How anthropics/claude-for-legal works
1. This is a content-first plugin marketplace, not a traditional app.
The repository packages legal workflows as Markdown/JSON bundles that can run either in Claude Cowork/Claude Code as plugins or as headless Managed Agents behind a separate orchestrator. There is no application backend or build system; the repo is the source of truth for prompts, skills, agents, hooks, and connector configuration.
2. The top-level boundary is the marketplace registry.
.claude-plugin/marketplace.json is the catalog entry point. It points users to install individual plugin directories like commercial-legal, litigation-legal, or legal-builder-hub. Each plugin is self-contained and versioned by the repo, so architecture diagrams should treat each plugin directory as an independent deployment unit.
3. Every plugin follows the same internal shape.
A plugin typically contains:
- .claude-plugin/plugin.json for packaging/metadata
- CLAUDE.md as the practice profile
- skills/ for slash commands and reusable workflows
- agents/ for scheduled or background workflows
- hooks/ for tool-event interception
- references/, data/, logs/, or matter folders for local state and templates
This repeating structure is the main architectural pattern across the repo.
4. The cold-start interview is the customization boundary.
Each practice-area plugin starts with a cold-start-interview skill that writes a tailored CLAUDE.md under the user’s plugin config. That practice profile becomes the shared policy and style layer that all other skills read from. In diagram terms, this is the personalization step between generic plugin code and firm-specific behavior.
5. Skills are the primary execution primitive.
Most user-facing workflows are implemented as skills/<skill>/SKILL.md, exposed as slash commands like /commercial-legal:review or /litigation-legal:claim-chart. Skills are mostly domain logic and output formatting; they depend on the practice profile, referenced templates, and available connectors. Some plugins also include non-command reference skills used internally by other skills.
6. Agents handle scheduled or event-driven work.
Background workflows live in agents/ and run on cadence or on a trigger. Examples include renewal watchers, docket watchers, regulatory feed monitors, and launch watchers. These agents usually read from an external system or local register, apply rules, and then write a digest, status update, or alert. They are the automation layer distinct from interactive skills.
7. Managed-agent cookbooks are the headless deployment model.
managed-agent-cookbooks/ mirrors certain agents with agent.yaml, steering examples, and depth-1 subagents. This is the deployable version used by scripts/deploy-managed-agent.sh. The repo also includes scripts/orchestrate.py, which shows the intended event flow for routing handoff_request events between agents through a separate orchestration layer.
8. Connectors form the data plane.
Each plugin’s .mcp.json defines MCP connections to external systems such as Slack, Google Drive, Box, Ironclad, DocuSign, iManage, Everlaw, CourtListener, Trellis, Aurora, Linear, Jira, Asana, and others. Skills and agents consume these connectors to read source material, verify citations, pull registers, or write outputs. This is the main boundary between the repo’s prompts and real enterprise data.
9. There is a strong trust and governance subsystem around community skills.
legal-builder-hub is the ecosystem’s trust layer for third-party skills. It includes registry browsing, skill installation, QA, update checks, and a registry-sync agent. The hub enforces allowlists, license checks, freshness checks, and re-scanning on update, so in an architecture diagram it should sit as a policy gate in front of community-sourced code.
10. The repo is organized by legal practice area, with each plugin reflecting a workflow cluster.
Examples:
- commercial-legal: contract review, amendment tracing, renewals, escalation routing
- corporate-legal: diligence, board consents, closing, entity compliance, integration
- employment-legal: hiring, terminations, investigations, leave tracking
- privacy-legal: DPA review, DSARs, PIAs, policy drift
- litigation-legal: matter lifecycle, holds, demand letters, claim charts, dockets
- law-student and legal-clinic: education, supervision, intake, deadlines, memos
Each plugin is a bounded workflow domain rather than a shared platform service.
11. State is mostly file-based and domain-local.
Several plugins keep their own local artifacts, such as logs, trackers, registers, matter folders, deadline YAML, or reference templates. These are not database-driven services; they are document-centric workflows where state lives in Markdown, YAML, or spreadsheet outputs and is updated by skills/agents over time.
12. The operational tooling is lightweight and repo-centric.
scripts/validate.py, scripts/lint-tool-scope.py, scripts/test-cookbooks.sh, and scripts/deploy-managed-agent.sh support validation and deployment of the cookbook/agent model. The GitHub Actions workflow likely serves as CI for these checks. This reinforces that the repo’s primary lifecycle is editing markdown, validating structure, and deploying definitions.
13. A useful architecture diagram should show three major surfaces.
First, interactive plugin use in Claude Cowork/Code via skills and commands. Second, background agents deployed through the Managed Agents API and an external orchestrator. Third, MCP-connected enterprise systems that supply facts and receive outputs. Everything else in the repo supports one of those three surfaces.
14. The main design principle is “same source, different runtimes.”
The same legal workflow content can be used as an installable plugin or as a managed agent cookbook. The repo intentionally keeps prompts, playbooks, and workflow structure in one place so the interactive and scheduled variants stay aligned.
