Antigravity OAuth Plugin for OpenCode (Revived & Extended)

English | Español

Note

Community Revival & Extended Edition: This repository is an active, open-source continuation of the archived project by NoeFabris (and earlier work by @jenslys and CLIProxyAPI). It fixes critical upstream bugs and unlocks the newest Google Antigravity models.

Important

Project Status (October 2026) - measured, not assumed.

Gemini 4 ("Argon") works, and the grant is per account. Measured 2026-10-31: an account that loadCodeAssist reports as free-tier received the full Argon grant (modelVersion: gemini-4-argon-trial plus its own Gemini 4 Argon weekly bucket), while two other accounts only received the gemini-4-argon-disabled placeholder and answer 404 for the real ids. Do not infer access from the plan name: enable model_discovery and confirm with antigravity_quota.
Argon has its own quota bucket. Two 32k-token Argon calls moved gemini-argon-wtu-weekly from 98.75% to 97.75% while gemini-weekly and gemini-5h did not move at all. Argon is not billed against Gemini Flash/Pro.
Rotation is entitlement-aware. Each account's model catalog is measured and cached, and selection skips accounts that provably lack the requested model. Argon only goes to the account that has it; gemini-3.8-flash keeps using every account.
Claude quota is per family, not per generation. Free accounts do have Claude quota, but it covers claude-*-4-6 and GPT-OSS; Claude 5.5 answers 404 there. antigravity_quota reports both generations separately.
Claude 4.6 is being retired. On the entitled account claude-sonnet-4-6 already answers 200 with "Claude Sonnet 4.6 is no longer available. Please switch to Claude Sonnet 5.5."; Free accounts still serve it. Google's announced date is 2026-11-02, and model_discovery surfaces the 5.5 successors without a plugin release.
Use Argon moderately. Measure it: ~1% of its weekly bucket per 65k input tokens, versus a Flash bucket that is effectively free at the margin. Keep Gemini 3.8 Flash as the high-volume workhorse.

What this fork changes

OpenCode v1 and v2 in one entrypoint. A callable v1 factory plus native v2 .setup(), with models, variants, tools and slash commands registered through the v2 registries.
Real Antigravity model catalog. The curated core is Gemini 3.8 Flash, the Claude 4.6 pair and GPT-OSS 120B; anything newer that your account is granted (Gemini 4 Argon, Claude 5.5) is registered by model_discovery from the backend catalog, one selector entry per family with low/medium/high variants.
Official CLI client signature. Requests present antigravity/cli/<version> (aidev_client; ...), resolved at startup from the CLI auto-updater. Without it the backend answers 404 NOT_FOUND for the newer Gemini lines - which is also why a stale signature used to look like "preview access not enabled".
Multi-account pool with entitlement-aware rotation. Per-account quota, health and measured model catalog; failover on 429, 403, transport drops and 404 misses; account cooldowns and circuit breakers so a bad account cannot poison the pool.
Dual-window quota reporting. antigravity_quota and /antigravity-quota show 5h and weekly windows per model family (Gemini, Claude/GPT, Gemini 4 Argon), short reset countdowns, a truthful USABLE flag and the measured capabilities of each account.
Thinking tiers and Claude signatures. Tiered backend ids, thinking budgets for Claude, signature caching and recovery for interrupted tool loops.
Safety controls. Configurable safety_level, plus safety_shield to surface and rotate on Google safety telemetry.
Session resilience. Historical tool-response pruning, transport-drop circuit breaker and retry instrumentation.
Request hygiene. Strips x-api-key/x-goog-api-key/x-goog-user-project, normalizes protobuf-incompatible tool schemas, and sanitizes payloads that the Cloud Code API rejects.

Enable OpenCode to authenticate against Antigravity (Google's IDE) via OAuth so you can use Antigravity rate limits to run gemini-3.8-flash, the Claude line (4.6 today, 5.5 on accounts that have it) and gpt-oss-120b, plus whatever newer model your account is granted - Gemini 4 Argon included - with your Google credentials.

