On 24 October 2026, I cannot use Antigravity in either the Gemini Mac app or the standalone Antigravity app. Google browser sign-in completes, but standalone account setup fails. In Gemini’s Antigravity mode, no models load.
Environment
macOS 26.7.1, build 25G241.
Standalone Antigravity 2.26.0.
Gemini Mac app 1.135.1.1291, with bundled Antigravity.
Error (local IP address and ephemeral port omitted):
Post "https://daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist": read tcp [LOCAL_ADDRESS]->172.217.118.4:443: read: socket is not connected
Steps to reproduce
Launch Antigravity 2.26.0 and sign in with Google.
Browser authentication succeeds, but the app shows “There was an unexpected issue setting up your account” with the error above.
Open Antigravity mode in Gemini Mac: “Error Loading Models” with the same loadCodeAssist failure.
Repeat with VPN disconnected and direct Wi-Fi: same failure. The direct Wi-Fi route was verified.
curl to https://daily-cloudcode-pa.googleapis.com/ and /v1internal:loadCodeAssist fails with exit 35: “Recv failure: Socket is not connected”; HTTP 000.
Explicit TLS 1.2 and HTTP/1.1 checks fail the same way.
Python’s certificate-validating TLS client times out during the handshake to this hostname.
https://cloudcode-pa.googleapis.com/v1internal:loadCodeAssist and https://oauth2.googleapis.com/token complete TLS verification and return HTTP 404 to unauthenticated GET, as expected for this test.
https://antigravity.google/ returns HTTP 200.
A same-IP TLS comparison succeeds for cloudcode-pa.googleapis.com and fails for daily-cloudcode-pa.googleapis.com.
Local DNS agrees with Google Public DNS for both Code Assist hostnames.
Managed endpoint network filters remain installed, including Microsoft Defender and GlobalProtect. Their presence is not proof of causation; no per-connection Defender block verdict was established. System HTTP/HTTPS proxy settings are empty. Root cause remains unconfirmed: hostname-specific network filtering or a Google endpoint/TLS issue may be involved.
Expected: successful sign-in, account setup and model loading.
Actual: Antigravity is unusable in both desktop surfaces.
Please investigate the daily-cloudcode-pa endpoint/TLS path, confirm the required domain allowlist, and provide a supported recovery path. No app binaries, authentication caches or security protections were modified. No credentials, tokens, raw diagnostic bundles or private project data are included in this report.
Thank you for reporting this issue and sharing your detailed diagnostic logs. Please check whether any custom environment variables or endpoint overrides are configured in your shell profile. We can confirm this issue is already being investigated by our engineering team to understand why this build attempted to route traffic to a staging host.
Update on 28 October after checking the edited reply and testing the latest clients:
No endpoint override was configured in my shell profiles, environment, launchctl environment or Antigravity config before these tests.
Updated the official clients to Antigravity App 2.28.0 and CLI 1.4.9. The default daily-cloudcode-pa.googleapis.com route still fails with the same TLS/socket error.
Testing CLOUD_CODE_URL=https://cloudcode-pa.googleapis.com restored sign-in and discovery of 13 models, but generation returned 429 RESOURCE_EXHAUSTED for Argon High, Gemini Flash and Sonnet Low. The refreshed app quota page showed Google AI Pro with 100% remaining weekly and five-hour quota.
I then tested this alternate Google endpoint, which is referenced in the installed client:
CLI Argon High and Sonnet Low each completed a strictly read-only task against a local repository: read a README and return its first heading. Both returned the expected completed response. Checked file hashes and Git status were unchanged.
The standalone desktop app completed a separate project-free Gemini Flash Medium test and returned DESKTOP_OK.
Logs still show an ancillary aicode.googleapis.com/v1alpha:fetchConfig socket failure followed by the client’s CCPA fallback; inference on the alternate endpoint nevertheless completed.
I have retained this as a reversible local environment workaround for terminal and GUI launches. Existing Google-account authentication, certificate validation and endpoint security controls remain in place. I did not patch any client binary, disable security software, switch to separately billed API-key mode, or change repository files.
Could engineering confirm whether this alternate endpoint is supported for personal Google AI Pro accounts and provide the intended permanent routing/configuration fix? The current default route remains broken on this Mac, so I am treating the result as a local workaround rather than a general fix.
Thank you for the thorough follow-up and for testing those endpoints on Antigravity App v2.28.0. Here is what we can confirm regarding your questions:
Setting CLOUD_CODE_URL=https://daily-cloudcode-pa.sandbox.googleapis.com (or https://cloudcode-pa.googleapis.com) is not an officially documented or supported configuration for personal Google AI Pro accounts—in our official documentation, custom endpoint configuration is only supported via GOOGLE_GEMINI_BASE_URL when using a Gemini API key with "modelProvider": "gemini" (Authenticate — Point the CLI to a custom endpoint). Using the default client configuration is required for supported routing, and cloudcode-pa.googleapis.com returns 429 RESOURCE_EXHAUSTED because personal Google AI Pro Antigravity quota entitlements (Plans) are not routed through that host.
Because your independent curl and Python TLS checks outside of Antigravity also fail with Recv failure: Socket is not connected specifically on daily-cloudcode-pa.googleapis.com (and aicode.googleapis.com/v1alpha:fetchConfig) while succeeding on cloudcode-pa.googleapis.com at the same IP, the connection is being reset on the local network or endpoint filter path rather than by a local Antigravity configuration file. Ensuring daily-cloudcode-pa.googleapis.com and aicode.googleapis.com are not intercepted or reset by local endpoint security rules will allow the default route to connect without an environment override.
I retested on 30 October after your updated reply. The official updaters offered newer builds: the installed App now reports 2.28.2 and CLI 1.4.11; update verification and the app code-signature check passed.
For a fresh, isolated CLI process I removed CLOUD_CODE_URL from its environment. The default route still fails at loadCodeAssist with the socket error. Model configuration does not load, and the requested Gemini Flash model is then reported as unrecognized. No inference completed on the default route.
Fresh credential-free curl checks, with normal certificate validation, return exit 35 / HTTP 000 for https://daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist and https://aicode.googleapis.com/v1alpha:fetchConfig. The https://cloudcode-pa.googleapis.com comparison completes TLS and returns HTTP 404 to unauthenticated GET.
The same minimal inference check with the existing temporary sandbox-host override returns the completed response ANTIGRAVITY_CASE_195017_OK. I understand this override is unsupported for personal Google AI Pro; it remains a temporary local workaround, not evidence the default routing problem is fixed.
I also submitted the sanitized findings once through the authenticated App’s Give Feedback screen, referencing #195017. Server-log and screenshot attachments were off to exclude unrelated private project data. The form cleared without an error; no feedback reference ID was shown. Please confirm receipt/linkage, or provide a way to submit an isolated, sanitized diagnostic log rather than the full app logs.
No endpoint protection, network filters, TLS checks, authentication credentials or billing mode were changed. Tests used an empty workspace with no project code or financial services. Could you provide the exact supported default endpoint requirements and the status of the routing investigation mentioned in your earlier edited reply? These checks confirm the hostname-specific failure, but do not identify which component resets the connection.
Thank you for testing on Antigravity App v2.28.2 and Antigravity CLI v1.4.11 and sharing those results.
Because your independent curl and Python TLS tests also fail with a socket reset on the same host while succeeding on cloudcode-pa.googleapis.com at the same IP—and because your endpoint network filters (Microsoft Defender and GlobalProtect) remained unchanged during testing—the connection is being reset by a hostname-specific filter rule on the Mac or local network path.
Please try the following workarounds to isolate and resolve the connection reset:
Allowlist *.googleapis.com and the Required Production Endpoints in Your Endpoint Filters:
Per our Enterprise Network Allowlists documentation, the documented endpoints over HTTPS (TCP port 443) for the Antigravity App, Antigravity CLI (agy), and Antigravity for IDEs are:
businessaicode.googleapis.com and regional businessaicode.us.rep.googleapis.com / businessaicode.eu.rep.googleapis.com (AI Developer Tools Orchestration)
aiplatform.googleapis.com (Vertex AI / Gemini Enterprise API)
discoveryengine.googleapis.com (Discovery Engine)
To prevent endpoint security software from resetting specific Google API subdomains used by the background language_server and agy processes, configure Microsoft Defender and GlobalProtect to allow outbound connections to *.googleapis.com over TCP port 443 without TLS inspection or interception (or test briefly with those endpoint filters paused / on a mobile hotspot to identify which component is resetting the socket).
Clear the Cached macOS Session State and Sign In Again:
Once *.googleapis.com is allowed through your endpoint filter:
Quit the Antigravity App completely.
Back up and remove ~/Library/Application Support/Antigravity to clear any cached session state.
In the CLI, run agy /logout, restart agy, and sign in again.
How to Share an Isolated, Sanitized Diagnostic Log:
If the issue persists after testing the steps above, you do not need to upload full application logs via the in-app feedback form. Per Sharing diagnostics with support, you can copy only the relevant sanitized error lines from these local plain-text log files on macOS and paste them directly in this thread so that we can look into it further.
Antigravity CLI (Linux and macOS): ~/.gemini/antigravity-cli/cli.log
Thanks. Here are isolated, sanitized diagnostics from 31 October, with Antigravity App 2.28.2 and CLI 1.4.11.
A fresh CLI process, with CLOUD_CODE_URL removed only from that child and using an empty workspace in plan/sandbox mode, still fails before model configuration and produces no model response. Its dedicated --log-file contains this error (socket addresses redacted):
Post "https://daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist": read tcp [LOCAL_IP:PORT]->[REMOTE_IP:443]: read: socket is not connected
Credential-free curl comparisons, all directed to the same currently resolved Google IPv4 address with normal TLS verification:
https://daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist
exit 35, HTTP 000, Recv failure: Socket is not connected
https://aicode.googleapis.com/v1alpha:fetchConfig
exit 35, HTTP 000, Recv failure: Socket is not connected
https://cloudcode-pa.googleapis.com/v1internal:loadCodeAssist
exit 0, HTTP 404; TLS completed
With the existing sandbox-host workaround, a separate isolated CLI session completed the exact response ANTIGRAVITY_CASE_195017_OK using Gemini Flash Medium. I understand this override is unsupported; the default route remains unresolved.
The hostname-specific behavior is reproducible, but these comparisons do not identify which component resets the connection. I have not paused endpoint protection, disabled inspection, or broadened allowlists. I also deferred clearing the working authenticated session because the credential-free TLS failure occurs independently of application session state. No project files or project services were accessed.
The linked allowlist documentation covers Gemini Enterprise and does not list the two failing hostnames above. This case uses a personal Google AI Pro account. Please escalate the sanitized results to engineering and confirm the expected default endpoints for this account type, including whether the daily host is intended. Please also specify any narrow hostname/process requirements and diagnostic evidence IT should check before considering policy changes.
Because curl -sv https://daily-cloudcode-pa.googleapis.com -o /dev/null and https://aicode.googleapis.com/v1alpha:fetchConfig return a socket reset (exit 35) while https://cloudcode-pa.googleapis.com completes the TLS handshake at the same IP address, a local firewall or Deep Packet Inspection (DPI) rule on your Mac or network is interrupting connections from the background processes (language_server and agy) to those two hostnames.
To resolve this on the default route without using an unsupported endpoint override:
Test launching the CLI with Go’s post-quantum TLS key exchange disabled (using the latest Antigravity CLI v1.4.12 , as some firewalls and DPI filters drop connections containing post-quantum cryptography packets:
If your network requires routing traffic through a proxy to reach those Google services, set HTTPS_PROXY / HTTP_PROXY in your environment so language_server and agy use the same path as your browser (and if you run commands inside the CLI terminal sandbox, you can configure sandboxProxy in ~/.gemini/antigravity-cli/settings.json).
Please let us know if the default route connects after testing these steps.
Thanks for clarifying which endpoints are expected for personal Google Account plans.
CLI 1.4.12 is already installed. On 2 November, a fresh isolated CLI process with neither CLOUD_CODE_URL nor GODEBUG set, using an empty workspace and plan/sandbox mode, still failed before model configuration. No model response completed. The isolated log contains:
Post "https://daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist": read tcp [LOCAL_IP:PORT]->[REMOTE_IP:443]: read: socket is not connected
A separate test with the existing sandbox-host workaround completed the exact response ANTIGRAVITY_CASE_195017_OK using Gemini Flash Medium. I still regard that unsupported override as a partial workaround.
Fresh credential-free curl comparisons at the same resolved Google IPv4 address also reproduce exit 35 / HTTP 000 for the daily and aicode hosts, while the standard cloudcode host completes TLS and returns HTTP 404.
No HTTP/S proxy environment variables or enabled macOS system proxies were observed, so no proxy setting was changed. Endpoint filters, certificate verification, security hooks and authentication remain unchanged; no project files or services were accessed.
I have not run GODEBUG=tlsmlkem=0 yet. A single-process compatibility test is prepared, but authorization to change the default post-quantum key-exchange behavior is pending. I will report its result separately if authorized. The default route remains unresolved.
Really appreciate how methodically you isolated this with clean curl checks across 2.26.0, 2.28.0, and 2.28.2, @Taras_Fog, and the endpoint context from @Rebbanpalli_Naveen.
Your finding that unauthenticated curl returns exit 35 (SSL connect error / read: socket is not connected) on daily-cloudcode-pa.googleapis.com and aicode.googleapis.com while completing the TLS handshake cleanly (HTTP 404) on cloudcode-pa.googleapis.com points to a transport-layer TLS/SNI reset or IPv4/IPv6 dual-stack routing issue specific to those two hostnames on your network path:
Compare IPv4 (-4) vs. IPv6 (-6) & Resolved Anycast VIPs: Run these two commands in Terminal to see whether daily-cloudcode-pa.googleapis.com is resolving to a local ISP cache/anycast VIP (172.217.118.4) that drops the TLS ClientHello for that SNI:
curl -4 -Iv https://daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist
curl -6 -Iv https://daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist
```Test Encrypted DNS / Alternate Google Anycast Edge (`8.8.8.8` / `1.1.1.1`): On macOS, when an ISP DNS resolver maps `daily-cloudcode-pa.googleapis.com` and `aicode.googleapis.com` to a regional Google edge subnet (`172.217.118.x`) where middlebox SNI inspection or PMTUD drops the handshake before `ServerHello`, Go's `net/http` client inside `language_server` logs `read tcp ...->172.217.118.4:443: read: socket is not connected`:
```bash
# Check which IP 1.1.1.1 returns vs your default resolver and test --resolve
dig +short @1.1.1.1 daily-cloudcode-pa.googleapis.com
curl -Iv --resolve daily-cloudcode-pa.googleapis.com:443:142.250.185.138 \
https://daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist
If the --resolve check against a clean 142.250.x.x Google edge completes the TLS handshake (HTTP 404 / 401 instead of exit 35), pinning that Google edge IP in /etc/hosts for daily-cloudcode-pa.googleapis.com and aicode.googleapis.com lets the default client route authenticate and load your Google AI Pro entitlements without touching CLOUD_CODE_URL.