New permission-based user API keys can't run CLI/SDK agents (Service Unavailable / HTTP 503); older key works

Where does the bug appear (feature/product)?

Cursor CLI

Describe the Bug

Since switching to The X Subscription (Premium) on 2026-10-13, user API keys created
with “Run agents” and “Sign in to the Cursor CLI and SDK” granted authenticate fine,
but every local agent run fails with “NonRetriableError: Service Unavailable”.
An older key on the same account (its page shows no permissions) runs normally.

Steps to Reproduce

Steps to Reproduce

  1. In Dashboard > API Keys, create a key with “Run agents” and
    “Sign in to the Cursor CLI and SDK”.
  2. CURSOR_API_KEY= cursor-agent about
    → Subscription Tier: Premium, correct account (the key is linked correctly)
  3. CURSOR_API_KEY= cursor-agent --list-models
    → lists models normally
  4. CURSOR_API_KEY= cursor-agent -p --mode plan --model composer-2.5
    –output-format json “Reply with exactly the text KEY_OK and nothing else.”
    → NonRetriableError: Service Unavailable Service Unavailable

Expected Behavior

Step 4 returns KEY_OK, as it does with my older key and with a browser login.

Operating System

MacOS

Version Information

Current Cursor Version
cursor-agent 2026.10.01-e373342 (latest per cursor-agent about)

Additional Information

Plan: The X Subscription, Premium.

Does this stop you from using Cursor

No - Cursor works, but with this issue

Hey @Eric_III, thanks for the report and the clear repro steps!

We can see this on our side: agent runs started with the new permission-based keys are failing on our servers, while sign-in and model listing work. That matches what you’re seeing. I’ve flagged it to the team.

In the meantime, either of these should keep you running:

  1. Keep using your older API key for CLI and SDK runs, or
  2. Sign in with your browser instead of the key: run cursor-agent login, then run your command without setting CURSOR_API_KEY.

Thanks @Colin

@Colin +1 — same failure class after the dashboard/key change. I run batch orchestration via Python cursor-sdk (Client.launch_bridge + Agent.create + local send), not just ad-hoc CLI.

Symptom: CURSOR_API_KEY set to a new dashboard grok_… key → Cursor.me() passes, but local agent runs fail (we see 503 Service Unavailable on the local SDK PONG path).
Constraints: I cannot mint legacy cursor_… keys anymore — only grok_…, and I don’t have an old key to fall back to (Colin’s workaround #1 doesn’t apply).

Colin’s workaround #2: I cursor-agent login’d — CLI cursor-agent -p -f works without CURSOR_API_KEY. Python cursor-sdk does not: Agent.create still requires CURSOR_API_KEY, and with grok_ the local agent path still fails. So orchestration scripts remain broken.

Env: Linux, cursor-sdk 1.0.37, cursor-agent 2026.10.14, Ultra.

Is there a fix ETA for new/grok_ keys + local SDK, or an supported way for SDK batch to use browser/OAuth auth the way CLI does?

Hey @Eric_III and @Dumuvault_Support, quick update: the change that fixes agent runs with the new permission-based API keys (the ones starting with grok_) is now live.

You can go back to using your new key with both the CLI and the Python SDK. The command from the original report (cursor-agent -p with CURSOR_API_KEY set to the new key) should return KEY_OK, and Agent.create plus local runs in the SDK should work with the same key, so you no longer need the older key or the browser login workaround.

Switched to cli

thanks.