AI Engineer World's Fair 2025

Agents Access and the Future of Machine Identity

Loading slides…

Read the talk

Agents, Access, and the Future of Machine Identity

A shirt-ordering MCP server shows how OAuth, persistent storage, and mutable authorization state let an agent act for a user while keeping the server in control.

From a talk by Nick Nisi and Lizzie Siegle

Before you start: Basic familiarity with APIs and tool-calling assistants is helpful; OAuth, JWTs, and the storage roles are explained as they appear.

What should an agent be allowed to do?

Lizzie Siegle built an agent to vote for her in the NBA finals. Eventually, it got blocked. Booking scarce tennis courts in San Francisco offers another attractive use for automation: let an agent handle the tedious work of getting something you want. Both examples lead to the same question: how do you control what an agent is allowed to do?

GitHub agenda titled “Agents, Access & The Future of Machine Identity,” with sections on the opening problem, WorkOS, and infrastructure.Open full source frame
The opening agenda asks how to control what agents are allowed to do.

For Nick Nisi, the appeal is managing GitHub: an MCP server can help an assistant read reviews, inspect diffs, and work with repositories. But his setup requires editing a JSON file and supplying a personal access token, or PAT. That is substantial friction for someone who simply wants a tool to act on their behalf. His experience describes the setup used at the time; GitHub’s current official MCP server also documents remote OAuth, with support depending on the host.

The starting point is familiar delegated authorization. WorkOS supplies authorization and user management for applications; agents need those foundations too, before permissions become more granular. OAuth lets a user authorize an agent to act on their behalf without turning credential setup into the user’s job. The agent becomes another participant in the application’s access model.

0:320:49
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:32 · section reference included

Give the agent compute, memory, and identity

An agent also needs somewhere to run and somewhere to remember things. Cloudflare Workers provides compute at the edge. Alongside it are hosted AI inference, the Vectorize vector database, a SQL database, and Durable Objects, which supply persistent memory in Cloudflare’s agents framework.

Workers bindings connect application code—including agents and MCP servers—to Cloudflare services. The wider platform also includes video streaming and image optimization, and applications can still call other companies’ products. Siegle points to the newly available Durable Objects free tier and to revenue-generating startups that pay Cloudflare nothing. That free tier covers SQLite-backed Durable Objects within usage limits, rather than unlimited hosting.

These pieces serve different needs: Workers executes the code, Durable Objects retains state, and the framework’s OAuth support establishes whom the worker or agent represents. Persistent memory and delegated identity can then participate in the same request.

GitHub agenda showing Infrastructure Foundation notes, including persistent stateful infrastructure, Durable Objects, AI bindings, and deployment steps.Open full source frame
Infrastructure notes describe persistent agents, Durable Objects, and bindings.
2:262:40
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

2:26 · section reference included

Deploy the shop and connect Claude

The companion agenda is a Markdown file in the repository. The runnable MCP Shop on Cloudflare example combines Cloudflare infrastructure with WorkOS authentication. Nisi deploys it with the project’s deployment script, which invokes Wrangler:

bash

npm run deploy

The deployed worker appears in Cloudflare’s compute dashboard. Local development is also supported, although the live explanation does not settle on a precise local command.

The connection sequence replaces manual token handling with a sign-in flow:

  1. Copy the Workers URL returned by the deployment command.
  2. Open Claude and choose Add Integration.
  3. Name the integration—the demonstration uses mcp.shop and mcp-shop—and append /mcp to the Workers URL.
  4. Select Connect, then sign in with GitHub.

After sign-in, Claude refreshes with the integration connected. The server now has a user on whose behalf the agent can act. These are the interface labels and endpoint used in the recording.

4:204:24
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:20 · section reference included

Turn a shirt request into a stored order

Nisi asks for a shirt in ordinary language. Claude recognizes the shop’s available tools and proposes listing inventory; Nisi permits the tool to run. The result is an MCP shirt offered for free in the demonstration, with three required inputs: size, company name, and mailing address.

He supplies size XL and a mailing address in Bellevue, Nebraska. Claude notices that the company is missing and asks for it. Nisi replies WorkOS duh, introducing a small but revealing ambiguity: is the final word part of the company name, or conversational emphasis?

The order completes and returns an order number. Nisi then asks for the order information to check how the company was recorded. Siegle calls this “Human in the loop”; Nisi reports that Claude interpreted the company correctly. The follow-up is useful because completing a tool call and capturing the intended field values are separate things.

