法律代理编排器

Prerequisites

• Claude Code installed and logged in. Max subscription strongly recommended — a single case can burn 200K+ tokens across subagents, and on metered API pricing that adds up. On Max, marginal cost is zero.
• macOS or Linux with git, bash or zsh, and python3 (3.10+).
• 법제처 Open API account. Free, sign up with an email. You'll get an LAW_OC key which the korean-law MCP server uses to query Korean statutes, precedents, and administrative interpretations in real time.

Getting Started

2. Install the four subordinate agents

./setup.sh

This script shallow-clones all four specialists' GitHub repositories into agents/ under their Agent ID names, tracking each one's main branch:

agents/
├── legal-research-agent/       ← Legal Research Specialist (general + game)
├── legal-writing-agent/        ← Legal Writing Specialist
├── second-review-agent/        ← Senior Review Specialist
└── data-protection-agent/      ← Data Protection Specialist

Each folder is an independent Claude Code agent with its own CLAUDE.md, skills/, knowledge base, and MCP configuration. When the orchestrator dispatches a case, it calls into these agents via Claude Code's Agent tool with cwd: agents/{agent-id}/, so each subagent runs in its own working directory with its own context.

Other setup.sh commands: - ./setup.sh update [agent-id ...] — fast-forward every agent, or only the listed agents, to the latest main of its upstream repository (idempotent; same as the default ./setup.sh when no agent IDs are supplied). - ./setup.sh status [agent-id ...] — show each selected agent's local SHA next to the upstream main SHA, and flag any that are behind / up to date / unreachable / symlink (dev mode) - ./setup.sh link [agent-id ...] — development mode: if you already have the selected agent repositories checked out under ~/코딩 프로젝트/, create symlinks instead of fresh clones so your local edits flow through immediately

During case execution the orchestrator now resolves route-specific sync targets with scripts/resolve-sync-targets.py and applies TTL-aware sync with scripts/sync-agents.py, so out-of-scope cases and repeated runs avoid unnecessary network work.

Each agent is shallow-cloned (--depth 1 --single-branch), so only the latest snapshot of main lives on disk — no git history, no other branches.

2. "Why not LangGraph or Agent SDK?"

Wrapping existing Claude Code agents in a web framework loses 40–50% of their capability: MCP breaks, the skills system needs reimplementation, knowledge-base browsing changes. You end up with a pretty demo producing legal opinions at half quality.

We inverted the tradeoff: Claude Code as the runtime, agents preserved 100% intact, and final delivery collapsed into a single case-report.md artifact instead of a web UI. Real legal work, not a demo.

3. Set your Korean Open Law API key

export LAW_OC=your_law_oc_key

⚠️ This is required every shell session. Claude Code does not auto-load .env, and without LAW_OC the korean-law MCP server will fail to answer the first statute lookup. The simplest solution is to put export LAW_OC=... in your ~/.zshrc or ~/.bashrc.

Comparison

---

FAQ

How does it handle client confidentiality? Everything runs locally on your machine under your own Claude Code session. No intermediate SaaS. Claude Code itself sends prompts to Anthropic for inference — whether that's acceptable for a given matter depends on your firm. output/, agents/, and .env are gitignored so case files and API keys don't leak into commits.

What does ./setup.sh actually do to my machine? It creates an agents/ folder inside this repository and clones four GitHub repositories into it, one per specialist agent. Nothing outside this directory is touched. No global package installs, no environment mutations beyond whatever git clone does. Each agent folder is roughly 10–80 MB depending on its knowledge base size.

Can I add my own specialist agent? Yes. Write it as a standalone Claude Code agent (its own CLAUDE.md, skills/, optional library/, optional .mcp.json), drop it under agents/ (or symlink it), add one line to the REPOS array in setup.sh, and add one row to skills/route-case.md so the router knows when to call it. No orchestrator code changes needed. The design is plugin-shaped.

How much does it cost per opinion? On Claude Code Max: zero marginal dollars. On metered API pricing: roughly $3–10 per opinion depending on complexity. The real cost is wall-clock time (5–15 minutes per pipeline).

---