Foundations

Claude Cowork & Skills for Legal Practice

A text-first course on Claude for legal work, firm Skills, MCP connectors, privilege posture, and the operating discipline that keeps AI output out of the firm record until it earns its way in.

10 modules~102 min totalFree, no signup

For: Practicing attorneys at solo to mid-size firms, firm operators, and legal-aid staff adopting Claude for the first time or trying to use it responsibly at scale.

Last revised June 19, 2026. Maintained as the legal AI industry evolves, on a best-effort basis as new releases and guidance land.

Independent educational material. FlowCounsel is not affiliated with or endorsed by Anthropic.

Use the frontier.

Keep the record.

Bound the context.

Verify sources.

Review before effect.

Before you build anything on Claude, you need to understand the 2026 legal-workflow direction Anthropic has publicly signaled and where the gaps remain. The legal AI landscape changed materially this year, and the operating discipline required of firms changed with it.

Claude Cowork Legal plugin coverage

Public coverage describes Anthropic moving legal workflow primitives closer to Claude rather than only exposing Claude as an API to legal-tech vendors. Treat those reports as the market signal, not as a substitute for reading Anthropic's current product documentation and your own contract terms. The load-bearing assumption remains that an attorney reviews every output before it leaves the firm.

Claude For Legal public-source read

The public anthropics/claude-for-legal repository and contemporaneous legal-tech coverage describe a suite of practice-area plugins, recurring-workflow agent patterns, MCP connectors, and managed-agent deployment concepts. Critical detail: this is not the same thing as a new model. It is an operating layer of skills, agents, connectors, and workflow patterns around Claude. Counts, connector availability, and product packaging can change, so treat the repository and official documentation as the current source of truth.

Model choice is not the control layer

Use the strongest appropriate Claude model available under the firm's approved deployment and contract terms. Better models may reduce some failure modes, but they do not remove the legal operating discipline: source verification, refusal when uncertain, attorney review, and provenance before anything enters the firm record.

The 12 practice-area plugins

Public-source review shows legal workflows grouped around commercial, corporate, employment, privacy, product, regulatory, AI governance, IP, litigation, education, clinic, and builder use cases. The practical lesson does not depend on a locked count: any plugin or skill layer is only as good as the setup discipline behind it. Firms need playbooks, escalation chains, risk calibration, house style, and review gates, not just a named plugin.

MCP connectors and firm-system boundaries

Connector availability changes. The durable point is architectural: MCP and adjacent connector patterns can let Claude read from or write to firm systems under scoped permissions. Treat every connection as a new security boundary. Confirm the current connector list, permissions, logging, retention, and write capability before connecting a legal workflow.

The Managed Agents API

Managed-agent concepts matter because they move legal AI from chat into scheduled or programmatic workflow. That creates leverage and risk at the same time. Any production deployment needs an explicit tool contract, authorization boundary, logging, refusal state, and attorney-review rule before outputs can affect firm records or client communications.

What is NOT in the box

Turnkey deployment. Bar association legitimacy. Firm-IP compounding (a layer that keeps the skills you build accruing to your firm rather than to a model provider). Network effect across firms. Cross-firm intake routing. Pro bono distribution. The technical and operational gap is real, and most of this course is about how to close it for your firm.

Key takeaways

Anthropic's legal-workflow push should be treated as a major market signal, but current product documentation and contract terms remain the source of truth.
Claude For Legal is better understood as an operating layer of skills, agents, connectors, and workflow patterns around Claude, not as a new model.
Each plugin has a setup interview that learns your firm. Skip the interview and the plugin underperforms.
Model quality does not eliminate the need for attorney review of every output.
The foundational primitives are not the same thing as firm operating infrastructure. Governance, deployment, firm IP, network effects, and review state still need a home.

Sources