Skip to main content

Recap: First five weeks of the Agent Readiness Rampage

Category
Company
Published

What we've shipped to improve the agent experience for you and your users.

For the last five weeks, Clerk's been focused on shipping updates for agents in an initiative we've been calling the Agent Readiness Rampage. New updates have landed on X twice a week as we build in public. Here's a summary of what we've shipped and why.

Your coding agent and Clerk

Each time Clerk needed a person to click, copy, or paste, your coding agent stopped and waited for you. The Wednesday updates went after those moments to reduce friction, so the agent can set up, run, and ship Clerk with fewer handoffs.

An MCP server for your account

The first update rebuilt Clerk's Model Context Protocol (MCP) server around your Clerk account. It used to only serve SDK snippets. Now you sign in once from your MCP client and approve what the agent may do. The agent can then create applications, change instance settings, manage users, check domains, and create production instances.

Each call that reads or changes your apps gets a token with only that tool's permissions, and the token expires within 5 minutes. No secret key goes into your MCP client config. The launch post covers sign-in and permissions in detail.

Before this, we pointed agents at the Clerk command-line interface (CLI). The CLI still works, and clerk mcp install connects your AI clients to the server. But an MCP server also reaches agents that never open a terminal, such as ChatGPT and Grok Bot.

Install the Clerk plugin to add the MCP server and Clerk's skills to your agent in one step:

npx plugins add clerk/skills

More of the Dashboard as tools

At launch, the MCP server focused on creating and deploying apps. Since then, we have added a variety of tools for you and your team to use. Your agent can find users by last sign-in, creation date, account status, or Organization, and update their details and metadata. It can list and search Organizations and their members, rename an application, and allowlist redirect URLs for single sign-on (SSO) in mobile and desktop apps. If you connected before November 12, sign out and reconnect to use the Organization tools.

Figure out why ada@example.com can't sign in.

Compare my dev and prod configs.

The MCP page has more prompt examples like these to explore and paste into your agent. The next tools come from what agents look for and can't find. The first set was user and Organization lookups.

The MCP server still doesn't cover everything the Dashboard does, but it's getting close, and we're working on this. For the gaps, Clerk's MCP server can open the Dashboard for your agent in a browser session. That session is read-only, so the agent can look around but can't change anything.

Plugins where you already work

The Clerk plugin for Codex gives it Clerk-specific guidance for adding auth, building custom sign-in flows, and setting up Organizations and Billing. It's listed in OpenAI's plugin directory, which also serves ChatGPT. Grok Bot can look up your users and Organizations through Clerk's plugin, and the Grok Build coding agent has a Clerk plugin too.

Clerk's MCP server can also send skills to your agent, so it learns how to work with Clerk without a separate install. Codex has early support for this today.

In Codex, the server also ships an MCP App. Ask about a user, and their details show up as a card in the chat instead of a wall of JSON.

Show me the user ada@example.com.

A site agents can read

Agents read clerk.com too, and now it's easier for them to find their way around. A missing page returns markdown when an agent asks for it. The homepage points agents to Clerk's API catalog and llms.txt with a Link header.

Clerk's MCP server card follows the format proposed for MCP, and the server's handshake sends a title, description, icon, and instructions. There's also a Personal Agent Protocol manifest at /.well-known/poppy.json. Clerk's docs register a WebMCP search tool, so browser agents can search the docs without clicking through pages.

Production without the friction

Going live used to be where agents handed the work back to you. Two updates changed that.

First, the CLI's deploy command now works for agents. In agent mode, it never prompts, and with --domain, it creates the production instance. Each run prints a JSON report with the next step, such as the Domain Name System (DNS) records to add. clerk deploy status --wait then rechecks whether DNS, SSL, and email DNS are verified.

npx clerk@latest deploy --mode agent --domain <your-domain>

Second, the MCP server can now give your agent a secret key without exposing it. When your agent creates an application or a production instance, it can send a public key, and the secret key comes back encrypted to it. Only your agent's private key can decrypt it, so the key never travels in plaintext through the MCP connection.

That's a straight shot from prompt to deployed app, with no keys to copy from the Dashboard. For an existing app, clerk env pull is still how you get its keys.

Your users' agents and your app

Your users are starting to send agents to your app. Some connect through an MCP server you run, from clients like Claude, ChatGPT, and Cursor. Others are personal agents that browse your site.

Each one needs to sign in as the right person with the right permissions, and you need to know what it ran into. The Friday X updates we've been sharing built those pieces.