Main entry points: marketplace.json Marketplace · managed-agent-cookbooks Cookbooks · commercial-legal Commercial · privacy-legal Privacy
Components 21
Registry
Marketplace entry points and shared packaging metadata.
- Marketplace · Top-level marketplace registry for installable plugins.
.claude-plugin/marketplace.json
Plugins
Installable legal workflow bundles used interactively in Claude.
- Commercial · Contract review, renewals, escalation, and stakeholder workflows.
commercial-legal - Corporate · Diligence, closing, entity compliance, and deal support workflows.
corporate-legal - Employment · Hiring, terminations, investigations, leave, and policy workflows.
employment-legal - Privacy · DPA review, DSARs, PIAs, policy drift, and regulatory analysis.
privacy-legal - Litigation · Matter lifecycle, holds, demands, claim charts, and docket work.
litigation-legal - IP · Portfolio, clearance, FTO, infringement, and takedown workflows.
ip-legal - Product · Launch reviews, feature risk, and marketing claims workflows.
product-legal - Regulatory · Regulatory monitoring, gaps, comments, and policy redrafting.
regulatory-legal - Builder Hub · Registry browsing, skill installation, QA, and allowlist enforcement.
legal-builder-hub - Legal Clinic · Client intake, deadlines, drafting, supervision, and handoff workflows.
legal-clinic - Law Student · Case briefing, outlines, flashcards, drills, and exam prep.
law-student - CoCounsel · External community plugin bundled as a separate installable unit.
external_plugins/cocounsel-legal
Managed Agents
Headless cookbook deployments for scheduled or triggered work.
- Cookbooks · Headless deployment model for managed agents.
managed-agent-cookbooks - Diligence Grid · Document extraction and diligence grid assembly pipeline.
managed-agent-cookbooks/diligence-grid - Docket Watcher · Deadline extraction and docket tracking pipeline.
managed-agent-cookbooks/docket-watcher - Launch Radar · Launch monitoring and risk classification pipeline.
managed-agent-cookbooks/launch-radar - Reg Monitor · Feed ingestion and materiality filtering pipeline.
managed-agent-cookbooks/reg-monitor - Renewal Watcher · Renewal deadline calculation and alert pipeline.
managed-agent-cookbooks/renewal-watcher
Tooling
Repo-level validation, orchestration, and deployment support.
- Scripts · Validation, orchestration, testing, and deployment helpers.
scripts
Enterprise Systems
External data sources and downstream systems accessed through MCP connectors.
- Enterprise Systems · External systems accessed through MCP connectors and outputs.
Connections 27
- Marketplace → Commercial: lists · Marketplace entry installs the commercial plugin.
- Marketplace → Corporate: lists · Marketplace entry installs the corporate plugin.
- Marketplace → Employment: lists · Marketplace entry installs the employment plugin.
- Marketplace → Privacy: lists · Marketplace entry installs the privacy plugin.
- Marketplace → Litigation: lists · Marketplace entry installs the litigation plugin.
- Marketplace → IP: lists · Marketplace entry installs the IP plugin.
- Marketplace → Product: lists · Marketplace entry installs the product plugin.
- Marketplace → Regulatory: lists · Marketplace entry installs the regulatory plugin.
- Marketplace → Builder Hub: lists · Marketplace entry installs the builder hub.
- Commercial → Enterprise Systems: connects · Skills and agents read/write via MCP connectors.
- Corporate → Enterprise Systems: connects · Skills and agents read/write via MCP connectors.
- Employment → Enterprise Systems: connects · Skills and agents read/write via MCP connectors.
- Privacy → Enterprise Systems: connects · Skills and agents read/write via MCP connectors.
- Litigation → Enterprise Systems: connects · Skills and agents read/write via MCP connectors.
- IP → Enterprise Systems: connects · Skills and agents read/write via MCP connectors.
- Product → Enterprise Systems: connects · Skills and agents read/write via MCP connectors.
- Regulatory → Enterprise Systems: connects · Skills and agents read/write via MCP connectors.
- Builder Hub → Enterprise Systems: validates · Trust layer governs community skill intake and freshness.
- Legal Clinic → Enterprise Systems: connects · Clinic workflows use connectors for documents and state.
- Law Student → Enterprise Systems: connects · Study workflows are content-first and connector-light.
- CoCounsel → Enterprise Systems: connects · External plugin integrates through MCP connections.
- Commercial → Cookbooks: mirrors · Same workflow can be deployed as a managed-agent cookbook.
- Litigation → Cookbooks: mirrors · Scheduled legal workflows have headless equivalents.
- Privacy → Cookbooks: mirrors · Content and policy workflows can run headlessly.
- Regulatory → Cookbooks: mirrors · Monitoring workflows are mirrored in cookbook form.
- Builder Hub → Scripts: uses · Repo tooling supports validation and deployment.
- Cookbooks → Scripts: uses · Deployment and orchestration tooling manages cookbooks.