SECURITY & PRIVACY · ANTIGRAVITY TOOLS
Antigravity Manager: Security, Privacy & Account Risks
Set up separate access for administrators and API clients, protect the account files on your machine, and check what becomes reachable when you enable LAN access or a tunnel. This guide also explains how to remove access when a credential is shared or a client is retired.
For version 5.0.0 · Updated
BEFORE CONNECTING AN ACCOUNT
Can my Google account be restricted?
Users have reported restrictions after using Manager. The available reports leave the frequency and cause uncertain, so consider that risk before connecting an account you depend on.
One user reported losing access to connected accounts; another reported using account switching without a ban. These are individual experiences. Account rotation, device binding and a lower request volume have no established ability to prevent restrictions.
Review the provider’s rules for the access method you intend to use. Google’s account-restriction guidance explains how its policies can affect one service or the whole account. The app’s code license is described separately in the project background.
If access stops, identify which layer rejected it
- Google sign-in says the account is disabled
- Open Google sign-in in your browser, read the explanation and use the appeal option provided for that account.
- Google sign-in works, but Antigravity or a model request is denied
- Read the product’s error message for an authorization, availability, quota or policy issue. Follow account-access troubleshooting with that exact message.
- Manager returns a local 401 or 403
- Check the API key, user-token restrictions and IP rules below. These checks can reject a request before it reaches the provider.
Use the recovery steps shown by the provider for your account. A verification trick described in a comment may apply to a different situation; keep passwords, tokens and authorization codes private throughout recovery.
FILES & REQUEST CONTENT
What does Manager store and send?
Manager stores account credentials and settings on its host, then uses those credentials to send model requests upstream. Protect both the stored files and the content your clients include in requests.
Protect account files, exports and backups
In version 5.0.0, account records are written as readable JSON in accounts/<id>.json. They contain OAuth credentials and account metadata. gui_config.json stores configuration, including the proxy API key and management password; other features use SQLite databases for logs, statistics and user tokens. The account-storage implementation shows the JSON write path.
The default data folder is ~/.antigravity_tools. A custom location or ABV_DATA_DIR can move it. Restrict read access to the people and processes that operate Manager, and keep exports and backups in protected storage. Use the data and backup guide to find the active folder before copying or deleting files.
Understand the request path
- Authorization: the browser and Manager communicate with Google to authorize the account, exchange tokens and obtain account profile information.
- Model use: the request can include prompts, conversation history, attachments and tool results. Manager forwards the required content to the configured upstream service. A localhost client URL points to the first step in this path.
- Optional services: a configured relay, upstream proxy or Cloudflare tunnel adds infrastructure to the connection. Review the destinations and settings you enabled.
- Saved diagnostics: logs and captures may retain request content locally. Sharing a capture or export creates another copy for its recipient.
Review which files and tool outputs your AI client includes before using it with confidential work. The key entered in that client authenticates its requests to Manager; Manager uses the connected account’s credentials for upstream access.
Prepare a log, HAR or capture for a support request
Work on a copy and keep only the requests needed to reproduce the problem.
- Remove Authorization headers, cookies, API keys in URLs, OAuth tokens, passwords and account emails.
- Read the bodies as well as the headers. Check prompts,
systemInstruction, tool inputs and results, attachments and response content. - Remove private paths, usernames, workspace names and conversation or trajectory IDs when they are unnecessary to diagnose the error.
- Inspect screenshots and embedded files, then reopen the cleaned artifact before sharing it.
Use a small sanitized example instead of an account JSON export. Follow the diagnostics guide for collecting the relevant error.
CONTROL WHO GETS ACCESS
Which password or token should I share?