A Dashboard home for agents

The first Friday update added an Agents section to the Clerk Dashboard. On its MCP servers page, you register your MCP server's URL, and access tokens Clerk issues for it work with MCP clients such as Claude Code. Once it's registered, the page gives you a prompt for your agent to add Clerk auth to the server.

A week later, the page started running configuration checks. Clerk tests your server's public discovery against the MCP authorization spec and shows what's missing. Since early November, development instances also accept localhost URLs, because that's where most servers start.

MCP auth your agent sets up

Setting up MCP auth means getting discovery documents, scopes, and token checks right. Your agent can do that work with Clerk.

Install the Clerk plugin and ask it to set up auth for your MCP server. The plugin walks it through the setup, including the /.well-known files that clients use for discovery. Clerk's MCP server can then register your server's URL and run the same checks as the Dashboard.

Set up auth for my MCP server with Clerk.

Your server can also use step-up authorization. Clients can connect with a small set of scopes. Your server asks for more only when the agent calls a tool that needs them, and the user keeps the access they already approved. The requireAdditionalScopes() helper in @clerk/mcp-tools builds the challenge for the MCP SDK. Step-up needs @clerk/mcp-tools 0.9.0 or later and @modelcontextprotocol/server 2.1 or later. Not every client supports step-up yet, and the step-up guide explains how they differ.

Feedback straight from agents

When an agent gets stuck in your product, it usually knows what would have helped. Until October 23, you had no built-in way to hear it.

Agent Feedback fixes that. Add the submit_feedback tool from @clerk/mcp-tools to your MCP server, and agents report when they're blocked, need a workaround, or find a tool confusing. Clerk stores the reports on your instance, and your own agent can read them through Clerk's MCP server.

Look at the blocked reports and come up with new tools to add.

A week later, reports got their own Feedback page under Agents in the Dashboard. Reports can also record the user the agent was acting for and your MCP server's version, so you can tell which release they came from. You get the user when your server verifies tokens with verifyClerkToken(), and the version when you pass it to submitFeedbackTool().

Clerk's own MCP server runs the same tool, and its reports land in our Dashboard the way yours land in yours. The agent feedback guide shows how to add it.

Signed-in browsers for agents

Agents that use a browser for your users reach your sign-in page and stop, or ask for a password. Since November 6, your MCP server can give them a one-time link instead. The create_browser_session tool from @clerk/mcp-tools returns a link that opens your app already signed in as the user who owns the MCP access token.

The user comes from the verified token, never from the tool's input. The agent picks only the page to open and a short task description. By default, Clerk's Frontend API refuses Organization switches and profile changes from that session.

The tool is in beta, and it needs its own scope that users approve for each client. See Sign agents in to your app with a one-time link to set it up.

Agent Auth for personal agents

Personal agents are a new use case. They're starting to adopt the Personal Agent Protocol (Poppy), an open protocol for how they discover apps, prove who they are, and report what blocked them.

Clerk for Agent Auth, built on OAuth and Poppy, entered private beta with a small group of partners in late October. Agents that speak the protocol can discover your app, Clerk verifies their identity and signs them in, and they can send you feedback when they hit a wall.

On November 13, the first piece opened to every app: Poppy's browser session endpoint. A person's agent can now open your app in a browser that's already signed in as them. Turn it on from the Personal Agent Protocol page under Agents in the Dashboard, then serve a poppy.json file so agents can find it. With @clerk/nextjs 7.10.0 or later, clerkMiddleware({ poppy: true }) serves the file for you.

Before you turn it on, restrict what agent sessions can do in your app. The Personal Agent Protocol guide covers setup, and Clerk will support more of the protocol over time.

Analytics for agent activity

Feedback shows where agents get stuck. Analytics shows how many show up.

In November, the Dashboard Overview got an Agents view. It shows active agents, split into new, retained, and reactivated, and the clients they use most, such as Claude, ChatGPT, or Cursor. Agent share shows how much of your active usage comes from agents.

Rampage on

Five weeks in, we're excited about the progress made so far and feedback received. If your agent runs into something Clerk can't do yet, have it tell us. Clerk's MCP server has a submit_feedback tool, and the reports go straight to our team. You can also find us in Discord.

As we continue to build in public, we'll keep sharing updates on X and in the changelog. If you want to learn more about how Clerk works with agents, check out the AI overview guide in our docs.

Ready to get started?

Start building
Author
Drew Sing

Share this article

Share to socials: