You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Breaking: tool results are now structured. Every tool declares an outputSchema and returns structuredContent alongside the JSON text
content, and list results changed shape from {count, projects|deployments}
to {pageCount, items, receipt}, where receipt reports the request's scopeKind, appliedFilters, and endpointProfile. Clients parsing the
previous text shape must update — hence 0.2.0 rather than 0.1.1.
Blank strings are no longer accepted for the optional search, projectId,
and state inputs (.min(1)). Previously a blank string was silently
treated as "filter not applied", widening the result scope.
Added
Read-only tool annotations (readOnlyHint, destructiveHint: false, idempotentHint, openWorldHint) on all four tools.
Politeness throttle on Vercel API calls: a minimum start-to-start interval
and a concurrency cap, configurable via VERCEL_MCP_MIN_INTERVAL_MS
(default 250) and VERCEL_MCP_MAX_CONCURRENT (default 4); invalid values
fall back to the defaults.
HTTP 429 responses with a numeric Retry-After of 10 seconds or less are
retried exactly once after waiting; all other 429s surface immediately.
Minimal response-shape validation on 2xx bodies before casting, so a
malformed upstream response surfaces as a shaped unexpected_response_shape error instead of an unhandled exception.
SECURITY.md with private reporting via GitHub Security Advisories.
Test suite grown from 35 to 64 tests; throttle and retry tests run on an
injected fake clock (no real-time waits).
Documentation: single-page listing scope for the list tools, throttle
configuration, contributor Node 22+ note, and a refreshed TESTING.md
that matches the current test files.