The Night My Agents Went Silent
On April 4, 2026, Anthropic quietly rolled out a billing change that broke third-party AI integrations everywhere. My OpenClaw gateway — the backbone of my AI agent stack — stopped responding at 10 PM with zero warning. No error. No log. Just silence.
If you're running AI agents through a Claude Max or Pro subscription, here's exactly what happened and how to fix it.
What Changed
Third-party harnesses like OpenClaw, custom Claude integrations, and any app accessing Claude via OAuth no longer bill through your subscription. Anthropic moved them to a separate system called Extra Usage — pay-as-you-go, billed independently from your plan.
The rollout was quiet. No email blast. No dashboard alert. Agents just started failing silently.
There's one piece of good news: Anthropic is offering a one-time credit equal to your monthly subscription price to help with the transition. You have until April 17 to claim it at claude.ai/settings/usage.
Step 1: Activate Extra Usage
This is step zero. Without it, every agent request fails silently.
Go to claude.ai/settings/usage and enable Extra Usage. The credit will be applied automatically if you activate before April 17.
Step 2: Migrate to the Claude CLI Backend
If you're running OpenClaw or a similar harness with a static OAuth token, it will time out after the billing change. The fix is migrating to the CLI backend:
openclaw models auth login --provider anthropic --method cli --set-default
This switches your provider from static OAuth to the Claude CLI as the execution layer.
Step 3: Configure cliBackends Manually
This is the part the documentation doesn't tell you.
The migration command updates your model paths but does not configure the backend. You need to manually add this to your openclaw.json under agents.defaults:
cliBackends: {
claude-cli: {
command: /usr/bin/claude,
args: [-p, --output-format, json,
--permission-mode, bypassPermissions],
output: jsonl,
input: stdin,
sessionMode: none
}
}
Without this block, OpenClaw has no idea how to invoke the CLI — and agents stay silent even after the migration command succeeds.
The Gotcha: Avoid These Flags
Two flags will silently break your output:
--verbose— causes raw debug JSON to leak into your chat channels instead of clean text--output-format stream-json— same problem, JSON fragments appear as chat messages
Stick with --output-format json. It works reliably and returns clean, parseable output.
Summary
| Step | Action |
|---|---|
| 0 | Claim the credit at claude.ai/settings/usage before April 17 |
| 1 | Enable Extra Usage in your Claude account settings |
| 2 | Run openclaw models auth login --provider anthropic --method cli --set-default |
| 3 | Add the cliBackends block manually to openclaw.json |
I spent 3 hours debugging this so you don't have to. If your agents went quiet this week, start with step 0 — Extra Usage — and work down the list.
Questions or running a different harness? Drop a comment below.