lavern  by AnttiHero

Multi-agent AI system for evidence-based legal document review

Created 1 month ago
270 stars

Top 95.1% on SourcePulse

GitHubView on GitHub
Project Summary

Summary

Lavern is an agentic legal system automating document review via multi-agent debate and verification. It targets technical users and legal professionals, offering a structured, evidence-backed, auditable AI process beyond the "junior associate" model. The system aims for superior output quality, though this remains an empirical question.

How It Works

The system uses 67 specialized AI agents and orchestrators in an evidence-backed debate protocol. Agents cite specific text for findings, which undergo three verification layers: an evaluator gate, adversarial debate, and a 10-pass quality check. Human gates precede critical decisions. It supports Anthropic, Mistral AI (EU), or local Ollama, using a persistent "Precedent Board" for memory.

Quick Start & Requirements

Installation uses platform scripts (curl, PowerShell) or manual git clone + npm install. Local mode requires no API keys, running the API on port 3000 and dashboard on 5173. Real engagements need Anthropic/Mistral API keys. A 60-second walkthrough is in QUICKSTART.md.

Highlighted Details

  • Features 67 agents, 21 MCP tools, and 9 workflows.
  • Includes 5 seeded legal datasets (CUAD, MAUD, ACORD, UNFAIR-ToS, LEDGAR).
  • Supports Anthropic, Mistral AI, and local Ollama inference providers.
  • Boasts 1,677 tests and clean TypeScript checks.
  • Offers Interactive, autonomous (Clawern), and EU modes.

Maintenance & Community

Developed over six months by a law firm founder, copyrighted 2025–2026 by Antti Innanen. Contribution guidelines exist, but explicit community channels or sponsorship details are absent from the README.

Licensing & Compatibility

Licensed under Apache 2.0, generally permissive for commercial use. Bundled datasets have separate licenses; one (ContractNLI) was removed due to incompatibility. Consult NOTICE for dataset specifics.

Limitations & Caveats

No public benchmark exists; quality claims are hypotheses. Multi-agent debate is imperfect. Lacks dense/vector retrieval (uses BM25). No durable task queue; server restarts mid-engagement require re-initiation. Counsel deliveries can be slow. EU mode has a specific limitation where the "Lavern Challenge" may access Anthropic.

Health Check
Last Commit

5 days ago

Responsiveness

Inactive

Pull Requests (30d)
1
Issues (30d)
0
Star History
31 stars in the last 30 days

Explore Similar Projects

Feedback? Help us improve.