Give API clients an API credential and keep the management password with the administrator. Set a distinct management password first: when it is empty, Manager accepts the proxy API key for administration too.
| Credential | Access it provides | Where to use it |
|---|---|---|
| Google OAuth tokens | Account access used for upstream authorization. | Manager’s account workflow and protected account storage. Keep them out of client configuration and support reports. |
| Proxy API key | Access to Manager’s model API. | The API-key field of a client you control. A shared key must be replaced for all its clients when revoked. |
| Management password | Web UI and management routes, including account and configuration access. | The administrator’s Web UI login. Use a value different from the proxy API key. |
| User token | Model API access with its own expiry, IP-count limit and curfew. | A separate client or user whose access you want to manage individually. For version 5.0.0 revocation, delete the token. |
| Cloudflare tunnel token | Permission for a connector to run its named tunnel. | The tunnel configuration. Configure visitor access separately at the edge and in Manager. |
Set the API key and management password
In API Proxy → Service Configuration, set API Key and Web UI Management Password to different, private values. Use the edit controls and save each change. In a Docker deployment, update the corresponding values in .env and recreate the container; environment-managed credentials are applied again at startup.
Create a separate credential for each API consumer
Open User Tokens → Create Token, enter a recognizable username and choose an expiry. Set IP Limit if you need to cap the number of bound addresses; zero means unlimited. Share the resulting token with that client through its API-key setting.
For Curfew (Service Unavailable Time), enter times in UTC+8 / Beijing time, which version 5.0.0 uses for this check. Plan for users whose network address changes, and keep the model API’s authorization enabled so expiry, IP and curfew checks run.
CHECK BEFORE SHARING
How do I check that authentication works?
With the proxy running, request its model list once without a key and once with an authorized key. Then check that the same API credential is rejected by a management route.
Choose an explicit authorization mode
In API Proxy → Service Configuration → Authorization → Mode, choose All (Strict) or All except Health for shared access. Changes to this mode are saved and applied to the running proxy.
- All (Strict) ·
strict - Requires authentication on model API routes and
/healthz. The login page, CORS preflight and certain service routes remain public. - All except Health ·
all_except_health - Protects the model API while allowing health and probe routes. Use it when a health monitor needs access without a key.
- Off (Open) ·
off - Allows model API requests without credentials and skips normal user-token restrictions. Keep shared services in an explicit protected mode.
- Auto (Recommended) ·
auto - Uses Off when Allow LAN Access is disabled, and All except Health when it is enabled. Set an explicit protected mode before opening a tunnel to a local-only instance.
Management routes require their own credential regardless of the model API mode. For headless deployment, also check ABV_AUTH_MODE and AUTH_MODE: a supplied override can change the startup mode.
Run a request without credentials
On Linux or macOS, run this command against your instance. Adjust port 8045 if needed:
curl --silent --output /dev/null --write-out '%{http_code}\n' \
http://127.0.0.1:8045/v1/modelsIn PowerShell, use curl.exe --silent --output NUL --write-out "%{http_code}" http://127.0.0.1:8045/v1/models. A direct request to a running Manager with protected authentication should return 401.
Check the allowed and denied requests
Use your HTTP client’s Authorization header field for credentials: Authorization: Bearer followed by the value. Keep credentials out of URLs and shared screenshots. Send the following requests directly to Manager and inspect the status:
- GET /v1/models with the proxy API key: expect 200 and a JSON model list.
- GET /v1/models with a user token: expect 200 when its expiry, IP and curfew rules allow the request.
- GET /api/accounts with no credential, then with the proxy API key or a user token: expect 401 when a distinct management password is set.
- GET /api/accounts with the management password: expect 200. Keep the returned account details private.
- GET /v1/models with the management password: expect 403 when that password differs from the API key.
A 503 response can mean the proxy is stopped. Open API Proxy, start forwarding, and repeat the request. If a keyless model-list request returns 200, close remote access and correct the mode before sharing the address again.
After these checks, use the first model request to confirm account-backed generation. Through a public hostname, also repeat the authorized and keyless requests using the exact path your client will use.
CHOOSE WHO CAN CONNECT
How should I expose Manager to other devices?
For personal remote access, start with a local-only service and the SSH tunnel setup. For shared access, configure authentication first, then choose the network route and verify it from another device.
LAN access
Allow LAN Access changes the listener from loopback to network interfaces. Set credentials and an explicit protected mode first, enable the option only when needed, then restart the service to apply the bind change.