Next, he opens Cloudflare’s dashboard, navigates to storage KV, and selects orders. The saved order data is visible there. The conversation has produced application state, not just a response describing an order. Workers KV provides the key-value storage for that record; the demonstration establishes persistence, not shipment fulfillment.

6:076:13
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:07 · section reference included

Ask what the authenticated agent knows

Nisi next asks what the assistant knows about him. A user-information tool returns mostly information from his JSON Web Token, or JWT. The response includes his name and email address, plus a custom favorite-song value: Careless Whisper. It also reports admin access from his roles and permissions. Authentication therefore supplies more than a successful login: the application can expose identity attributes and permission information to its tools.

8:468:53
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:46 · section reference included

Change stored state, change the authorization result

This MCP server also uses a Durable Object, allowing it to store data directly with the running object. Siegle describes objects provisioned per user and proximity as a reason for fast storage and retrieval. Placement is more specific than that shorthand suggests: current documentation describes placement near the initial request, not continuous relocation to follow a user. The demo does not measure latency.

Nisi asks the agent to change the demo mode to banned. A server tool updates that mode in the context associated with the object. He then requests another shirt and selects always allow for the tool call in Claude. The request is still rejected. Permission to invoke a tool in the client does not override the server’s authorization decision.

The server can combine the authenticated identity with mutable state stored in the Durable Object. Nisi describes that state as unique to each user; the companion agenda describes an object per session, so the demonstrated behavior should not be read as a guarantee of account-wide state across sessions.

LayerWhat it contributes
JWT informationIdentity, roles, and permissions
Durable Object stateMutable demo mode, including banned
Client tool approvalPermission for Claude to attempt the call

Together, these explain why an authenticated user can approve an attempted call while the server still refuses the requested operation.

Nisi then asks nicely. The demo has a Pretty Please tool available, and after it runs he reports that another shirt has been ordered. This is an intentionally playful state-changing capability: the server exposes a tool that can change the outcome. Polite wording is not itself an OAuth permission or a general way to bypass authorization.

9:329:48
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:32 · section reference included

Start with tools, then protect them

Cloudflare’s click-to-deploy route provides a short path to a custom MCP server: generate a GitHub repository, clone it, and edit the supplied starter tools. The starter Siegle describes is authless, which is why she explicitly recommends adding authorization. An authenticated deployment remains a quick setup; its extra work establishes who can use the capabilities being exposed.

11:3111:48
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

11:31 · section reference included

Make delegated actions accountable

Nisi characterizes the foundation as adding an OAuth flow to MCP, which is effectively an API. The familiar controls used for human-facing applications are a starting point for tools acting on a person’s behalf. From there, he sees authorization becoming more specific:

  • Changes: permission for individual line changes.
  • Tools: permission for particular tool changes or operations.
  • Connections: permission governing network connections between components.

These are directions for finer control, beyond what the shopping demonstration implements.

As agents potentially perform thousands of tasks for a person, authorization also needs an audit trail. Nisi’s questions are concrete: why did an interaction happen, on whose behalf did it happen, what was its result, and where did it fail? The scale is a forecast, while the questions describe what an operator would need to reconstruct an agent’s behavior.

Siegle frames the actors as deputies with access to tools. That access enables both use and misuse; her experience inviting live-demo input is a reminder that people do not always submit friendly requests. A delegated tool needs boundaries even when the person directing it has successfully signed in.

GitHub agenda with Authority Controls and Takeaways sections, including “Think deputies, not users” and questions about controlling and authorizing agent actions.Open full source frame
Authority controls and takeaways emphasize deputies and action authorization.
12:0912:20
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:09 · section reference included

Use the workflow or customize the server

The closing screen provides two routes: the top QR code offers the shirt workflow, and the second points to the code. Nisi names mcp.shop as the public entry point for the same ordering flow, but distinguishes it from the live demo: the public version does not include the Pretty Please tool.

Two QR codes in a GitHub document under “Get a shirt with MCP!” and “Check out the code,” with presenters inset at lower left.Open full source frame
QR codes offer routes to get a shirt with MCP and check out the code.

At the time of the recording, the public offer was a zero-dollar MCP shirt. Today, that address redirects to the WorkOS Shop; the storefront does not establish whether a separate free MCP promotion remains available. The code is the route for adding a custom Pretty Please tool—or a more useful application-specific capability—with the closing invitation focused on building MCP servers with authorization.

13:2813:40
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

13:28 · section reference included

Resources

From the talk

Updates since the talk

  • Current documentation for object placement, location hints, jurisdictions and latency considerations.

  • GitHub's official server for repository and pull-request workflows, with current remote OAuth and PAT setup instructions.

Transcript download
Full transcript

Loading transcript…