AGY-API Admin Panel
✨ Automated Setup: After successful authentication, the system will automatically enable the required API services for your project.
- Gemini Cloud Assist API
- Gemini for Google Cloud API
Note: No need to enable APIs manually — the system handles these configuration steps automatically for a seamless auth experience.
📁 Advanced: Google Cloud Project ID
(optional — you can skip this and just click the button below)
▶
Authorization Link:
Instructions:
- Click the authorization link above — a Google OAuth page will open in a new window
- Sign in with your Google account and grant permissions
- After authorization, you'll be redirected to localhost:11451 with a success page
- Close the OAuth window and return to this page
- Click the "Get Credential File" button below to complete the flow
🚀 Can't reach localhost? Try the shortcut
▼
Credential File Contents:
🚀 Antigravity Auth Mode
Obtain Google Antigravity credentials.
Antigravity Authorization Link:
Instructions:
- Click the authorization link above and complete Google authorization in a new window
- After authorization, you'll be redirected to localhost with a success page
- Close the OAuth window and return to this page
- Click the "Get Credentials" button below to complete the flow
🚀 Can't reach localhost? Try the shortcut
▼
Antigravity Credential Contents:
Batch Upload Credential Files
Upload credential files in bulk.
📤 Credentials Batch Upload
Click to select files or drag and drop here
Supports .json and .zip files
ZIP files will be automatically extracted for JSON credentials inside
Selected Files:
Upload Progress:
0%
📤 Antigravity Credentials Batch Upload
Click to select files or drag and drop here
Supports .json and .zip files
ZIP files will be automatically extracted for JSON credentials inside
Selected Files:
Upload Progress:
0%
Credential Management
Manage all credential files — view status and perform operations.
📡 API Endpoints (click to copy):
| OpenAI format | |
| Claude format | |
| Gemini format | |
💡 About the Verify Feature:
Click the "Verify" button on any credential to re-fetch the Project ID and check the account's tier (free/pro/ultra).
✅ A successful verify can recover credentials from 403 errors and get them working again.
Use this when you encounter 403 errors.
Click the "Set Preview" button on any credential to configure the Preview channel.
✅ Setting Preview can resolve 404 errors when calling preview models with certain credentials.
Use this when you get 404 errors calling models like gemini-3-pro-preview.
0
Total
0
Active
0
Disabled
Batch Operations
Loading credentials...
Antigravity Credential Management
Manage all Antigravity credential files — view status and perform operations.
📡 API Endpoints (click to copy):
| OpenAI format | |
| Claude format | |
| Gemini format | |
💡 About the Verify Feature:
Click the "Verify" button on any credential to re-fetch the Project ID and check the account's tier (free/pro/ultra).
✅ A successful verify can recover credentials from 403 errors and get them working again.
Use this when you encounter 403 errors.
0
Total
0
Active
0
Disabled
Batch Operations
Loading credentials...
Configuration
Manage system configuration — changes take effect immediately after saving.
Loading configuration...
Server Settings
Host address to listen on. 0.0.0.0 listens on all interfaces.
Port to listen on. Requires a server restart to take effect.
Password for the chat API. Used to authenticate OpenAI and Gemini API endpoints.
Password for this control panel web interface.
(Legacy) If set, overrides the two passwords above. Leave blank to use separate password settings.
Basic Settings
Directory path where credential files are stored.
HTTP/HTTPS/SOCKS5 proxy endpoints. One per line. Leave blank to disable.
Endpoint Settings
Google Cloud Code Assist API endpoint URL.
Google OAuth2 API endpoint. Used for token retrieval and refresh.
Google APIs endpoint. Used for API service calls.
Google Cloud Resource Manager API endpoint. Used for project management.
Google Cloud Service Usage API endpoint. Used for enabling services.
External WebSearch API endpoint. When set, web_search tool requests are intercepted and forwarded here. Leave empty to disable.
Auto-Ban Settings
Automatically disable credentials when specified error codes are encountered.
Comma-separated list of error codes that trigger auto-ban.
Anti-Ban Protection (Antigravity)
Version string used in User-Agent and client identification headers. Update when new Antigravity binary versions are released.