What You Get

Gemini 3.8 Flash, Claude Opus/Sonnet 4.6, GPT-OSS 120B via Google OAuth, plus Gemini 4 Argon and Claude 5.5 for the accounts that are granted them (discovered per account, not hardcoded)
Multi-account support — add multiple Google accounts, auto-rotates when rate-limited
Request Normalization & Clean Feedback — optimized parameters prevent false-positives on development tasks; clean one-line message on filter notices
Slash Commands — /antigravity-quota to inspect dual-window limits, and /antigravity-stats for instant local metrics
Native Tools — antigravity_quota, antigravity_stats, and google_search tools registered directly in the agent runtime
Thinking models — extended thinking for Claude and Gemini 3 with configurable budgets / thinking levels
Google Search grounding — enable web search for Gemini models (auto or always-on)
Auto-recovery — handles session errors and tool failures automatically
Plugin compatible — works alongside other OpenCode plugins (oh-my-opencode, dcp, etc.)
Dual-Version Support — native compatibility with both OpenCode v1 (opencode-ai) and OpenCode v2 (@opencode/cli)

OpenCode v1 & v2 Architecture Compatibility

This plugin implements a polymorphic dual-compatibility entrypoint in index.ts:

OpenCode v1 (1.18.x / opencode-ai): Loads as a callable factory function:
```
export default async function ({ client, directory }: PluginContext): Promise<PluginResult>
```
Returns { auth, event, tool } hooks with full multi-account rotation and header interception.
OpenCode v2 (2.0.x / @opencode/cli): Loads as a native v2 plugin definition object:
```
export default {
  id: "opencode-antigravity-auth",
  setup: async (context: V2Context) => cleanup
}
```
Registers native tools (context.tool.transform), commands (context.command.transform), models (context.model.transform) and the Google Antigravity OAuth integration (context.integration.transform), while sharing the unified account pool and config. This makes opencode auth login offer OAuth with Google (Antigravity) and appends every newly authorized account to the multi-account rotation pool.

No configuration changes are required when transitioning between OpenCode v1 and v2.

Terms of Service Warning — Read Before Installing

[!CAUTION] Using this plugin (and any proxy for Antigravity) violates Google's Terms of Service. A number of users have reported their Google accounts being banned or shadow-banned (restricted access without explicit notification).

By using this plugin, you acknowledge:

This is an unofficial tool not endorsed by Google

Your account may be suspended or permanently banned

You assume all risks associated with using this plugin

Installation

Option A: Direct from GitHub (Recommended)

Add to plugins in ~/.config/opencode/opencode.json:

{
  "plugins": ["github:JoshRob297/opencode-antigravity-auth"]
}

OpenCode v2 will automatically clone, build, and update the plugin directly from the official repository.

Option B: Local Directory (Development / Offline)

Clone and reference locally:

git clone https://github.com/JoshRob297/opencode-antigravity-auth.git /path/to/plugin
cd /path/to/plugin && npm install && npm run build

Then in ~/.config/opencode/opencode.json:

{
  "plugins": ["/path/to/plugin"]
}

Quick Start (Zero-Config in OpenCode v2)

Add the plugin to ~/.config/opencode/opencode.json:

{
  "plugins": ["github:JoshRob297/opencode-antigravity-auth"]
}

Authenticate with Google:
```
opencode auth login
```
Select Google → OAuth with Google (Antigravity). (In SSH / Headless environments, the CLI automatically provides a direct URL to open on your browser and paste the verification code).
Ready to use! In OpenCode v2, all models, variants, and thinking budgets are auto-registered dynamically. You can immediately run:
```
opencode run "Hello" --model=google/antigravity-gemini-3.8-flash --variant=high
```
(Optional: You can run /antigravity-setup inside OpenCode anytime to cleanly auto-format your opencode.json with recommended whitelists).

Models

Model Reference

Model	Variants	Description
`antigravity-gemini-3.8-flash`	`low`, `medium`, `high`	Gemini 3.8 Flash with thinking tiers (Default)
`antigravity-claude-sonnet-4-6`	`low`, `medium`, `high`, `max`	Claude Sonnet 4.6 (retiring 2026-11-02)
`antigravity-claude-opus-4-6-thinking`	`low`, `medium`, `high`, `max`	Claude Opus 4.6 with extended thinking (retiring 2026-11-02)
`antigravity-gpt-oss-120b-medium`	`low`, `medium`, `high`	GPT-OSS 120B open-source model

Newer families are not hardcoded. model_discovery registers them from the account's own catalog; measured 2026-10-31 on an entitled account: antigravity-gemini-4-argon (low/medium/high), antigravity-claude-sonnet-5-5 and antigravity-claude-opus-5-5. Limits come from the backend (Claude 5.5: 1M context / 128k output).