Use HTTPS or an encrypted private connection for traffic crossing untrusted networks. Restrict the reachable port at your firewall and repeat the keyless and authorized requests from the intended client. See connection troubleshooting for timeouts or connection refusals.
Headless mode permits network listening unless ABV_BIND_LOCAL_ONLY is enabled. Docker adds its own host port mapping; the Docker network example keeps host access on loopback while letting the bridge reach the container.
IP allowlists and blocklists
Open IP Management. Add intended addresses or CIDR ranges under Whitelist → Add Trusted IP before turning on Enable Whitelist Mode in Security Config. Add unwanted addresses under Blacklist, enable the required protection and choose Save Changes.
Whitelist mode accepts listed addresses and rejects others. Whitelist Priority (Overrides Blacklist) lets a listed address take precedence over a matching blacklist rule. Allowed clients still need a valid credential.
Keep your administration address allowed, then test one allowed client and one denied client. In version 5.0.0, IP matching trusts X-Forwarded-For and X-Real-IP supplied with the request. When using a reverse proxy, restrict direct access to Manager and have the trusted proxy replace these headers. Keep authentication enabled alongside IP rules.
Recover from an IP-rule lockout
Use an administration connection that you included in the allowlist to restore the previous rules. If every route is blocked, close remote ingress, stop Manager, and restore the previous security configuration from a protected backup before restarting locally. Confirm local access and the intended rule set before reopening the remote route.
Cloudflare Tunnel
Quick mode creates a temporary tunnel. Auth mode uses a connector token for a named tunnel. Visitor authentication is configured through Manager and, where needed, Cloudflare Access policies.
- Set Manager to an explicit protected auth mode before starting the connector, including when its origin binds only to localhost.
- Publish only the API routes your clients need. Keep management
/api/and internal/internal/paths on a separate private administration route. A local connector can reach loopback-only routes on the origin. - Configure the intended edge access policy. Check that your API client can supply its required authentication as well as Manager’s API credential.
- Test the public address from a separate session or device. Check both the intended client and a request with no credentials. An edge login page should be handled before expecting a JSON API response.
- When finished, stop the connector, remove the public route and confirm that the hostname no longer reaches Manager.
REMOVE ACCESS
How do I revoke a key, token or account?
Replace or remove the credential that granted access, then test its old value against the relevant endpoint. Each type of access has its own revocation step.
USER TOKENS · VERSION 5.0.0
Delete a user token to revoke it
A user token marked Disabled can still authenticate in version 5.0.0. Delete the token in User Tokens, keep API authentication enabled, and confirm that its old value returns 403 from /v1/models. Use this deletion procedure when retiring a client or replacing an exposed token.
- Proxy API key
- Replace API Key, update your intended clients, and check that the old key is rejected. User tokens are managed separately, so review and delete those too if their access should end. If management used the API-key fallback, set a separate management password.
- Management password
- Replace Web UI Management Password and check that the old value returns 401 from a management endpoint. If credentials come from Docker environment variables, change the deployment values and recreate the container.
- Google authorization
- Open your Google Account’s linked apps, find the relevant connection under Access to your Google Account, review its details and choose Remove access. Follow Google’s linked-app instructions to identify the correct connection. Then remove the account from Manager if you are finished using it there.
- Stored account data
- Stop Manager and review its active data folder, account exports, backups and client copies. Delete the copies you no longer need using the data-management guide. Review provider authorization separately from local file removal.
- Cloudflare connector token
- Replace the connector token in Cloudflare and update the intended connector. Review active connectors and public routes. Rotate Manager’s own credentials separately if those were exposed.
If you suspect unauthorized access, close the remote route while changing credentials. Restore access only after the replacement works and the old value is rejected. Clients may need their settings updated or an account authorized again.
Before sharing the service
Confirm the intended client works, then check these four controls on the route you plan to share:
- Requests without a valid API credential are rejected.
- API clients are denied management access.
- Account files, backups and diagnostics have the intended read access.
- You can revoke a client’s credential and still administer the service.