--- slug: "bmw-lex" source_type: "readme" source_url: "https://cdn.jsdelivr.net/gh/itsamirkhon/lex@main/README.md" repo: "https://github.com/itsamirkhon/lex" source_file: "README.md" branch: "main" --- # Lex — BMW Legal AI Agent Platform > **Munich Hacking Legal Challenge** — an AI agent platform for the BMW Group legal department. > Lex coordinates specialized AI agents (contract review, compliance, research, risk) and returns > structured, traceable, source-cited deliverables to human lawyers. Built on the [Pi](https://github.com/mariozechner/pi) coding-agent runtime and adapted for BMW legal workflows. --- ## Quick Start ### Install ```sh # One-line install (clones into ~/.local/share/lex) curl -fsSL https://raw.githubusercontent.com/itsamirkhon/lex/main/install.sh | bash # OR via npm npm install -g bmw-lex ``` ### Configure ```sh lex auth # paste your OpenRouter API key (sk-or-v1-...) ``` Get a key at [openrouter.ai/keys](https://openrouter.ai/keys). Lex uses Claude Sonnet 4.5 by default. ### Run ```sh lex # interactive REPL lex contract-review samples/acme-supplier-nda-draft.md --jurisdiction de lex compliance-check "Acme GmbH" --check-type sanctions,lksg lex legal-research "force majeure pandemic" --jurisdiction de,uk lex-web # web UI on http://localhost:3000 ``` > **Name conflict on Linux?** A system tool `/usr/bin/lex` (a lexer generator) may shadow ours. > Fix with `export PATH="$HOME/.local/bin:$PATH"` or use the `bmw-lex` alias. --- ## What Lex Does A corporate legal department handles thousands of recurring tasks: reviewing supplier contracts, checking entities against sanctions lists, researching jurisdictional differences, drafting from templates. Lex automates the **decomposable parts** while keeping a human lawyer in the loop for judgment calls. Every workflow: 1. **Classifies** the request (contract type, jurisdiction, urgency) 2. **Plans** the approach in `outputs/.plans/.md` 3. **Delegates** subtasks to specialist agents in parallel 4. **Synthesizes** findings into a single deliverable 5. **QA-checks** all citations and jurisdiction coverage 6. **Gates** on human review before final delivery 7. **Returns** the artifact + a `.provenance.md` sidecar with full audit trail --- ## Architecture ``` ┌─────────────────────────────────────────────────────────────┐ │ CLI / Web UI │ │ bin/lex.js • bin/lex-web.js (Next.js on :3000) │ └────────────────────────────┬────────────────────────────────┘ │ ┌──────────────▼───────────────┐ │ Lex Orchestrator │ │ .lex/SYSTEM.md │ │ Pi runtime + extensions │ └──────────────┬───────────────┘ │ ┌─────────────────────┼─────────────────────┐ │ │ │ ┌───▼────┐ ┌──────▼──────┐ ┌──────▼─────┐ │Workflow│ │ Subagents │ │ Tools │ │Prompts │ │ .lex/agents│ │ extensions/│ └────────┘ └─────────────┘ └────────────┘ File-based handoff via outputs/.plans, outputs/.drafts ``` ### Specialized agents | Agent | Role | Output | |-------|------|--------| | `contract-agent` | Clause-by-clause review, redlines, BMW-template comparison | `contract-analysis.md` | | `compliance-agent` | OFAC/EU sanctions, LkSG, ESG, GDPR checks | `compliance-report.md` | | `research-agent` | Statute & case law (DE / US / UK / FR / EU) | `legal-research.md` | | `risk-agent` | Risk register with CRITICAL/HIGH/MEDIUM/LOW grading | `risk-assessment.md` | | `negotiation-agent`| BATNA, position tiers (preferred → walk-away), redlines | `negotiation-brief.md` | | `knowledge-agent` | Internal precedent / template / memo search | `kb-results.md` | | `qa-agent` | Citation verification, jurisdiction completeness | `qa-report.md` | ### Workflow commands | Command | Description | |---------|-------------| | `/contract-review [--jurisdiction ]` | Risk-grade a contract clause-by-clause | | `/compliance-check [--check-type sanctions\|lksg\|esg\|gdpr\|all]` | Multi-framework compliance assessment | | `/legal-research [--jurisdiction ]` | Cross-jurisdiction statutory + case law research | | `/risk-assessment ` | Comprehensive legal risk register | | `/contract-draft [--jurisdiction ]` | Draft from BMW standard templates | ### Custom legal tools Registered as Pi tools, available to all agents: - `legal_web_search` — pre-filtered for `gesetze-im-internet.de`, `eur-lex.europa.eu`, `bailii.org`, etc. - `sanctions_check` — searches OFAC SDN, EU Consolidated, UN Security Council lists - `eurlex_search` — EU regulations, directives, decisions - `document_parse` — PDF/DOCX → plaintext for contract analysis - `document_upload` — copies files into a matter workspace --- ## Project Layout ``` . ├── bin/ # CLI entry points (lex, lex-web) ├── dist/ # Compiled TypeScript (cli, paths, runtime, setup) ├── extensions/ │ ├── legal-tools.ts # Pi extension entry │ └── legal-tools/ # Sub-modules (search, document, matter, header) ├── prompts/ # Workflow templates (slash commands) │ ├── contract-review.md │ ├── compliance-check.md │ ├── legal-research.md │ ├── risk-assessment.md │ └── contract-draft.md ├── .lex/ │ ├── SYSTEM.md # Orchestrator system prompt │ ├── settings.json # Pi packages config │ └── agents/ # 7 specialized subagent definitions ├── knowledge-base/ # BMW internal legal corpus (mock) │ ├── policies/ # Primary BMW policy sources (Code of Conduct) │ ├── templates/ # Standard NDA, supply, service agreements │ ├── precedents/ # Approved clause language │ └── memos/ # LkSG checklist, sanctions procedure ├── samples/ # Demo contracts for testing ├── web/ # Next.js 15 web UI │ └── app/ │ ├── page.tsx # Main UI (slash-command input + artifact viewer) │ └── api/ # task / status / upload / artifacts routes ├── scripts/ │ └── patch-embedded-pi.mjs # Runtime patches to embedded Pi (tool name filter, etc.) ├── outputs/ # Generated at runtime │ ├── .plans/.md # Matter plans (orchestrator scratch) │ ├── .drafts/-*.md # Subagent intermediate artifacts │ └── matters// # Final deliverables + provenance ├── install.sh # One-line installer └── package.json ``` --- ## Provenance & Auditability Every final deliverable is paired with a `.provenance.md` sidecar: ```markdown - Matter ID: bmw-acme-nda-review - Date: 2026-04-25 - Jurisdiction(s): DE - Governing Law: BGB / GeschGehG - Task Type: contract-review - Agents used: contract-agent, risk-agent, knowledge-agent, qa-agent - Sources consulted: 14 | Sources verified: 12 - QA result: PASS WITH NOTES - QA MAJOR findings: Liability cap §8.2 references unverified market norm - Human review gate: APPROVED - Privilege status: PRIVILEGED ``` This enables full audit trails for regulatory inquiries, BAFA reporting, and internal governance. --- ## Demo Scenarios ### 1. Contract Review (CLI, ~2 min) ```sh lex contract-review samples/acme-supplier-nda-draft.md --jurisdiction de ``` The sample NDA contains 5 intentional defects. Expected output: - Clause-by-clause risk table - 2 HIGH risks flagged (unlimited liability, wrong governing law) - Negotiation positions per contested clause - Comparison against `knowledge-base/templates/bmw-nda-standard.md` ### 2. Compliance Check (CLI, ~1.5 min) ```sh lex compliance-check "supply chain partnership XYZ Corp China" --check-type sanctions,lksg ``` - OFAC SDN / EU / UN sanctions screening - LkSG due-diligence checklist (10+ items) - PASS / FAIL / REQUIRES_REVIEW per framework ### 3. Multi-Jurisdiction Research (CLI, ~2 min) ```sh lex legal-research "force majeure pandemic clause enforceability" --jurisdiction de,uk ``` Comparative table with BGH and English High Court citations. All URLs verified by `qa-agent`. ### 4. Web UI (~2 min) ```sh lex-web # open http://localhost:3000 ``` Drag-and-drop contract upload, live agent activity log, artifact viewer with markdown rendering, provenance panel. --- ## Tech Stack - **Runtime**: Pi coding-agent (`@mariozechner/pi-coding-agent`) on Node.js 20+ - **LLM**: Claude Sonnet 4.5 via OpenRouter (configurable to other providers) - **Extension language**: TypeScript (loaded via `tsx` at runtime) - **Web UI**: Next.js 15, React 19, react-markdown - **Storage**: File-based (no database) — `outputs/`, `~/.lex/agent/` - **Tool framework**: TypeBox JSON Schema for tool parameters ### Why file-based storage? For legal work, the audit trail of intermediate artifacts is a feature, not a workaround. Plans, drafts, QA reports, and provenance sidecars all sit on disk in human-readable Markdown — making the system fully auditable, resumable across sessions, and easy to inspect without a database. --- ## Development ```sh git clone https://github.com/itsamirkhon/lex.git cd lex npm install lex --version # Run from source node bin/lex.js --help ``` ### Adding a new agent 1. Create `.lex/agents/your-agent.md` with YAML frontmatter (`name`, `description`, `tools`, `output`) 2. Reference it in workflow prompts under `prompts/` 3. Restart Lex — it auto-discovers agents at startup ### Adding a new tool Edit `extensions/legal-tools/your-tool.ts`, register via `pi.registerTool()`, and re-export from `extensions/legal-tools.ts`. --- ## Open Questions for Production Deployment - **Real Otto-Schmidt / Beck-Online integration** — currently uses targeted web search as fallback - **Vector DB for KB search** — `knowledge-agent` currently uses grep, fine for MVP - **MCP wrapping** — tools are inline Pi extensions; MCP server form would enable cross-agent reuse - **Authentication** — web UI runs on localhost only; production needs SSO + role-based access - **DOCX redline output** — currently emits Markdown; lawyers paste into Word --- ## License MIT — built for the Munich Hacking Legal Challenge by BMW Group. ## Links - Repository: https://github.com/itsamirkhon/lex - Pi framework: https://github.com/mariozechner/pi - Munich Hacking Legal: https://munich-hacking-legal.bmw.com (challenge brief)