Manual models configuration (Optional / Advanced)

In OpenCode v2 this is not required because the plugin registers all models and variants automatically via model.transform. However, if you wish to pin custom token limits or override model names manually, you can add this to ~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugins": ["github:JoshRob297/opencode-antigravity-auth"],
  "providers": {
    "google": {
      "models": {
        "antigravity-gemini-3.8-flash": {
          "name": "Gemini 3.8 Flash (Antigravity)",
          "limit": { "context": 1048576, "output": 65536 },
          "modalities": { "input": ["text", "image", "pdf"], "output": ["text"] },
          "variants": {
            "low": { "thinkingLevel": "low" },
            "medium": { "thinkingLevel": "medium" },
            "high": { "thinkingLevel": "high" }
          }
        },
        "antigravity-claude-sonnet-4-6": {
          "name": "Claude Sonnet 4.6 (Antigravity)",
          "limit": { "context": 200000, "output": 64000 },
          "modalities": { "input": ["text", "image", "pdf"], "output": ["text"] }
        },
        "antigravity-claude-opus-4-6-thinking": {
          "name": "Claude Opus 4.6 Thinking (Antigravity)",
          "limit": { "context": 200000, "output": 64000 },
          "modalities": { "input": ["text", "image", "pdf"], "output": ["text"] },
          "variants": {
            "low": { "thinkingConfig": { "thinkingBudget": 8192 } },
            "medium": { "thinkingConfig": { "thinkingBudget": 16384 } },
            "max": { "thinkingConfig": { "thinkingBudget": 32768 } }
          }
        },
        "antigravity-gpt-oss-120b-medium": {
          "name": "GPT-OSS 120B (Antigravity)",
          "limit": { "context": 131072, "output": 8192 },
          "modalities": { "input": ["text", "image", "pdf"], "output": ["text"] }
        }
      }
    }
  }
}

Model Discovery

The Antigravity backend reports which models each account can use. With model_discovery enabled the plugin registers the ones newer than the curated catalog on top of it, so a model released to your plan becomes selectable without waiting for a plugin release:

// ~/.config/opencode/antigravity.json
{
  "model_discovery": true
}

If any account in the pool has Argon or Claude 5.5, they show up. Gemini 4 Argon and the Claude Sonnet/Opus 5.5 successors are on the forward-only rule, and the catalog is taken as the union of every account: one entitled account is enough for the selector entry to appear. Access is granted per account, not by plan name - measured on 2026-10-31, an account reporting free-tier to loadCodeAssist still received the full Argon grant (modelVersion: gemini-4-argon-trial plus its own Gemini 4 Argon quota bucket). Discovery never claims support: it mirrors what the backend grants.
One selector entry per family, tiers as variants. The backend exposes one id per thinking tier (gemini-4-argon-low|medium|high, claude-sonnet-5-5-low|medium|high); the plugin registers a single model (antigravity-gemini-4-argon, antigravity-claude-sonnet-5-5) with a low|medium|high variant picker instead of three near-identical rows. Limits come from the backend when it declares them (Claude 5.5 reports 1M context / 128k output).
Rotation is entitlement-aware. Each account's catalog is measured from fetchAvailableModels and cached on disk, and account selection skips any account that provably lacks the requested model - so an Argon request never lands on a Free/AI Plus account that would answer 404 NOT_FOUND, while that same account keeps serving Gemini 3.8 Flash, which every plan has. A 404 that still happens rotates to the next account (bounded to one attempt per account) instead of surfacing a misleading "request preview access" error.
Quota notice on frontier models: Be advised that invoking Gemini 4 Argon burns through your 5-hour and weekly Antigravity quotas much faster than Flash models. Use it moderately for architectural decisions or tough problems, and keep Gemini 3.8 Flash as the high-throughput everyday engine.
Claude 5.5 carries an account-side safety monitor. Measured 2026-10-31: a long agent prompt (17k tokens of security-oriented system instructions) is answered with finishReason: SAFETY and an empty text part. Argon and Flash serve the same prompt normally. The plugin now turns that empty payload into an explicit "blocked by the model's filters" message instead of an opaque provider error.
Forward-only. Anything older than the curated workhorses (gemini-3.8-flash, Claude 4.6) stays out, curated ids are never duplicated, and ids measured as dead, retired, internal or -disabled are filtered by a denylist. A 200 from the backend is not treated as proof of usability.
Best-effort. If the catalog lookup fails, the curated catalog is used and startup is unaffected.

