Anthropic released Claude for Legal — a single Apache 2.0 GitHub repository containing 13 vertical plugins, 70+ named agents, and 30+ MCP connectors covering every major legal practice area. With a built-in Westlaw connector maintained by Thomson Reuters and output formats native to lawyers (Word tracked changes, multi-sheet Excel), this isn't a SaaS product — it's a free scaffold that commoditizes the multi-billion-dollar legal AI market and templates a pattern for every regulated vertical that follows.
Legal AI is a multi-billion-dollar category. Companies like Harvey, Habia, Spellbook, and CoCounsel sell six-figure subscriptions for contract redlining, case summarization, and model prompting. Harvey alone has 60+ Am Law 100 firms on contract.
The disruption — that entire market just got a free reference implementation from Anthropic, with Westlaw access included.
The repo — anthropic/claudeforlegal, Apache 2.0 license. Fully open source.
Everything is markdown and JSON — no build step. Skills are just files. "You can read the entire product with cat."
Key insight: "Legal AI is a multi-billion-dollar category. That market just got a free reference implementation from Anthropic with Westlaw access included."
The repository maps directly to the legal practice landscape with 13 vertical plugins covering every major area of law.
Practice areas — Commercial, Corporate, Employment, Privacy, Product, Regulatory, AI Governance, IP, Litigation
Law Clinic plugin — designed for clinical program staff in legal education
Law Student plugin — includes a skill called "Socratic Drill" that asks questions, pushes back, and never gives the answer. "Anthropic shipped a 1L professor."
30+ MCP connectors — every legal data system you've heard of, pre-wired
3 The Thomson Reuters Play — Westlaw Goes Free ▶ 1:36
The move the legal tech industry didn't see coming: Thomson Reuters, the company that owns Westlaw, builds and maintains the CoCounsel legal MCP connector — and it's distributed through Anthropic's marketplace for free.
What it connects to — case law, statutes, regulations, practical law
Verified citations — three jurisdictions per run
The biggest legal data moat in the industry — now accessible as a free MCP connector
Key insight: "The biggest legal data moat in the industry, distributed through Anthropic's marketplace, for free."
Every plugin's first command is a cold start interview — a pattern developers should steal for any vertical AI deployment.
The interview — asks about your playbook, house style, escalation rules. Reads your seed documents.
The output — writes a customized Claude.md practice profile
The mechanism — every single downstream skill in that plugin reads from this one file. "One file, the customization mechanism is the plugin."
Output format is the differentiator — tracked changes in Microsoft Word sidebar, Excel with multi-sheet and sources tab sorted by cancel-by date. "Lawyers don't live in VS Code, they live in .docx."
Key insight: "The model isn't the differentiator anymore. The output format is. Read the room — Anthropic did."
5 Security — Legal Builder Hub and Four Gates ▶ 2:54
Community skills flow through the Legal Builder Hub, where every inbound skill must pass four security gates.
Gate 1 — Hidden content scan
Gate 2 — Prompt injection detection
Gate 3 — Restrictive-by-default allow list
Gate 4 — Rescan on every update
Non-lawyer routing — a non-lawyer gets routed to their attorney contact, not an "install anyway" button
6 Anthropic's Real Strategy — Win the Infrastructure ▶ 3:23
Why ship this for free? Because Anthropic isn't trying to win the legal app — they're trying to win the legal infrastructure.
Dual deploy — every skill runs two ways from one source: as a Claude Co-Work plugin, or as a managed agent API deployment behind your own orchestrator. "Same system prompt, same skills, you choose where it runs."
The translation — the legal app layer is now a commodity. The model and the agent runtime — that's where Anthropic gets paid.
This does NOT replace lawyers — the repo's first rule: "Every output is a draft for attorney review, not legal advice." Source attribution on every citation, conservative privilege defaults, jurisdiction assumptions surfaced, explicit send gates before anything ships.
Key insight: "The legal app layer is now a commodity. The model and the agent runtime — that's where Anthropic gets paid."
The video closes with four architectural patterns from the repo that apply to any vertical AI deployment — not just legal.
Skills as markdown files — greppable, reviewable, forkable. No compiled artifacts, no build steps.
Claude.md as practice profile — one file that every skill reads from. The cold start interview writes it, all downstream skills consume it.
MCP for tools — vendor-neutral plumbing. Swap data sources without changing skills.
Dual deploy — same skill works as a plugin or a managed agent. Pick where it runs.
Key insight: "Anthropic just templated vertical AI. Medical, finance, ops — every one of those gets this pattern next."
🎯 Key Takeaways
70+ free legal AI agents — Anthropic released anthropic/claudeforlegal under Apache 2.0, covering 13 practice areas with 30+ MCP connectors pre-wired.
Westlaw access for free — Thomson Reuters maintains the CoCounsel MCP connector with verified citations across three jurisdictions, distributed through Anthropic's marketplace at no cost.
Skills are just markdown — no build step, no compiled artifacts. The entire product is readable with cat. This makes it greppable, reviewable, and forkable.
Cold start interview pattern — every plugin interviews the user to create a customized Claude.md practice profile. All downstream skills read from this one file. Steal this pattern.
Output format is the differentiator — Word tracked changes, multi-sheet Excel with sources tabs. Anthropic understood that lawyers live in .docx, not VS Code.
Four security gates — hidden content scan, prompt injection detection, restrictive allow list, rescan on update. Non-lawyers get routed to attorneys, not "install anyway" buttons.
Dual deploy architecture — same skill runs as a Claude Co-Work plugin or a managed agent API behind your own orchestrator. One source, two runtimes.
The app layer is now commodity — Anthropic's play isn't the legal app, it's the infrastructure. The model and agent runtime are where they get paid.
Augmentation, not replacement — every output is a draft for attorney review. Source attribution, conservative privilege defaults, jurisdiction surfacing, and explicit send gates are baked in.
This templates every vertical — medical, finance, ops. The four patterns (markdown skills, practice profiles, MCP tools, dual deploy) apply to any regulated industry.