Loading
0xB0Lesson 12 of 16

Keep state with handles, and listen for changes

Carry state across calls with explicit handles, and get notified when lists or resources change.

20 min 5-question quiz 1 code exercise
By the end of this lesson you can
  • Design a tool that returns and accepts a state handle
  • Bind handles to the authenticated user
  • Use subscriptions/listen for change notifications

MCP has no protocol session, so a server can’t tie two tool calls together by connection. When it needs state - a shopping basket, an open browser tab, a database transaction - the spec’s guidance is an explicit handle:

  1. a creation tool returns a handle, e.g. {"basket_id": "bsk_a1b2c3"},
  2. the model passes it as an argument to later calls (add_item(basket_id, sku)),
  3. the server looks up the state under that key.
handles.py
1import secrets
2baskets = {}
3
4def create_basket(user):
5    basket_id = "bsk_" + secrets.token_hex(8)
6    baskets[(user, basket_id)] = []
7    return basket_id
8
9def add_item(user, basket_id, sku):
10    if (user, basket_id) not in baskets:
11        return "error: unknown or expired basket"
12    baskets[(user, basket_id)].append(sku)
13    return f"{len(baskets[(user, basket_id)])} item(s)"
14
15basket = create_basket("ana")
16print(add_item("ana", basket, "book"), "|", add_item("mallory", basket, "tv"))
Output
1 item(s) | error: unknown or expired basket

Good handles are:

  • unguessable - random (like a UUIDv4), not basket-1, basket-2;
  • bound to the user - stored as user:handle with the user taken from the verified token, so a leaked handle is useless to anyone else (a handle is a name, not a password);
  • limited in lifetime, with the policy stated in the tool description (“baskets expire after 24 hours”);
  • clear when expired - return a tool execution error saying so, so the model can create a new one.

Subscriptions and long-running work

  • Change notifications. A client opens a long-lived subscriptions/listen request, choosing what to hear about (tool, prompt or resource list changes, or updates to specific resource URIs). The response is a stream that stays open; notifications carry a subscriptionId in _meta.
  • Progress. A request can include a progressToken in _meta to receive notifications/progress while it runs.
  • Cancellation. On Streamable HTTP, closing a request’s response stream cancels it; on stdio, the client sends notifications/cancelled.
  • Tasks for work that takes minutes or hours now live in the official io.modelcontextprotocol/tasks extension, with tasks/get for polling.

Key takeaways

  • No sessions: carry state with explicit handles passed as tool arguments.

  • Make handles random, user-bound and expiring - never a substitute for authorization.

  • subscriptions/listen streams change notifications; progressToken streams progress; Tasks handle long-running work.

Lesson quiz

5 questions · pass with 4 correct · up to 50 XP

Passing this quiz completes the lesson and keeps your streak going. Questions you miss come back in review sessions later.

Practice: build an MCP server

Write Python in the editor and run it against sample inputs. Python runs locally in your browser using a WebAssembly runtime.

Exercise 1

Implement user-bound handles

+25 XP

Process commands until the input ends. Each line is USER create HANDLE, USER add HANDLE ITEM or USER show HANDLE. State is stored per (user, handle).

  • create: start an empty basket and print created HANDLE.
  • add: if the user owns that basket, add the item and print added ITEM; otherwise print error: unknown basket.
  • show: print the items joined by commas (or empty), or error: unknown basket.
  • Owner and intruder
  • Empty and unknown
main.py
Loading editor…

Python runs in a sandboxed browser worker with a 60 second time limit. Its runtime loads from the Pyodide CDN; your code stays in this browser.

Questions about this lesson

Stuck? Ask. Figured something out? Share it. Explaining is one of the best ways to learn.

Loading posts…