Full behavior, limits and caveats: docs/CONFIGURATION.md.

Hiding Foreign Google Models

models.dev publishes a google provider with the full Gemini lineup. Because the Antigravity OAuth credential makes that provider available, every one of those models shows up in the picker even though only the curated Antigravity models can be served. With hide_foreign_models enabled the plugin disables every google model that is not part of its curated catalog:

// ~/.config/opencode/antigravity.json
{
  "hide_foreign_models": true
}

Opt-in on purpose. If you have your own Google API key (GEMINI_API_KEY) the full models.dev list is valid, so the default is false.
Re-evaluated on every load, so a catalog refresh can no longer push the full Gemini lineup back into the selector.
Fail-open. An editor that cannot list models, or any error, leaves the list as is.

Multi-Account Setup

Add multiple Google accounts for a higher combined quota. The plugin automatically rotates between accounts when one is rate-limited.

opencode auth login  # Run again to add more accounts

Account management options (via opencode auth login):

Configure models — Auto-configure all plugin models in opencode.json
Check quotas — View remaining API quota for each account
Manage accounts — Enable/disable specific accounts for rotation

For details on load balancing, dual quota pools, and account storage, see docs/MULTI-ACCOUNT.md.

Request Handling, Safety & Feedback

Safety Levels (`safety_level`)

By default, the plugin mirrors Google's native moderation baseline (medium / BLOCK_MEDIUM_AND_ABOVE) to protect your accounts from telemetry anomalies.

You can configure safety_level in ~/.config/opencode/antigravity.json:

{
  "safety_level": "medium"
}

Available levels:

"medium" (Default) — Standard Google safety filters (BLOCK_MEDIUM_AND_ABOVE). Maximum account protection.
"high" — Blocks only high-probability harm (BLOCK_ONLY_HIGH). Recommended for extensive development, vulnerability analysis, and debugging.
"none" — Disables external safety classifiers (BLOCK_NONE) and bypasses prompt injection filters (HARM_CATEGORY_JAILBREAK).

Scope: safety_level is a Gemini control. It is never sent on Claude requests, because the Antigravity backend ignores safetySettings for the Claude line. Claude 5.5 is filtered by the provider's own monitor on the account: it can answer finishReason: SAFETY with empty content even with safety_level: "none" set, and the block is decided on the account side, not by this plugin. Measured 2026-10-31: the same request that mentions frida-server or "exploit generation" in the system prompt is blocked on claude-sonnet-5-5 and claude-opus-5-5 while claude-sonnet-4-6 and Gemini answer normally. Remove or rephrase the flagged context (usually AGENTS.md or project instructions), or use another model; the plugin surfaces the block with an explicit message instead of the bare "Provider blocked the response".

DISCLAIMER: Setting safety_level to "none" is strictly for authorized security research and advanced workflows. Using "none" is done solely at your own discretion and risk. We accept no responsibility or liability for account reviews, suspensions, or bans enacted by Google.

OPSEC Safety Shield (`safety_shield`)

When operating with safety_level: "none", Google's backend still evaluates your queries and attaches safetyRatings probabilities (NEGLIGIBLE, LOW, MEDIUM, HIGH) to responses.

