anthropics/claude-for-legal: A suite of plugins for legal workflows · GitHub

Reference agents, skills, and data connectors for the legal workflows we see most — in-house commercial, privacy, product, corporate, employment, litigation, regulatory, AI governance, IP, and the learning side of the practice (law school clinics and students).

New here? Start with QUICKSTART.md — install in 60 seconds. This README is the full reference.

Everything here is available two ways from one source: install it as a Claude Cowork or Claude Code plugin, or deploy it through the Claude Managed Agents API behind your own workflow engine. Same system prompt, same skills — you choose where it runs.

What’s in the repo:

• Practice-area plugins covering in-house, firm, and academic legal work — each one built around a cold-start interview that learns your playbook and a CLAUDE.md practice profile that every skill reads from.
• Managed-agent cookbooks for the scheduled, eyes-on-the-feed workflows (renewal watcher, docket watcher, regulatory feed monitor, diligence grid, launch radar).
• MCP connectors across general productivity (Slack, Google Drive, Box) and legal-specific systems (Ironclad, DocuSign, iManage, Everlaw, CourtListener, and more).
• Named agents — end-to-end workflow agents (Vendor Agreement Reviewer, DSAR Responder, Termination Reviewer, Claim Chart Builder, …) with job-style names and a single command to run each one.

Each agent is named for the workflow it runs. They’re the most common surface — start with the ones that match your work, then tune the underlying skill, the practice profile, and the connectors to how your team does it.

For Managed Agent deployment — agent.yaml, leaf-worker subagents, steering-event examples, and per-agent security notes — see managed-agent-cookbooks/.

commercial-legal/         # in-house commercial — vendor/NDA/SaaS review, renewals, escalations
corporate-legal/          # M&A diligence, closing checklists, board consents, entity compliance
employment-legal/         # hire/term review, worker classification, leave, investigations
privacy-legal/            # DPA, DSAR, PIA, privacy triage, policy monitor
product-legal/            # launch review, marketing claims, "is this a problem?" triage
regulatory-legal/         # reg feed watcher, policy diff, gap tracker, NPRM comments
ai-governance-legal/      # AI use-case triage, AIAs, vendor AI review, AI reg gap-check
ip-legal/                 # trademark clearance, FTO, C&D, DMCA, OSS, IP clauses, portfolio
litigation-legal/         # portfolio, matters, holds, demands, depo prep, claim charts
legal-clinic/             # clinic setup, student ramp, intake, deadlines, memos, handoffs
law-student/              # Socratic drilling, outlining, IRAC, bar prep, flashcards
legal-builder-hub/        # community skill discovery and install with a trust gate
external_plugins/         # partner-built plugins maintained by their vendors
  cocounsel-legal/        # Thomson Reuters — Westlaw Deep Research via the CoCounsel Legal MCP
managed-agent-cookbooks/  # Claude Managed Agent cookbooks — one dir per scheduled agent
  diligence-grid/
  docket-watcher/
  launch-radar/
  reg-monitor/
  renewal-watcher/
scripts/                  # deploy-managed-agent.sh · validate.py · orchestrate.py · lint-tool-scope.py
.claude-plugin/
  marketplace.json        # plugin registry

Each plugin directory has the same shape:

/
  .claude-plugin/plugin.json
  CLAUDE.md               # template practice profile — filled in by /:cold-start-interview
  README.md
  skills/                 # skills — each is a /: slash command
  agents/                 # scheduled agents (if any)
  hooks/                  # pre- and post-tool hooks (if any)

In Cowork:

1. Open the Cowork tab.
2. Click Customize in the left sidebar.
3. Click Browse plugins and install the ones you want, or upload a custom plugin file (any plugin directory zipped up).

After install, skills fire automatically when relevant, slash commands are available via /, and the scheduled agents run on the cadence set in their frontmatter.

