Skip to content
Blog

Your agent belongs in two directories

Cosette CresslerContent & Product Marketing Lead

November 11, 20266 min read

You can reuse the same agent across your app, ChatGPT, and Claude by serving it once as a remote MCP server and listing that server in the Claude Connectors Directory and the ChatGPT plugin directory. The agent and its permissions stay in one codebase; each directory adds only an OAuth callback, tool metadata, a test pass, and a listing.

An agent that lives only inside your product reaches the people who already open your product. The same agent, published to both directories, reaches people inside the tools they already have open.

Most teams stop at "built." They ship an agent behind a chat box in their own UI and move on once a slice of users finds it. Frameworks talk about building and running agents, while directories talk about submitting them. Nobody owns the step in between, so distribution falls through the gap.

What does it mean to publish an agent?

Publishing an agent means making it callable from inside someone else's interface, with each user's own permissions, without shipping a new copy of the agent for each interface.

MCP makes this possible. Claude connectors and ChatGPT plugins both speak it. Serve your agent as a remote MCP server, and Claude and ChatGPT both call the same endpoint with the same tools, signing each user in through the same OAuth flow. The agent never learns which surface asked.

So you build the agent once and register it in two places.

How one agent serves your app, Claude, and ChatGPT

Take the invoicing agent from How to build an agent for your product and serve it everywhere. It answers questions about invoices and payment status, and it can create and send invoices too. It runs on AgentOS behind AuthKit on a public MCP endpoint.

Now follow one customer, a finance lead at a company that uses your product, through a Tuesday.

At nine she opens your app to reconcile last week's payments and asks the chat which invoices cleared. That's the surface you built, and it has the smallest audience: people who already logged in and already came looking.

At eleven she's in Claude drafting a renewal proposal for Northwind and needs a number. She asks, without leaving the draft, "what does Northwind still owe us?" Claude picks lookup_invoices, walks her through your OAuth sign-in once, and answers with her data. Claude ran the conversation while your agent looked up the balance.

At four she's in ChatGPT, where her team runs its weekly ops thread, and opens your plugin: send Northwind the overdue invoice. ChatGPT calls manage_invoices, asks her to confirm because you marked the tool destructive, and the invoice goes out. Same agent, same server, same permissions she already had.

Between nine and four, the only thing that changed was where her request came from.

What do the Claude and ChatGPT directories add to an MCP server?

Both directories add the same four things on top of the server you already run: an OAuth configuration that accepts their callback, tool metadata their reviewers can read, a test pass through their own client, and a listing with a privacy policy and a support contact. Neither touches the agent.

Tool annotations are the one place where the directories reach back into how you build. An agent published as a single plain-language tool can usually read and write, so since Agno 3.0.2, AgentOS marks it destructive by default, which means Claude asks for confirmation before every call and ChatGPT can too. The finance lead's morning is mostly lookups, and a confirmation on every lookup wears thin by lunch. Publish two tools instead of one:

mcp=MCPConfig(
    tools=[
        invoice_lookup_agent.as_tool(
            name="lookup_invoices",
            title="Look up invoices",
            description="Answers questions about invoices, balances, and payment status. Send plain language.",
            annotations={"readOnlyHint": True, "destructiveHint": False, "openWorldHint": False},
        ),
        invoice_agent.as_tool(
            name="manage_invoices",
            title="Manage invoices",
            description="Creates, updates, and sends invoices. Send plain language.",
        ),
    ],
    default_tools=False,
)

Lookups run without a prompt. Writes still ask. Both directories read the same annotations, so you make this decision once and it holds on both.

ChatGPT's review requires all three of readOnlyHint, destructiveHint, and openWorldHint on every tool, and tests each claim against what the tool really does. AgentOS states all three for anything published through as_tool, so you only override the ones your tool contradicts.

How do the Claude and ChatGPT directories differ?

The server stays the same, but the review doesn't. Claude's directory checks each submission against its requirements and lists approved connectors with a Community label, and Anthropic may escalate a listing to a hands-on Verified review. ChatGPT sends every plugin to a reviewer, who needs a verified publisher and written test cases before anything is listed.

Claude Connectors DirectoryChatGPT plugin directory
Who can submitAnyone on a paid Claude plan (on Team and Enterprise, an Owner submits)A verified individual or business publisher on the OpenAI Platform
What review checks firstThe submission requirements; Anthropic may escalate a listing to Verified review, where a reviewer tests each toolFive positive and three negative test cases, a video walkthrough of them, and a written justification for every annotation
After approvalListed automaticallyYou select Publish
Published review timelineNone; times vary with queue volumeNone

Neither directory publishes a timeline. Plan for weeks. Submit to both on the same day, and don't tie your launch date to either approval.

The path

Three guides cover the how, each built on the same invoicing agent:

If your agent isn't on MCP yet, How to build an agent for your product and serve it everywhere gets you there first.

Why publish to both directories

Every framework helps you build an agent, and most help you run one. Publishing is where the agent stops being a feature inside your product and becomes a presence inside the tools your customers open first. The directories are young and the forms look like paperwork, so most teams never file. That leaves the shelf space to the ones who do.

A month from now the invoicing agent answers questions in Claude for customers who never opened your app that day and sends invoices from ChatGPT for customers who did. The agent itself hasn't changed at all. Start with the guide for whichever directory your customers already use most.

Last updated: November 19, 2026. Both directories revise their review criteria often.

Frequently asked questions

Serve the agent as a remote MCP server with per-user OAuth, then add it to Claude as a connector and to ChatGPT as a plugin. Both read the same tools from the same endpoint; only the OAuth callback and the listing details differ. With AgentOS, MCPConfig and as_tool() expose an existing agent over MCP without changing it.

No. One MCP server serves both directories. The Claude connector guide and the ChatGPT plugin guide walk the same server through each submission.

Submit a plugin that wraps your MCP server through OpenAI's plugin portal. You need a verified publisher, a public HTTPS endpoint, OAuth that accepts ChatGPT's redirect URI, annotated tools, and five positive and three negative test cases, plus a video walkthrough of those test cases.

Submit your MCP server through Anthropic's developer portal at claude.ai/directory/manage. Anyone on a paid Claude plan can submit, and on Team and Enterprise plans an Owner submits. You need OAuth that Claude can register with, a title and annotation on every tool, public documentation, a privacy policy, and a test account with real data in it.

Both name the failed check and accept a resubmission of the same listing. How to fix a rejected Claude connector or ChatGPT plugin walks through each rejection reason and its fix.

Others also liked...