Base44
Build and edit Base44 apps from an AI assistant, and query data your app's users submit.
Give a coding agent Clerk's own auth code — hooks, components, server-side patterns and seven framework quickstarts.
Opens Agent Studio, where connecting is one click. The connector URL below works in any MCP client.
Paste it into any MCP client. Setup docs
The Clerk MCP server hands a coding agent Clerk's own authentication code — hooks, components, server-side patterns and seven framework quickstarts — so it writes Clerk integrations from current examples rather than training-cutoff memory. It exposes 2 tools and a large documentation library addressed by URI. No account or key is required.
Verified connector
Listed by Anthropic as a partner connector in its Connectors Directory.
Connection checked by Agentman on .
Anthropic states this reflects the level of review a connector received, not a security audit.
Get Clerk SDK code snippets and patterns.
List all available Clerk SDK snippets and bundles. Filter by tag to find specific functionality.
Read from the server on 2026-08-18, including each tool's own safety annotations.
annotations object is absent on both, so there is no machine-readable safety claim to rely on in either direction. The bound on this surface comes from the schemas, not from a vendor assertion.clerk.com/docs.clerk_sdk_snippet rejects quickstart slugs. In a client that does not surface MCP resources, those seven documents are unreachable.clerk_sdk_snippet does not search. An unrecognised slug returns an "unknown snippet" message, not a nearest match. Call list_clerk_sdk_snippets first. And omitting slug entirely yields the b2b-saas default — roughly 27,000 characters on 2026-08-18, a real context cost for an agent that meant to ask something narrow.initialize, so there are no server-supplied templates to invoke.initialize, tools/list, resources/list, resources/templates/list, prompts/list, two resources/read calls and four tools/call invocations against the two catalogue tools. Every claim above comes from those responses, from the schemas, or from Clerk's own published pages.No. Neither live tool takes an instance identifier, an API key or a user identifier. Their only parameters are a snippet slug and a tag, and every resource is a static markdown document under the clerk:// scheme. We confirmed it on 2026-08-18 with an anonymous handshake that returned the full surface without a credential. This connector ships documentation, not account access.
No. Anthropic lists the connector as authless, and we verified it on 2026-08-18: an anonymous handshake returned both tool schemas and all resource entries, and a zero-argument call to list_clerk_sdk_snippets executed and returned the catalogue with no credential and no authentication challenge. The server does publish an OAuth descriptor, but nothing on the surface we exercised required it.
A tool is a call the model makes mid-conversation; a resource is context the client fetches by URI and holds. Clerk's two tools search and return snippets on demand. Its resources are markdown documents at addresses like clerk://sdk/use-auth that a client can load directly. Resources carry no safety annotations because nothing invokes them, so calling them read-only would be a category error.
Because the value is in the catalogue behind them, not in the tool count. The pair is a list-then-fetch design: list_clerk_sdk_snippets enumerates 17 snippets, 6 bundles and 30 tags, and clerk_sdk_snippet retrieves one by slug. Anthropic's directory lists exactly those two names and the live server returns exactly those two, so there is no drift on the tool surface.
Seven have dedicated quickstart resources: Next.js 15 App Router, React with Vite, Expo, Astro, Remix, Express.js and Go. The snippet documents are broader, since the hook examples name React, Next.js, Expo, React Router, Chrome Extension and TanStack Start as supported imports. Anything outside those seven falls back to Clerk's own documentation site rather than the connector.
Neither tool carries any annotation at all. The annotations object is absent on both, so Clerk has made no machine-readable safety claim in either direction, and silence is not a guarantee. What we can say comes from the schemas: the only inputs are a snippet slug and a tag string, and the observed outputs are markdown documents, so there is no parameter through which either tool could reach your data.
No. Clerk's documentation opens with a beta callout stating the feature is in beta and that functionality may change before general availability, and the January 2026 changelog announcing it says public beta. Treat the tool names, snippet slugs and resource URIs on this page as a snapshot of 2026-08-18 rather than a stable contract, and re-probe before depending on any identifier.
.md, an affordance Clerk's robots.txt advertises. Source for the beta callout, the clerk mcp install route and its ten supported clients, the per-client manual setup, the two-row tool table, the example prompts, the six snippet bundles and all three FAQ answers including the SSE statement · retrieved 2026-08-18initialize against https://mcp.clerk.com/mcp (2026-08-18, HTTP 200) — protocol 2025-06-18, serverInfo name "Clerk MCP Server" version 0.0.0, declaring tools and resources capabilities and no prompts capability. Answered an anonymous request with no credential and no challenge; this is what sets authVerified · retrieved 2026-08-18tools/list against the same endpoint (2026-08-18) — 2 tools with titles, descriptions and input schemas, and no annotations object on either. Source for every parameter, default and annotation-absence claim above. Schemas committed at docs/marketing/data/tool-schemas/clerk.tools.json · retrieved 2026-08-18resources/list and resources/templates/list against the same endpoint (2026-08-18) — 48 entries resolving to 24 unique clerk:// URIs, all text/markdown, plus 2 URI templates. Source for the quickstart/sdk split, the per-prefix counts and the duplication finding · retrieved 2026-08-18resources/read of clerk://sdk/use-auth and clerk://quickstarts/nextjs-app-router (2026-08-18) — the document sizes, framework import comments and middleware/layout content above · retrieved 2026-08-18tools/call against list_clerk_sdk_snippets (no arguments, and with tag set to billing and organizations) and clerk_sdk_snippet (no arguments, and with slug set to use-user, organizations and nextjs-app-router), 2026-08-18 — source for the 17-snippet / 6-bundle / 30-tag catalogue, the bundle sizes, the tag-filter behaviour and the unknown-slug rejection. Every call was a catalogue read returning markdown; none created, changed or deleted anything · retrieved 2026-08-18prompts/list against the same endpoint (2026-08-18) — JSON-RPC error -32601 Method not found, consistent with the absent prompts capability · retrieved 2026-08-18https://mcp.clerk.com/.well-known/oauth-protected-resource (2026-08-18, HTTP 200) — an RFC 9728 descriptor naming https://clerk.clerk.com as the authorization server, with PKCE S256 and OAuth scopes. The server publishes an OAuth descriptor while answering anonymously, which is unusual and worth recording. We observed no path on which it was required: discovery and execution both succeeded unauthenticated. https://mcp.clerk.com/robots.txt returns 404 · retrieved 2026-08-18mcp-remote, the converging re-install and clerk doctor · retrieved 2026-01-20https://clerk.com/robots.txt (retrieved 2026-08-18, HTTP 200) — Allow: / for the general agent with a short disallow list that does not include /docs/guides/, and Content-Signal: search=yes, ai-input=yes, ai-train=yes. ai-input=yes is an explicit opt-in to AI synthesis of this content, so the prose above is synthesised from Clerk's pages with their stated permission. The file also blocks four named crawlers outright and advertises /llms.txt, /llms-full.txt and per-page .md suffixes. https://clerk.com/llms.txt (HTTP 200) is the machine index that located the changelog digests · retrieved 2026-08-18.md suffix returns 404 here, so the policy was read from HTML · retrieved 2026-08-18Connect Clerk once and your agents call these tools on their own: on a schedule, in a workflow, with nobody at the keyboard.