Anthropic внедряет строгие правила проверок и доверенных подключений для юридических клиентов

Based on the comprehensive documentation you provided, here is a structured synthesis of the Claude for Legal ecosystem, its architecture, safety protocols, and available capabilities. This summary organizes the raw data into a functional guide.

A modular ecosystem of plugins designed to integrate legal workflows directly into existing tools (Word, Excel, Slack, etc.) while enforcing strict trust, safety, and jurisdictional compliance.

The system prioritizes security and accuracy over speed, specifically addressing the risks of unvetted community code.

Before any community skill runs, it passes a multi-gate security review: * Security Review: Hidden-content scanning, injection detection, structural trust checks. * Allowlist Gate: Restrictive-by-default. Non-lawyers are routed to attorneys; "install anyway" buttons are disabled for unvetted code. * License Gate: Deployment context awareness (Personal vs. Firm vs. Product). * Freshness Gate: Ensures bundled reference content (statutes/regulations) has passed verification windows. * Re-scan at Update: Detects code poisoning between versions (e.g., v1.0 clean $\to$ v1.1 poisoned). * Audit Logs: Every install is recorded (Source, License, Review Verdict).

Connector-First: Citations from model training knowledge alone are flagged [verify].
Authoritative Sources: Plugins ship with connectors (CourtListener, Trellis, Descrybe) that pull from live databases.
Visual Tagging: Citations are tagged with their source. If no connector is active, a reviewer note explicitly states sources were not verified.

Focuses on managing AI usage within firms and compliance with regulations. * Key Skills: * use-case-triage: Classifies AI use cases (Approved/Conditional/No). * vendor-ai-review: Reviews vendor AI terms against firm governance. * reg-gap-analysis: Compares new regulations against internal posture. * policy-starter: Drafts firm-wide AI usage policies. * matter-workspace: Manages governance-specific matter tags. * Agents: policy-monitor (keeps policies current), reg-gap-analysis (continuous scanning).

Focuses on contract lifecycle, due diligence, and corporate structuring. * Commercial (commercial-legal): * vendor-agreement-review: Reviews MSAs, NDAs, SaaS terms with tracked changes. * renewal-tracker: Identifies contracts expiring within 90 days. * amendment-history: Traces changes across base agreements and amendments. * Corporate (corporate-legal): * tabular-review: Generates .xlsx reports (one row per doc) with citations. * diligence-issue-extraction: Extracts M&A issues from VDRs. * closing-checklist: Identifies critical path blockers. * entity-compliance: Tracks compliance registers across jurisdictions.

Jurisdiction-aware handling of hiring, termination, and benefits. * Key Skills: * worker-classification: Classifies contractors vs. employees against state tests. * wage-hour-qa: Q&A on overtime/FLSA per jurisdiction. * investigation-open: Manages internal investigations with privileged memo scaffolding. * leave-tracker: Monitors FMLA/CFRA/PFL deadlines. * expansion-kickoff: Plans international expansion logistics.

Key Skills:
- clearance: Trademark knockout and similarity search.
- fto-triage: Freedom-to-operate triage (not a formal opinion).
- cease-desist: Drafts C&Ds or triages received ones.
- takedown: DMCA notice drafting and §512(g) counter-notices.
- oss-review: Open source license compliance checks.
- portfolio: Tracks renewal dates and expiration alerts.
Key Skills:
- matter-intake: Creates matter.md and history files upon intake.
- demand-draft: Drafts demand letters with FRE 408 gate checks.
- subpoena-triage: Analyzes burden, privilege, and scope.
- legal-hold: Manages hold issuance, refreshment, and releases.
- claim-chart: Builds cause-of-action charts (civil/patent).
- chronology: Builds event timelines from sources.

Built on ABA Formal Opinion 512 principles. Learning mode, not answer mode. * Key Skills: * client-intake: Structured intake with issue spotting. * memo: IRAC-scaffolded analysis with research gaps flagged. * supervisor-review-queue: Manages review queues for formal supervision. * semester-handoff: End-of-term case transfers.

Capabilities: Socratic drilling, case briefing, IRAC grading, cold-call prep, flashcards, exam forecasting.

Skills require specific data connectors to function. These are configured via .mcp.json or the CLI.

Category Connector Provider Function Collaboration Slack All Plugins Read/search channels, send messages, create canvases. Cloud Storage Google Drive All Plugins Read docs, sheets, slides via link. Research CoCounsel Legal Thomson Reuters Westlaw Deep Research (caselaw, statutes, Practical Law). Contracts Box Corporate Read VDRs and matter rooms. Contracts Ironclad Commercial Read contract register, renewals, clauses. E-Sign DocuSign / CLM Commercial Envelope status, executed contracts. DMS iManage Corporate/Lit Read matter workspaces, doc versions. Discovery Everlaw Litigation E-discovery productions, tagged sets. Dockets CourtListener Public/Optional Federal dockets and opinions. Dockets Trellis Litigation State court dockets and motions. Clinic Aurora Litigation Clinic-style matter management. Drafting Definely Commercial In-doc drafting and defined-term checks. Legal Tech Courtroom5 Clinic Self-represented litigant workflow. Research Descrybe Clinic/IP/Law Case law research and summarization. Patents Solve Intelligence Corp/IP Patent drafting and prosecution. Routing TopCounsel Corp/Lit Matter routing and outside counsel panels. Project Mgmt Linear Product Launch tracking, issue tracking. Project Mgmt Jira Product Launch tracking. Project Mgmt Asana Product Project tracking.

> Note: Connectors marked "Customer Subscription" require the user's own API key.

Every plugin begins with a Cold-Start Interview (/<plugin>:cold-start-interview). * Function: Reads seed documents (MSAs, playbooks, escalation matrices) to build a Practice Profile. * Output: Writes the profile to ~/.claude/plugins/config/claude-for-legal/<plugin>/CLAUDE.md. * Benefit: Subsequent skills immediately adapt to your specific terminology, house styles, and escalation rules.

Small Fixes: Edit CLAUDE.md directly (e.g., updating a threshold or policy).
Material Shifts: Run the interview again (e.g., entering a new jurisdiction or CLM system).
Templates: Drop branded templates and terms into references/ within the plugin directory.
Scheduled Agents: Add cron-style agents to agents/ for automated sweeps (e.g., renewal-watcher, playbook-monitor).
Connector Swapping: Point .mcp.json at new systems (e.g., switching from Ironclad to a different CLM). Skills handle disconnection gracefully.
Word/Excel Integration: Skills output to the Claude for Word sidebar with tracked changes or generate clean .xlsx files for Excel.
SaaS: Use Anthropic's API directly.
On-Prem/Private Cloud: Use claude-for-msft-365-install for Vertex AI, Bedrock, or internal gateways.

Install Claude for Legal via Microsoft AppSource (for Word/Excel) or via CLI (for macOS/Linux/Windows terminal).
Configure Connectors: Run /claude connect or edit .mcp.json to authorize research tools (CourtListener, Trellis, etc.) and internal systems (Box, Ironclad).
Run Cold-Start: Execute the specific interview for your practice area (e.g., /commercial-legal:cold-start-interview).
Enable Plugins: Turn on desired plugins in settings.
Verify Research: Ensure a research connector is active before relying on citations; check for [verify] flags.

This architecture ensures that legal professionals can leverage powerful AI capabilities without compromising on data security, citation accuracy, or ethical compliance.