Forty-odd legal agents that know they aren't lawyers

What it does

Anthropic published a sprawling catalog of Claude plugins and managed agents for legal workflows—NDA triage, vendor contract review, DSAR responses, claim charts, and dozens more. Each agent is named like a job title and reads from a practice-specific CLAUDE.md profile so it can adopt your house style rather than sounding like generic AI. The same skills run either inside Claude Cowork and Claude Code as plugins or behind your own workflow engine via the Managed Agents API.

The interesting bit

The guardrails are the product. Every output is explicitly framed as a draft for attorney review, with source attribution on citations, conservative privilege defaults, and hard gates that block filing or sending until a human signs off. The documentation is refreshingly honest about what the agents cannot do: they do not represent Anthropic’s legal positions and are not a substitute for a lawyer.

Key highlights

• Covers ten practice areas: commercial, corporate, employment, privacy, product, regulatory, AI governance, IP, litigation, and academic legal clinics.
• Includes both on-demand agents and scheduled “eyes-on-the-feed” cookbooks like docket watchers, renewal watchers, and regulatory feed monitors.
• Ships MCP connectors for legal-specific systems (Ironclad, iManage, Everlaw, CourtListener, DocuSign) alongside general productivity tools (Slack, Google Drive, Box).
• Each practice area uses a cold-start interview to learn your playbook, then grounds every skill against that profile.
• The same system prompts and skills run inside Anthropic’s chat products or via the Managed Agents API.

Verdict

In-house counsel and legal-ops teams drowning in repeatable review work should look here first; developers outside the legal world will find it a well-documented example of domain-specific agent design, but little else.