Relex Claude — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited Relex Claude (Plugin) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Legal Workspace — One source of truth for any Agent — confidential by design
Keep legal knowledge, matter context and saved progress in Relex, independently of the assistant you use. Authorize another compatible agent to continue from the same stored context, without rebuilding the background in another chat.
intake and document workflows.
https://relex.legal/api/mcp and authorizeyour own Relex account. Installing a package does not authorize private data.
its conclusions when finished. A second authorized client can then use that continuing record.
Portability covers information saved in Relex, not automatic import of private chat histories or a model's internal memory. Client-side identity encryption, de-identification and MCP access controls protect the supported workflows; de-identified legal facts may still be sensitive. Review what you authorize.
Use Legal Workspace for persistent legal context; the Legal SDK for building your firm's or legal department's own platform; and the Legal Marketplace to discover published professional profiles or make an AI-first law firm discoverable across specialties.
An agent may help find a professional and prepare a reference-only request. The user must review and approve sharing in Relex. Discovery is not engagement, a completed conflict check, payment or a guarantee of professional availability.
Client support depends on the host product, plan and administrator settings. Gemini CLI support does not imply support in every Gemini web experience. Harvey BYOMCP is a customer-admin connection path, not a claim of Harvey Connector Library listing or approval. Check the current connector guides and portable-context guide.
Claude connects to a remote MCP server that exposes thirteen tools over Streamable HTTP (matter workflows plus generic search / read_api / execute). search({query,tag,method}) discovers endpoints from the OpenAPI spec; read_api({path,query}) reads one GET endpoint and execute({method,path,query,body}) creates, updates or deletes through one — the server validates each call against the spec, applies the PII guard, and runs it with the user's auth.
That MCP server is served by the Relex backend at https://relex.legal/api/mcp (the SvelteKit /api/* proxy forwards /api/mcp → backend /v1/mcp). Auth is a browser sign-in over OAuth 2.1 + PKCE: on the first tool call the server returns an RFC 9728 challenge, the client opens the user's browser to sign in with Google/Apple and approve, and a per-connection credential is provisioned — no key paste. A static Relex API key (Settings → API Keys) works as a CI/headless fallback. No separate service and no DNS to provision.
Party data — names, IDs, contacts — is sealed client-side with a key derived from the user's PII password; the server stores only ciphertext and cannot decrypt it under any circumstance — a cryptographic guarantee, not a policy choice. Document content is read and redacted by the Relex Private Model on Relex servers by default (browser reading mode keeps it on the device); Claude only ever receives redacted source text. On top of that, read_api and execute refuse any call that would return party or document plaintext and instead hands the user a deep link to do it securely in the Relex UI — enforced at the API boundary, not by convention.
The hosted connector at https://relex.legal/api/mcp exposes thirteen tools: list_matters, read_matter_context, read_matter_source, diagnose_matter_sources, save_matter_work_product, correct_matter_ontology, conclude_matter_session, find_legal_professionals, read_legal_professional, prepare_professional_request, search, read_api, and execute. Auth is OAuth 2.1 + PKCE. The default connector scopes are relex.cases.read relex.cases.write relex.draft; these are the scopes the 401 WWW-Authenticate challenge requests. relex.agent is an optional fourth scope for agent features; the challenge omits it by design, and the OAuth metadata (scopes_supported) and the MCP server card list all four.
Claude.ai / Claude desktop callback: https://claude.ai/api/mcp/auth_callback (also claude.com). Claude Code uses loopback http://localhost/callback / http://127.0.0.1/callback (any port).
relex-claude/
├── plugin/ Claude Code plugin "relex-legal"
│ ├── .claude-plugin/plugin.json
│ ├── .mcp.json Remote MCP connector (https://relex.legal/api/mcp; OAuth sign-in)
│ ├── commands/ /relex-setup (guided onboarding) + /relex-connect
│ ├── skills/
│ │ ├── relex/ Drive Relex over MCP + PII discipline (auto-loaded)
│ │ ├── relex-steering/ The steering-session protocol: directives → steering block → conclude/distill
│ │ ├── relex-counsel/ Senior-counsel role + oversight loop over the harness
│ │ ├── relex-ontology/ The read→audit→repair→acquire→converge collaboration loop
│ │ ├── relex-research/ You discover (web + public legal MCPs); the harness grounds
│ │ ├── relex-citations/ 3-tier epistemic labels + hard citation locks + anchors
│ │ ├── relex-matter/ Deadlines, timeline, conflicts, comms log, closing
│ │ ├── relex-participants/ Who's who as labels; sealed parties vs app participants; Claude-in-Slack
│ │ ├── relex-intake/ End-to-end client intake: request → agreement → e-sign → invoice
│ │ ├── relex-partner/ Guides a firm through partner-program registration (to charge clients)
│ │ └── jurisdictions/ Per-jurisdiction packs: US DE CH UK FR IT ES RO EU CA JP AU
│ ├── agents/relex-guide.md Onboarding subagent
│ └── references/ interop · deliverables (claude-for-legal etc.)
├── scripts/validate_skills.py Skill + jurisdiction-pack validator
├── SECURITY.md
└── docs/
├── install.md
├── connect-claude-code.md
├── connect-claude-desktop.md
├── connect-claude-tag.md
├── connect-codex.md
└── positioning.mdClaude acts as senior counsel + steering layer; the Relex backend agents are the execution harness. Claude doesn't produce the case work itself — it runs steering sessions: each case_req turn is a structured directive on a private branch of the case timeline; the case agent executes and teaches Claude the platform back through a steering block (guidance, missing data, suggested next calls); when Claude concludes the session, a distilled high-fidelity summary lands on the main case thread, attributed to the user "via Claude". They also collaborate through the case ontology (the shared understanding — Claude audits it for gaps, repairs it, and directs acquisition). Division of labour: you discover public authority (your own web search + public legal MCPs — CourtListener, EUR-Lex, entscheidsuche, NeuRIS…), and the harness grounds it — a draft may only cite law whose verbatim text sits in the cache (the server verifier enforces it), so citations don't hallucinate. Every legal assertion is labelled secured / verify / never-use; hard citation locks (no proprietary blind-cites, no docket numbers from memory) keep filings clean. Twelve jurisdiction packs carry the citation schema, official sources, compliance limits, method and limitation heuristics per system. Designed to run hand-in-hand with Anthropic's claude-for-legal (generic drafting playbooks) — Relex adds the confidential workspace + grounding.
The MCP handler itself (the search/read_api/execute tools, descriptor validation, and the PII guard) lives in the Relex backend (/v1/mcp), not in this repo. This repo is just the plugin + docs.
/plugin marketplace add relexlegal/relex-claude then /plugin install relex-legal@relex.paste); Claude walks you through PII password → knowledge → auto-created encrypted parties → first case. Full guide: docs/install.md (per-client: connect-claude-code.md, connect-claude-desktop.md, connect-claude-tag.md, connect-codex.md).
On Pro / Max you add the custom connector yourself (https://relex.legal/api/mcp). On Team / Enterprise an owner or admin adds it once in the organisation’s Connectors settings; each member only opens Customize → Connectors, finds Relex already listed, and clicks Connect. Details: docs/connect-claude-desktop.md.
Same MCP server, different packaging: relex-gpt · relex-grok · relex-gemini · relex-mcp (generic).
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.