BigLaw is an experimental research project, not production-hardened software. It handles credentials, client matter data, and privileged legal communications, and it has not undergone a formal independent security audit. Before anything touches real client data, read the security notice and the legal notices & disclaimers — they are not boilerplate — and engage an independent security review. AUTH_ENABLED=false is the local-dev default; never expose the API on a shared network without enabling authentication. Nothing this software produces is legal advice.

BigLaw is a cross between a platform, an experiment, and an art project.

As a platform, it is the most comprehensive open legal AI stack that exists — spanning research, drafting, redlining, e-signatures, briefing, docketing, billing, and collaboration across a bench of 100+ agents in a structured multi-round debate architecture.

As an experiment, it is an ongoing attempt to answer a genuine engineering question: how much of the $50,000–150,000 per-lawyer-per-year legal tech stack can be replicated with open models, open protocols, and open code? The answer so far is: most of it.

As an art project, it is a provocation. The cost math is not a sales pitch. It is a statement about who gets access to tools and who doesn't, and what happens when that changes. It is deliberately maximalist, deliberately opinionated, and deliberately not finished.

You are not buying a product. You are picking up a thing that is still being built and deciding what to do with it.

BigLaw isn't a chatbot with a legal prompt. It's an orchestration engine that replaces a stack of vendor contracts with a single open-source platform. It runs DyTopo rounds of granular epistemic, conceptual, and writing agents over an in-process vector agent registry — and puts a debate + verification protocol between every finding and the page. Low-confidence or challenged findings stop at a human gate before they reach final synthesis.

Big Michael is the agent that lives inside your firm's collaboration channels. @-mention him in Teams or Slack and he dispatches tasks to BigLaw's bench, surfaces matter status and client briefings, and posts back when work is done — turning the platform into a conversational layer on top of everything else the firm already uses.

The tab in your browser you never click is a $300,000 invoice.

Westlaw + CoCounsel, Practical Law, Contract Express, LexisNexis + PSL, Kira, iManage, Everlaw, Ironclad, Clio Insights, Solve Intelligence: $50,000–152,000 per lawyer, per year. BigLaw covers the same ground for the price of your model-provider API bill — at typical usage for a ten-lawyer firm, roughly $2,400/year against $500,000+.

The line-by-line cost chart, the math by firm size, and the feature-by-feature comparison: Why BigLaw.

Every tool in that chart is a subscription you can cancel the day you run setup.sh. Not all at once. One at a time. Start with whatever costs the most. Run the matter through BigLaw. Compare the output. Keep what you cancel.

Go. Do likewise.

A real matter, mid-flight — the bench self-organising, then the cited result.

Round-by-round communication graph	Cited, verified synthesis

Live admin · DyTopo depth, modes, DocuSeal	Convene a matter — client/matter numbering

Screenshots are captured from the running web console against a live matter (client 10482 · matter 10482-014). The interface, DyTopo communication graph, human gates, and per-round agent routing are exactly as the system produced them.

curl -fsSL https://raw.githubusercontent.com/discover-legal/BigLaw/main/setup.sh | bash

Needs git + Docker. Clones if needed, seeds .env, brings up the three-container stack, and waits for the REST API at http://localhost:3102. Add your QWEN_API_KEY (or local-inference settings) to .env — unconfigured connectors degrade gracefully.

bash setup.sh                                        # already cloned? same thing
go run ./biglaw-go/cmd/biglaw                        # native (Go 1.25+), REST on :3101
go run ./biglaw-go/cmd/biglaw demo                   # end-to-end tour: seed → tabular review →
                                                     # CP checklist → counter-redline (~$0.03)
cd ui && npm install && \
  BIG_MICHAEL_API=http://localhost:3102 npm run dev  # web workbench on :5173

Full install paths, the workbench tour, and your first task: Getting started. Opening this repo in Claude Code registers BigLaw as an MCP server: MCP guide.

• 100+ agents, 4 tiers, DyTopo rounds — Need/Offer matching builds each round's comm graph; every finding survives CitationGate → adversarial debate → verification → human gate → Architecture
• Verbatim citation grounding by construction — hybrid RAG + staged extraction + multi-pass writer took a local 7B from ~0% to 94% grounded citations → Grounding & coverage
• The negotiation stack — counter-redline loop with per-change rationale cards, Redtime per-clause negotiation timelines with silent-edit detection, Integrity Check on every inbound document → Negotiation
• Tabular review — 50 docs × 30 columns with a citation-verification ladder (exact → tolerant → paraphrase judge → ensemble) and a verified tally on the export → Tabular review
• Document production — .docx generation, tracked-changes redlining, PDF/OCR, DocuSeal e-signing → Document production
• Four-tier playbook cascade (client > matter > personal > firm) driving playbook-aware redlines → Playbooks
• Big Michael in Teams & Slack — matter status, hub-and-spoke client briefings, task dispatch from an @-mention → Big Michael
• Court deadline calculator — FRCP / UK CPR / EU Competition, calendar vs business days, cited → Deadlines
• Research engines — CourtListener-backed citation checking, headnote extraction, precedent generation → Research engines
• Billing & LPM — automatic 6-minute billable units, LEDES 1998B, pre-bills, OCG compliance, daily status reports, docket/regulatory/budget monitors → Billing & LPM
• 32 connectors across 15 providers — CourtListener, Westlaw, Everlaw, Ironclad, iManage, Clio (OAuth), Drive, Box, Slack, and more → Connectors
• Court-ready audit trail — hash-chained JSONL with every tool call attributed to the responsible lawyer → Audit trail
• Per-call cost tracking with cache-aware pricing and local power estimates → Cost tracking
• Lawyer voice fingerprinting — drafting agents mirror the assigned lawyer's style from their own writing → Tone profiles
• Cloud or fully local — Qwen by default, any OpenAI-compatible endpoint, or air-gapped Ollama / LM Studio / vLLM → Local inference

On Harvey LAB (60-criterion all-pass rubric, rubric hidden from the agents; judge claude-sonnet-4-6): claude-haiku-4-5 raw, in Harvey's own harness, scores 41/60; the same model on the BigLaw pipeline scores 50/60, and a cross-vendor GLM-5.2 run tops the board at 52/60 — all clean, zero starved rounds. A free, on-prem qwen2.5:14b scores 39/60 (a new local record) on the resilience-wave build. Spot-checked citations came back verbatim on the pipeline runs where the raw run fabricated statutory penalty figures. The full ladder, including the confounded runs that were found and fixed along the way, and the technique-by-technique account: Benchmarks.

Everything lives in docs/, one page per topic:

Apache-2.0 (LICENSE) with an express patent grant — use it, modify it, embed it, run it as a service; attribution per NOTICE is all that is asked. The document tools are a clean-room reimplementation with a published spec and an executed attestation record (docs/clean-room-attestations.md) — the full story: Provenance & licensing.