# Add the marketplace (use the absolute path to this repo or a GitHub URL)
/plugin marketplace add <path-to-this-repo>

# Install a plugin — pick the ones that match your practice
/plugin install commercial-legal@claude-for-legal
/plugin install privacy-legal@claude-for-legal
/plugin install corporate-legal@claude-for-legal

# Restart Claude Code, then run setup for each plugin you installed.
# This writes your practice profile to ~/.claude/plugins/config/claude-for-legal//CLAUDE.md
/commercial-legal:cold-start-interview
/privacy-legal:cold-start-interview
/corporate-legal:cold-start-interview

Run the cold-start interview first. Every other skill in a plugin reads from the practice profile it writes. Skipping setup is the single most common reason a skill produces generic output. The interview takes 10–20 minutes per plugin and will ask you to point at seed documents (a signed MSA, a playbook, a prior review memo — whatever fits the plugin). More seed material is better; a quick start option is available if you want to be productive in 2 minutes and refine later.

Start by connecting a research tool. Everything else is better with one, and citations are unverified without one. See MCP Connectors below for the full list — CourtListener, Trellis, Descrybe, and Solve Intelligence are the research tools the citation guardrails look for.

Updates: /plugin update.

For the scheduled agents — regulatory feed monitor, renewal watcher, docket watcher, diligence grid, launch radar — deploy behind your own orchestrator:

export ANTHROPIC_API_KEY=sk-ant-...
scripts/deploy-managed-agent.sh reg-monitor
scripts/deploy-managed-agent.sh renewal-watcher
scripts/deploy-managed-agent.sh docket-watcher
scripts/deploy-managed-agent.sh diligence-grid
scripts/deploy-managed-agent.sh launch-radar

Each template under managed-agent-cookbooks/ references the same system prompt and skills as its plugin counterpart. The deploy script resolves file references, uploads skills, creates leaf-worker subagents, and POSTs the orchestrator to /v1/agents. See scripts/orchestrate.py for a reference event loop that routes handoff_request events between agents via your own orchestration layer.

Research Preview: subagent delegation (callable_agents) is a preview capability and supports a single delegation level. See per-agent READMEs for security tier and handoff guidance.

Everything is markdown and JSON. No build step.

Grouped by where the work sits. Each plugin’s cold-start interview is what tailors it to your team — start there.

Plugins under external_plugins/ are built and maintained by their vendors. They install from this marketplace like any other plugin, but the vendor owns the code, the connector, and the support channel.

The community is building legal skills fast — registries like LegalOps Consulting’s lpm-skills and Lawvable already list dozens. But nobody certifies community skills, and a lawyer installing a random skill from GitHub is installing code that runs with access to their matter files, their practice profile, and their research connectors.

legal-builder-hub gives the ecosystem the trust layer it’s missing:

• Security review — hidden-content scan, injection detection, structural trust check on every install
• Allowlist — restrictive-by-default source gate (registries, publishers, connectors, licenses)
• License gate — deployment-context-aware license policy (personal / firm-internal / product-embedding)
• Freshness gate — tracks whether bundled reference content (regulations, statutes, procedures) has passed its verification window, and warns at invocation
• Re-scan at update — a skill that was clean at v1.0 and poisoned at v1.1 gets caught
• Install log — an auditable record of what’s installed, from where, under what license, with what review verdict

The allowlist is restrictive by default. Permissive mode is an explicit choice. A non-lawyer gets routed to their attorney contact, not an “install anyway” button.

Community skills go through the same design review (/legal-builder-hub:skills-qa) as the first-party plugins. If you build for lawyers, run the QA against your own skill before publishing. It’s the review a lawyer would do if they could read code.

These plugins ship connectors for the systems legal teams live in. A connector gives Claude the ability to read from and (where scoped) write to your data; the skills and commands use them.

Connectors marked “customer subscription” need the customer’s own account and API key. Configure them in each plugin’s .mcp.json or via claude mcp in your Claude Code setup.

