Claude Marketplace Is Live: List Your Connector or Plugin in 2026

What the Claude Marketplace actually is

The Claude Marketplace is a directory in three parts, and each part has a different door. The connectors and plugins section - "Over 2,000 connectors and plugins available", per the Marketplace site - is where your software becomes something Claude can call on a customer's behalf. The agents and products section lists finished "Legal, finance, and engineering tools powered by Claude, that can count toward your existing Anthropic commitment". The third section lists service partners who implement Claude for enterprises. This is not an app store with a billing SDK bolted on; it is a discovery surface wired into enterprise procurement.

• •Connectors and plugins "connect your apps, data and tools to Claude"; Anthropic tells builders to "create a connector or plugin using the Model Context Protocol (MCP) and Agent Skills".
• •Named among the connector and plugin partners at launch: Atlassian, Google, Microsoft, Notion and Salesforce.
• •Named among the Claude-powered agents and products: CrowdStrike, Cursor, Harvey, Legora, Lovable, Snowflake, Vercel, CodeRabbit, ThoughtSpot and Gamma.
• •Service partners come from the Claude Partner Network - Accenture, Boston Consulting Group and Deloitte are named, with Caylent and Slalom Consulting listed at Preferred tier.
• •Anthropic calls MCP and Agent Skills "open standards pioneered by Anthropic", so a connector you build on MCP is not locked to a single client application.

Who can list what: connectors, plugins, products and service partners

These four routes are not variations of one submission form. They differ in what you build, who reviews it, and whether money changes hands. A connector is a remote MCP server your customers attach to their own Claude account. A Claude Code plugin is a directory of skills, agents, hooks and MCP configuration that developers install into their editor. A Claude-powered product is your finished software, listed so enterprise buyers can set it against commitment they have already signed. Where Anthropic's public pages state no rule, the table below says so rather than guessing.

Why a Claude Marketplace listing beats another SEO page

Think about where the decision happens. A blog post competes for a click on a results page the buyer must visit first. A connector sits inside the tool list the model consults the moment someone types "pull last month's invoices" - no click, no landing page, no bounce. That is the whole distribution argument, and it explains why the launch roster reads like infrastructure companies rather than marketers. The second argument is procurement: letting a buyer set a Claude-powered product against commitment already signed removes the slowest step in an Indian enterprise sale, raising a fresh purchase order.

What changed in the 2026-07-28 MCP specification?

On 28 July 2026 the Model Context Protocol published its 2026-07-28 revision, and the MCP blog put it bluntly: "MCP is transforming from a bidirectional stateful protocol into a request/response stateless protocol." Sessions are gone, along with the initialize handshake and the Mcp-Session-Id header. Servers needing state across calls now mint explicit handles and take them back as ordinary tool arguments - which is what lets a server scale horizontally instead of pinning a user to one process. A server you shipped in 2025 still works, but it is now on a published clock.

• •server/discover is new and mandatory: servers MUST implement it to advertise supported protocol versions, capabilities and identity, and mismatches return UnsupportedProtocolVersionError.
• •The HTTP GET endpoint and resources/subscribe are replaced by subscriptions/listen, one long-lived POST-response stream clients opt into by notification type.
• •Server-initiated requests such as roots/list, sampling/createMessage and elicitation/create give way to Multi Round-Trip Requests: the server returns resultType "input_required" with inputRequests, and the client retries carrying inputResponses.
• •Tasks moved out of core into the io.modelcontextprotocol/tasks extension, replacing the blocking tasks/result with tasks/get polling and adding tasks/update.
• •Roots, Sampling and Logging are deprecated - migrate to tool parameters or resource URIs, direct LLM provider APIs, and stderr or OpenTelemetry respectively. ping and logging/setLevel are removed outright.
• •tools/list, prompts/list, resources/list and resources/read must now return ttlMs and cacheScope so clients and gateways can cache instead of polling.

Migration checklist for a 2025-era MCP server

Work through this in order. Most teams find the first three steps are the real project and the rest is mechanical. Budget a sprint if your server was genuinely stateful, far less if you already treated every tool call as independent.

Security and auth: CIMD, OAuth scopes and least privilege

The most consequential change for anyone running an authorisation server is that OAuth 2.0 Dynamic Client Registration is deprecated in favour of Client ID Metadata Documents. Under CIMD the client identifier is an HTTPS URL resolving to a JSON document: the spec requires the https scheme and a path component, such as https://example.com/client.json, and the document must carry at least client_id, client_name and redirect_uris, with client_id matching the URL exactly. Servers advertise support by publishing client_id_metadata_document_supported: true. The practical win is portability - CIMD client IDs work across authorisation servers without re-registration.

• •The spec's priority order is pre-registered credentials, then CIMD, then Dynamic Client Registration as a fallback, then prompting the user for client details.
• •Authorisation servers MUST check that a fetched document's client_id matches its URL, and MUST validate redirect URIs against the document.
• •Client credentials are bound to their issuer: key them by issuer identifier, never reuse them with a different authorisation server, and re-register when it changes.
• •Scope your connector's tools the way you would scope an API key - read-only first, writes behind explicit approval, every call logged.
• •Anthropic's caution on its own plugin directory applies to whatever you publish: it "does not control what MCP servers, files, or other software are included in plugins".

What should you build first - a connector or a plugin?

Build a read-only connector to the data your customers already log in to see. It is the shortest path to something a buyer can feel, and the version that survives a security review. Resist shipping write actions in version one. Once the read side is in real use, your logs will show which three tools people actually call - promote those into approval-gated actions. Agent Skills are the cheap second layer: a SKILL.md carries procedural knowledge your tool schemas cannot express, and Anthropic puts the standing cost of an untriggered Skill at roughly 100 tokens.

• •Weeks 1-2: a Streamable HTTP MCP server exposing three to five read-only tools over your existing API, with OAuth and per-call logging.
• •Week 3: a SKILL.md that tells Claude how your domain works - the vocabulary, the sequence, the edge cases - so the tools get used correctly.
• •Week 4: instrument it. Which tools get called, which arguments fail validation, where the model guesses wrong and needs a better description.
• •Only then: decide whether the same capability should also ship as a Claude Code plugin, which reaches engineers in their editor rather than business users in Claude.

Mistakes that sink a first connector or plugin

Four failure patterns show up repeatedly in a connector project's first month. Each is cheap to avoid and expensive to fix afterwards.

How Tech Arion helps with Claude Marketplace connectors and MCP servers

We build and maintain MCP servers and Claude connectors for product companies and IT services firms in Hyderabad, Chennai and Bengaluru. A typical engagement starts with a read-only connector over your existing API - three to five tools, OAuth, per-call logging - shipped in weeks rather than quarters, then extended with Agent Skills and approval-gated write actions once usage data shows what people actually call. If you already run a 2025-era server, we take it through the 2026-07-28 migration: stateless core, server/discover, subscriptions/listen, CIMD authorisation and the Tasks extension. That build work sits under techarion.com/services/vibe-coding, with strategy and governance under techarion.com/services/ai-consulting. We run our own AI-native delivery platform at techarion.com/services/ticket-agent, so client servers are the same shape as the ones we operate ourselves. Maintenance is part of the deal: MCP now has a published feature lifecycle with a twelve-month minimum deprecation window, and somebody has to own that calendar.

Frequently asked questions about the Claude Marketplace

The questions founders and CTOs ask us most often since the 23 September 2026 launch.