cap on the firm's liability? — you get the answer cited to the exact clause, quoted from the document. More things to try in demo/README.md.
To work on your own documents: create a matter, upload a .docx, and ask. Everything —
conversations, documents, the index — persists on your machine.
What you can do
- Every engagement gets its own private workspace. Matters are separate, self-contained workspaces — each holds its documents, conversations, and a private search index, and nothing is shared between them.
- Ask in plain English; get citations or redlines. One conversation handles both:
questions come back with the clause cited and quoted, and edit requests land as
tracked changes in the
.docxitself. - Pick an assistant — or let Auto choose. Switch between focused assistants — Q&A, Redline, Research — or leave it on Auto and doc.haus routes each message to the right agent as the conversation moves between asking, editing, and researching.
- Run a full multi-agent review in one click. Pick a workflow like Full review and a reviewer, an adversarial challenger, and a summarizer each read every document, then hand back one combined report — the multi-agent engine doc.haus inherits from OpenCode, pointed at contracts.
- Review every contract in the matter as a grid. Define question-columns once — Liability cap, Payment terms, Governing law — and doc.haus answers them for every document in the matter, side by side.
Video demos of each on doc.haus.
Why firms can trust it
doc.haus is self-hosted on infrastructure you control — no doc.haus cloud, no multi-tenant service, no vendor holding your clients' privileged documents. Document text and embeddings never leave your disk; the only thing that touches the network is the prompts the agents send to the model providers you chose, and those can be local models so nothing need leave at all. Everything binds to localhost by default; to put it in front of a team, front it with the reverse proxy and SSO your firm already trusts. MIT licensed — embed it, modify it, ship it. Full posture and vulnerability reporting in SECURITY.md.
Built on OpenCode
doc.haus is a true fork of OpenCode that retargets its agent harness from code onto legal documents — so it inherits a real, battle-tested agent engine (multi-agent review, permissions, skills, 75+ model providers) and keeps pulling upstream improvements via merge. The concepts map 1:1:
| Legal concept | OpenCode primitive |
|---|---|
| Matter | project (a directory) |
| Document | a file in the matter dir |
| Conversation | session |
| Legal agent | agent |
| Multi-agent review | primary agent + Task subagents |
| Retrieval / citation | a custom tool |
Built by Nick Watson from SureScale.ai on OpenCode by the Anomaly team, MIT licensed. The heavy lifting — the agent engine itself — is theirs; doc.haus is the legal layer on top. doc.haus is not affiliated with or endorsed by the OpenCode team.
Open to working with firms on implementations, customizations, and integrations — reach out on LinkedIn.
Disclaimer
doc.haus is software, not a law firm. Its output is not legal advice and, like all AI output, it can be wrong. Every answer cites its source so it can be checked — check it, and review anything doc.haus produces with a licensed attorney before relying on it.
Learn more
- doc.haus — the website, with video demos.
- docs/architecture.md — how it works: the three processes, the agents and tools, models and providers, running pieces by hand, mergeability.
- docs/providers.md — connect Anthropic, OpenAI, Google Vertex, or a fully local model.
- demo/README.md — the demo matter and what to try.
- docs/atlas/README.md — The Atlas: a live doc↔code navigator with search, drift detection, journeys and anchored feedback.
- CONTRIBUTING.md · SECURITY.md · FUTURE.md