Skip to main content
Before deploying, check your runtime, credentials, spending settings, and error handling.

​
Deployment checklist

1

Choose a supported runtime

Use ESM on Node.js 22.18 or later. See compatibility for verified Node and framework versions. The Claude Agent SDK also needs its supported process environment.
2

Store credentials on the server

Set AGENTMUXER_API_KEY in your deployment’s secret store, or pass the key through agentmuxer({ apiKey }). Keep your model-provider key separate. Never send an application key to a browser, mobile bundle, or model prompt.
3

Set spending and approval policy

Applications default to no spending limit and can use their owner’s available balance. Set a lifetime application cap when you need a hard cumulative limit. Configure payment confirmation for an interactive or unattended workflow. See billing and approvals.
4

Record useful run information

Use your agent framework’s tracing to connect a user request to model and tool activity. Retain request and invocation identifiers from AgentMuxer responses for investigation. Keep API keys and authorization headers out of logs and traces.

​
Connection lifecycle

Keep each Vercel connection alive for the full run that uses it. If you share a connection, closing it while another run is active interrupts that run. For HTTP streaming, cleanup must run after completion, failure, or cancellation of the stream.

​
Timeouts, cancellations, and retries

Set request and run limits using your agent framework. A stopped run or dropped connection doesn’t mean a purchase failed. AgentMuxer may already have accepted it.
Do not retry a purchase with a new idempotency key after a timeout. Preserve the original key when replaying the same purchase request. A new key represents a new purchase.
The Vercel helper disables automatic MCP tool-call retries. When retrying a paid call, keep the original arguments and idempotency key. If you’re unsure whether it completed, save the invocation ID and check the existing purchase.

​
Handle tool responses

Keep cost, invocation ID, and any refusal or confirmation details when processing tool results.

​
Logging and support

The SDK works with your framework’s logging and tracing; you do not need an AgentMuxer-specific tracing setup. Capture the SDK and framework versions, your own run identifier, and request/invocation identifiers returned by tools. Avoid logging entire connection options because they can contain your key. For help, send a minimal reproduction and those identifiers to founders@amorphiclabs.com. Remove secrets and sensitive tool inputs first.

Applications and keys

Rotate credentials and manage application access.

Compatibility and upgrades

Check supported versions before an upgrade.