CloudZero Desktop Agent 0.3.7

Release Notes

Release 0.3.7 addresses a security gap, adds ChatGPT as a sign-in provider alongside Claude, and improves the accuracy of recorded usage and cost.

The local debug interface no longer accepts requests originated by a web page: CloudZero's monitor exposes a debug interface, reachable only from the machine itself, that can switch on a capture debugger. It previously accepted cross-origin browser requests without authentication. Those requests are now rejected, and the interface no longer presents itself to web pages as reachable. No user or administrator action is required, and no configuration changes.

Sign in to ChatGPT: CloudZero can now connect to a ChatGPT account in addition to Claude. Setup now begins by asking which providers you use, and shows the steps that apply to your choice. You can also connect ChatGPT later from Settings.
ChatGPT credit balance and usage windows: Once connected, CloudZero shows your ChatGPT credit balance and current usage windows next to the existing Claude spend card, with a warning treatment as you approach a limit. Seats on a plan with no spending limit are shown as unlimited rather than reported as having no access.
Spend cards say which provider they describe: Every spend and usage card now carries a provider badge. Previously a card could read "Monthly Spend" with nothing on it identifying whose spend it was. Cards for Anthropic are labelled by the product used.

CloudZero recovers when an upgrade leaves the previous network extension running: An upgrade could replace the application while the network extension from the previous version kept running indefinitely, with nothing detecting the split — so a machine could carry old capture code for days while reporting itself healthy. CloudZero now compares the running extension against the one it ships with and re-requests activation when they differ.
Streamed responses are recorded with their full token counts: A streamed Anthropic response that arrived without network pacing could be recorded before its closing frames, billing its opening token tally instead of the final one — in one observed case 8 tokens instead of 215. Recording now waits for the response delivered to your application to finish.
Streamed output is no longer delayed for unusual response types: When a provider labelled a streamed response with an unexpected media type, output could be held back rather than passed through as it arrived. Streamed model output now reaches your application as it arrives in that case too.

Calls that could not be priced are distinguished from calls that were genuinely free: Cost rollups previously counted an unavailable cost estimate as zero, which understated spend without saying so. A reported cost now always means complete billable usage at a known rate — including a real zero — and a call the collector could not price is reported as unpriced rather than free.
Consistent cost estimation across providers: Anthropic, Vertex, Bedrock-shaped, and Codex traffic now share one cost estimator, so pricing behaviour cannot drift between them. Failed, incomplete, and unrecognized responses remain unpriced.
Response detail is retained across more provider formats: Request identifiers, organization and rate-limit facts, usage, and completion state — including partial and capped responses — are now carried through a shared path rather than a provider-specific one, so newer provider formats publish the same detail as established ones.