The plugin provides an active defense engine configured in ~/.config/opencode/antigravity.json:

{
  "safety_level": "none",
  "safety_shield": {
    "enabled": true,
    "log_ratings": true,
    "show_toast": true,
    "auto_rotate_threshold": 2
  }
}

log_ratings: Emits detailed warnings in logs whenever Google assigns HIGH harm or MEDIUM jailbreak probability to an answer.
show_toast: Displays non-blocking OPSEC warnings in OpenCode's interface ([OPSEC Shield] Google flagged response: ...).
auto_rotate_threshold: Number of consecutive high-risk responses on a single account before triggering an automatic, preventive rotation to the next Google account (Account Shield). This prevents risk clustering on individual personal accounts. Set to 0 to disable auto-rotation.

Concise Notification Handling

If a query triggers a backend moderation notice, the plugin intercepts the verbose response (which contains lengthy legal disclaimers, policy links, and bug report URLs) and replaces it with a clean, actionable notice:

[Solicitud bloqueada por filtros de seguridad de Gemini. Por favor, intenta reformular tu prompt o enfoque.]

This prevents chat clutter, session lockups, and unrecoverable error cascades.

Troubleshooting

Quick Reset: Most issues can be resolved by deleting ~/.config/opencode/antigravity-accounts.json and running opencode auth login again.

Configuration Path (All Platforms)

OpenCode uses ~/.config/opencode/ on all platforms including Windows.

File	Path
Main config	`~/.config/opencode/opencode.json`
Accounts	`~/.config/opencode/antigravity-accounts.json`
Plugin config	`~/.config/opencode/antigravity.json`
Debug logs	`~/.config/opencode/antigravity-logs/`

Windows users: ~ resolves to your user home directory (e.g., C:\Users\YourName). Do NOT use %APPDATA%.

Custom path: Set OPENCODE_CONFIG_DIR environment variable to use a custom location.

Windows migration: If upgrading from plugin v1.3.x or earlier, the plugin will automatically find your existing config in %APPDATA%\opencode\ and use it. New installations use ~/.config/opencode/.

Multi-Account Auth Issues

If you encounter authentication issues with multiple accounts:

Delete the accounts file:

rm ~/.config/opencode/antigravity-accounts.json

Re-authenticate:
```
opencode auth login
```

403 Permission Denied (`rising-fact-p41fc`)

Error:

Permission 'cloudaicompanion.companions.generateChat' denied on resource 
'//cloudaicompanion.googleapis.com/projects/rising-fact-p41fc/locations/global'

Cause: Plugin falls back to a default project ID when no valid project is found. Fixed in v1.7.0+ by automatic discovery of the account's real managedProjectId.

Solution:

Go to Google Cloud Console
Create or select a project
Enable the Gemini for Google Cloud API (cloudaicompanion.googleapis.com)

Add projectId to your accounts file:

{
  "accounts": [
    {
      "email": "your@email.com",
      "refreshToken": "...",
      "projectId": "your-project-id"
    }
  ]
}

Note: Do this for each account in a multi-account setup.

Gemini Model Not Found

Add this to your google provider config:

{
  "providers": {
    "google": {
      "npm": "@ai-sdk/google",
      "models": { ... }
    }
  }
}

Gemini 3 Models 400 Error ("Unknown name 'parameters'")

Error:

Invalid JSON payload received. Unknown name "parameters" at 'request.tools[0]'

Causes:

Tool schema incompatibility with Gemini's strict protobuf validation
MCP servers with malformed schemas
Plugin version regression

Solutions:

Update to latest release from GitHub:

{ "plugins": ["github:JoshRob297/opencode-antigravity-auth"] }

Disable MCP servers one-by-one to find the problematic one

Add npm override:

{ "providers": { "google": { "npm": "@ai-sdk/google" } } }

MCP Servers Causing Errors

Some MCP servers have schemas incompatible with Antigravity's strict JSON format.

Common symptom:

Invalid function name must start with a letter or underscore

Sometimes it shows up as:

GenerateContentRequest.tools[0].function_declarations[12].name: Invalid function name must start with a letter or underscore

This usually means an MCP tool name starts with a number (for example, a 1mcp key like 1mcp_*). Rename the MCP key to start with a letter (e.g., gw) or disable that MCP entry for Antigravity models.

Diagnosis:

Disable all MCP servers in your config
Enable one-by-one until error reappears
Report the specific MCP in a GitHub issue

"All Accounts Rate-Limited" (But Quota Available)

Cause: Cascade bug in clearExpiredRateLimits() in hybrid mode (fixed in recent beta).