Building a connector? See CONNECTORS.md for what a good legal MCP server looks like and how to submit yours for inclusion.

Lawyers live in Word and Excel. Every contract-touching skill in this repo is authored to work in the Claude for Word sidebar, with tracked changes as the output mode. That’s commercial-legal:review (vendor agreements, NDAs, SaaS subscriptions), commercial-legal:amendment-history, ip-legal:ip-clause-review, ai-governance-legal:vendor-ai-review, privacy-legal:dpa-review, and the diligence extraction in corporate-legal. A reviewer accepts or rejects each change exactly as they would for a human markup — numbering, defined terms, cross-references, and styles are preserved.

The Excel-facing skills produce workbooks that open cleanly: corporate-legal:tabular-review writes a multi-sheet .xlsx with a sources sheet, litigation-legal:claim-chart writes an element-by-element claim chart with citation columns, corporate-legal:entity-compliance writes the compliance register with deadline columns, and commercial-legal:renewal-tracker exports the renewal register sorted by cancel-by date.

Install Claude for Microsoft 365 from Microsoft AppSource. Once installed, the skills from any plugin you’ve enabled are available from the sidebar via /, and connectors are reachable from the same surface. A single thread can span Word, Excel, PowerPoint, and Outlook.

For IT admins deploying the add-in against your own cloud (Vertex AI, Bedrock, or an internal gateway) rather than Anthropic’s API, see the separate claude-for-msft-365-install tooling.

These are reference templates. They get better when you tune them to how your team works — and the customization mechanism is the plugin itself, not a config file buried in a repo.

• Run the cold-start interview. It is the customization mechanism. It asks how your practice works, reads your seed documents, and writes your practice profile. Every other skill reads from that profile. A /commercial-legal:cold-start-interview with five signed MSAs, your playbook, and your escalation matrix will make the review skills noticeably sharper.
• Edit the practice profile. Your profile lives at ~/.claude/plugins/config/claude-for-legal//CLAUDE.md. Edit it directly for small fixes — a wrong escalation threshold, a new integration, a policy update. It survives plugin updates.
• Re-run setup. /:cold-start-interview again for a full re-interview when your practice shifts materially (new jurisdiction, new CLM, new policy).
• Swap connectors. Point .mcp.json at your CLM, DMS, e-discovery platform, launch tracker, HRIS. Skills fall back gracefully when a connector isn’t configured — no silent no-ops.
• Bring your playbook and templates. Drop your terminology, house style, and branded templates into the plugin’s CLAUDE.md and references/. The skills will pick them up.
• Fork skills for house style. Every skill is a markdown file under skills/. Edit the steps, the gates, the output format.
• Add scheduled agents. The agents under /agents/ are markdown with a cron-style schedule. Add your own for the watchers your team needs.

No build step. Everything is markdown and JSON.

The full map across all plugins. The cold-start interview is the first thing to run in any plugin.

Everything here is markdown and JSON. Fork, edit, PR.

• New skill → add it under /skills//SKILL.md with the frontmatter the existing skills use (name, description, argument-hint). Keep the description under 1024 characters — it’s the trigger signal. The skill is invokable as /:. Mark pure-reference skills user-invocable: false.
• New agent → add /agents/.md with scheduling frontmatter and the system prompt. Add a matching managed-agent-cookbooks// if you want headless deployment.
• Community skills → use /legal-builder-hub:skill-installer to test a community skill in your environment. The hub runs /legal-builder-hub:skills-qa against every skill before installing — it scores the skill against the Legal Skill Design Framework (nine design parameters, three legal failure modes, a trust-surface check) and rejects anything that fails.
• Validate cookbooks before pushing → bash scripts/test-cookbooks.sh dry-runs every managed-agent cookbook and lints orchestrator tool scope.

Licensed under the Apache License, Version 2.0.

Copyright 2026 Anthropic PBC.