Platform string used in User-Agent header (e.g., "windows/amd64", "linux/amd64", "darwin/arm64").
Prepend the official Antigravity system prompt to every request for ban prevention. This is the primary anti-ban mechanism. Enabled by default.
The system prompt prepended to Antigravity requests. Mimics the real IDE signature. Leave blank to use the built-in default.
Add a delay before each Antigravity API request to reduce rate-limit and detection risk. Disabled by default.
Delay in milliseconds before each request when throttle is enabled (50-5000ms).
Enforce per-credential requests-per-minute limit to prevent rate-limit bans. Disabled by default.
Maximum requests per minute per credential (supports decimals, e.g. 0.5 = one request every 2 minutes).
How long to wait for an RPM slot when all credentials are at limit (seconds).
Minimum time gap between requests on the same credential (seconds). Prevents burst traffic. 0 = disabled. Requires RPM Limit enabled.
Maximum number of successful auto-credit retries per credential. 0 = unlimited (no cap). Only applies to antigravity credentials with auto_credit enabled.
Retry Settings
Automatically retry on error.
Maximum number of retry attempts on error.
Wait time between retry attempts.
Credential Rotation Settings
Choose how credentials are selected for each request. ✓ Hot-reload supported
How it works: Each request independently picks a random credential from the available pool. Provides uniform distribution over time without any ordering overhead.
Use when: You want simple, stateless load balancing with no coordination between workers.
Use when: You want simple, stateless load balancing with no coordination between workers.
How it works: Credentials are selected by ascending
On 429/503 errors: The rate-limited credential is automatically pushed to the back of the rotation queue so the next request uses a fresh credential.
Use when: You need predictable, ordered credential cycling for easier monitoring or debugging.
rotation_order, cycling through all credentials in a predictable sequence.
On 429/503 errors: The rate-limited credential is automatically pushed to the back of the rotation queue so the next request uses a fresh credential.
Use when: You need predictable, ordered credential cycling for easier monitoring or debugging.
How it works: Credentials are bucketed by
On 429/503 errors: Failure counters self-update, naturally deprioritizing the affected credential without any explicit queue manipulation.
Use when: You want automatic load balancing based on actual usage, preventing individual credentials from being overloaded.
floor((success_count + failure_count) / N) where N is the Calls Per Rotation Bucket. The credential in the lowest bucket is chosen, with random tie-breaking within a bucket.
On 429/503 errors: Failure counters self-update, naturally deprioritizing the affected credential without any explicit queue manipulation.
Use when: You want automatic load balancing based on actual usage, preventing individual credentials from being overloaded.
Compatibility Settings
Converts all system messages to user messages and disables system_instructions. ✓ Hot-reload supported
⚠️ Note: This option may reduce the model's comprehension ability, but can prevent empty streaming responses.
Use when: The model returns no content or empty responses during streaming.
Use when: The model returns no content or empty responses during streaming.
When enabled, the model's reasoning chain is included in the response; when disabled, it is filtered out. ✓ Hot-reload supported
💭 Info: Some models (e.g., Gemini 2.0 Pro) support thinking mode and output their reasoning before the final answer. Enable to see the thinking process; disable to show only the final answer for cleaner output.
When enabled, non-streaming requests use the streaming API internally and collect the full response. ✓
Hot-reload supported
🔄
Info: Optimization for Antigravity mode. Even if the client requests a non-streaming response, the backend uses the streaming API to collect the full response before returning.
Use when: The streaming API is more stable than non-streaming in your environment.
Default: Enabled
Use when: The streaming API is more stable than non-streaming in your environment.
Default: Enabled
When enabled, always send thinkingConfig even when tool calls are present in conversation history. Strips injected thinking blocks instead of removing thinkingConfig. ✓ Hot-reload supported
💭 Info: Fixes the issue where thinking/extended thinking stops working after tool calls appear in the conversation. When enabled, the model will always use thinking mode regardless of tool call history.
When enabled, uses Anthropic-standard
signature field name and emits signature_delta streaming events for thinking blocks. ✓ Hot-reload supported
💡 Info: Standard Anthropic clients use
signature instead of thoughtSignature. Enable this for full compatibility with the official Anthropic API protocol. Input always accepts both field names regardless of this setting.
