Lavern

An agentic law firm. Yours.

Lavern v0.15.0 is the first open-source cut. Apache 2.0. 67 specialist AI agents that review documents through evidence-backed debate, with mandatory human gates and a 10-pass verification loop. Runs on a Mac Mini (local), Claude (frontier), or Mistral (EU sovereign).

curl -fsSL lavern.ai/install.sh | sh

Or:

git clone https://github.com/AnttiHero/lavern.git
cd lavern && npm install && (cd viz && npm install)
npm run serve:dev      # API on :3000  (demo mode, no API key needed)
cd viz && npm run dev  # Dashboard on :5173

What's actually inside

• 67 agents — 59 domain specialists + 7 orchestrators + 1 base prompt. Each is a system prompt with its own MCP tool permissions and a slot in the debate protocol.
• 21 MCP tool modules — debate board, scoring, verification, memory, risk pricing, baselines, knowledge base, report cards, quality checks, handoffs, feedback loop, document reader.
• 9 workflows — counsel, review, adversarial, roundtable, legal-design, full-bench, pre-engagement, verification, tabulate.
• 5 legal datasets — CUAD, MAUD, ACORD, UNFAIR-ToS, LEDGAR (FTS5-indexed for retrieval).
• 2 LLM providers — Anthropic Claude, Mistral (EU sovereign).
• Clawern — autonomous mode that watches a folder, reviews documents overnight, and pings your phone when something matters.
• 1,665 tests across 108 files. CI-green on this commit.

The protocol, briefly

What makes Lavern more than 67 prompts on Claude:

1. Debate is the product. Agents cite evidence from the parsed document for every finding. Unsupported findings don't enter the board.
2. Three verification layers. Evaluator gate → adversarial debate → 10-pass mechanical verification (clause grounding, defined-term integrity, monetary preservation, jurisdiction integrity, …). Each layer fails closed.
3. Human gates are mandatory. Critical findings pause for real human approval before delivery. Auto-approve is a CLI flag, not a default.
4. Precedent Board. Findings that recur get reinforced; stale ones decay. The next review starts smarter than the last.

The architecture deep-dive: lavern.ai/architecture/

Two modes

• Interactive. Brief the firm; it works. Briefing → Strategy → Team → Working → Delivery. Live dashboard, real-time agent activity, gate approvals via UI.
• Autonomous (Clawern). Point Lavern at a folder. It watches, reviews on a schedule, files deliveries next to the original docs, and notifies via Telegram / email / macOS Notification Center when something matters.

EU sovereignty

Set LAVERN_PROVIDER=mistral and MISTRAL_API_KEY=... and every model call routes through Mistral. Embeddings, agent dispatch, and the precedent board all stay on-disk in ~/.lavern/. No third-party telemetry. No data leaves the EU.

Local mode

Without an API key, the dashboard, demo cinematic, Clawern monitor, and agent profiles all work — fully wired UI, fully wired auth-skipping local-user mode. You see exactly what a real engagement looks like before spending a token.

What this isn't

• Not legal advice. Lavern produces document analysis and design transforms. The disclaimer is on every output. Verify with a qualified lawyer.
• Not a Claude plugin. claude-for-legal is plugins. Lavern is a firm — agents that debate, gates that pause, artifacts that ship.
• Not autonomous in the "no human" sense. The human gates are the architecture. Agentic — the agents work — but you sign off on the calls that matter.

Repo

• Discussions: enabled — community questions go there
• Issues: bugs
• License: Apache 2.0
• Status: v0.15.0 is the first publicly tagged release. Everything before this was development inside the OSS-prep branch.

Thanks

To everyone who reviewed builds, broke things, and pushed back on claims that didn't survive scrutiny. The launch is better for it.