Default thinking budget for Antigravity mode when no model suffix is specified. ✓ Hot-reload supported
🧠 Info: Sets the default thinking budget used when no override is present.
Priority: Model suffix (-max, -high, -medium, -low, -minimal) > Client thinkingBudget > This config value.
Default: Medium (8192)
Priority: Model suffix (-max, -high, -medium, -low, -minimal) > Client thinkingBudget > This config value.
Default: Medium (8192)
When enabled, forces the thinking budget to 1024 for all Antigravity requests, overriding model suffix, client thinkingBudget, and budget_tokens. Prevents credential bans caused by clients sending excessively high thinking budget values. ✓ Hot-reload supported
⚠ Warning: When active, the thinking budget is always fixed at 1024 regardless of any client or model-suffix override. Disable when high thinking budgets are intentionally needed.
Default: Disabled
Default: Disabled
When enabled, sends a heartbeat SSE chunk every ~15 seconds of upstream silence to prevent reverse proxies and clients from dropping idle connections. Auto-enables when a proxy is configured. ✓ Hot-reload supported
💓 Info: Applies to all backends (GeminiCLI and Antigravity). Useful when nginx, Cloudflare, or other proxies drop idle SSE streams before the model responds. Automatically enabled when a proxy is configured; disable explicitly if not needed.
Controls text_delta event smoothing in Anthropic-format streaming responses (Antigravity only). Simple splits large chunks; Adaptive uses a drip buffer for continuous smooth output. ✓ Hot-reload supported
When enabled, 429 QUOTA_EXHAUSTED and RESOURCE_EXHAUSTED errors will set model-level cooldown. When disabled, these errors are treated as normal retryable errors (retry with different credential, no cooldown set). ✓ Hot-reload supported
⏱ Info: Disable this if Google's rate-limit algorithm has changed and the default cooldown timer is too aggressive. Credential rotation will handle 429s without setting per-model cooldowns.
When enabled, credentials that return matching error codes during "Test All Disabled" are permanently deleted (not just disabled). ✓ Hot-reload supported
Comma-separated list of HTTP error codes that trigger permanent credential deletion.
⚠ Warning: This is a destructive action. Credentials deleted by this feature cannot be recovered from within the application. Only enable this if you are certain that the configured error codes indicate permanently revoked credentials. Default error code: 403 (Forbidden).
Anti-Truncation Settings
Maximum continuation attempts when truncated output is detected.
The system instruction injected to require the model to emit the
[done] marker when its response is complete. Must contain [done].
Note: Anti-truncation is now controlled via the model name:
- Select a model with the "-stream-anti-truncation" suffix to enable
- This feature only applies during streaming
- Example: "gemini-2.5-pro-stream-anti-truncation"
Custom System Prompt
When set, this text is appended to the system instruction of every request. Leave blank to disable.
Telegram Alert Settings
Telegram bot token from @BotFather. Required for alert notifications.
Alert when active credentials drop to this number or below. Default: 1.
Keepalive Settings
When set, the service will periodically send GET requests to this URL to stay alive. Leave blank to disable. ✓ Hot-reload supported
Time between keepalive requests. Range: 5–86400 seconds. Default: 60. ✓ Hot-reload supported
TLS HTTP Mode
Choose which HTTP version the TLS fingerprint binary uses for upstream requests. ⚠ Requires server restart to take effect
HTTP/2 mode uses
tls_config_h2.json. HTTP/1.1 mode uses tls_config.json. The change takes effect after a server restart.
Hot-Reload Reference
🔥 Hot-reload settings (take effect immediately):
- Network: Proxy, endpoints, HTTP timeouts, max connections
- API: Rotation count, retry settings, auto-ban config
- Passwords: API password, panel password, universal password
- Features: Anti-truncation max retries
🔄 Settings requiring a restart:
- Server: Host address, port
- Directories: Credentials directory, Code Assist endpoint
Live Logs
View real-time system log output with filtering and auto-scroll.
Connection: Disconnected
Waiting for log stream connection...