Solutions:

Update to latest beta version
If persists, delete accounts file and re-authenticate
Try switching account_selection_strategy to "sticky" in antigravity.json

Session Recovery

If you encounter errors during a session:

Type continue to trigger the recovery mechanism
If blocked, use /undo to revert to pre-error state
Retry the operation

Using with Oh-My-OpenCode

Important: Disable the built-in Google auth to prevent conflicts:

// ~/.config/opencode/oh-my-opencode.json
{
  "google_auth": false,
  "agents": {
    "frontend-ui-ux-engineer": { "model": "google/antigravity-gemini-3.8-flash" },
    "document-writer": { "model": "google/antigravity-gemini-3.8-flash" }
  }
}

Infinite `.tmp` Files Created

Cause: When account is rate-limited and plugin retries infinitely, it creates many temp files.

Workaround:

Stop OpenCode
Clean up: rm ~/.config/opencode/*.tmp
Add more accounts or wait for rate limit to expire

OAuth Callback Issues

Safari OAuth Callback Fails (macOS)

Symptoms:

"fail to authorize" after successful Google login
Safari shows "Safari can't open the page"

Cause: Safari's "HTTPS-Only Mode" blocks http://localhost callback.

Solutions:

Use Chrome or Firefox (easiest): Copy the OAuth URL and paste into a different browser.
Disable HTTPS-Only Mode temporarily:
- Safari > Settings (⌘,) > Privacy
- Uncheck "Enable HTTPS-Only Mode"
- Run opencode auth login
- Re-enable after authentication

Port Conflict (Address Already in Use)

macOS / Linux:

# Find process using the port
lsof -i :51121

# Kill if stale
kill -9 <PID>

# Retry
opencode auth login

Windows (PowerShell):

netstat -ano | findstr :51121
taskkill /PID <PID> /F
opencode auth login

Docker / WSL2 / Remote Development

OAuth callback requires browser to reach localhost on the machine running OpenCode.

WSL2:

Use VS Code's port forwarding, or
Configure Windows → WSL port forwarding

SSH / Remote:

ssh -L 51121:localhost:51121 user@remote

Docker / Containers:

OAuth with localhost redirect doesn't work in containers
Wait 30s for manual URL flow, or use SSH port forwarding

Configuration Key: `plugins` (v2) or `plugin` (v1)

In OpenCode v2, use the native plural key:

{
  "plugins": ["github:JoshRob297/opencode-antigravity-auth"]
}

OpenCode v1 uses the legacy singular key "plugin". The updater detects both automatically and preserves whichever convention your configuration uses.

Migrating Accounts Between Machines

When copying antigravity-accounts.json to a new machine:

Ensure the plugin is installed in ~/.config/opencode/opencode.json ("plugins" or "plugin")
Copy ~/.config/opencode/antigravity-accounts.json
If you get "API key missing" error, the refresh token may be invalid — re-authenticate

Known Plugin Interactions

For details on load balancing, dual quota pools, and account storage, see docs/MULTI-ACCOUNT.md.

Plugin Compatibility

@tarquinen/opencode-dcp

DCP creates synthetic assistant messages that lack thinking blocks. List this plugin BEFORE DCP:

{
  "plugins": [
    "github:JoshRob297/opencode-antigravity-auth",
    "@tarquinen/opencode-dcp@latest"
  ]
}

oh-my-opencode

Disable built-in auth and override agent models in oh-my-opencode.json:

{
  "google_auth": false,
  "agents": {
    "frontend-ui-ux-engineer": { "model": "google/antigravity-gemini-3.8-flash" },
    "document-writer": { "model": "google/antigravity-gemini-3.8-flash" },
    "multimodal-looker": { "model": "google/antigravity-gemini-3.8-flash" }
  }
}

Tip: When spawning parallel subagents, enable pid_offset_enabled: true in antigravity.json to distribute sessions across accounts.

Plugins you don't need

gemini-auth plugins — Not needed. This plugin handles all Google OAuth.

Configuration

Create ~/.config/opencode/antigravity.json for optional settings:

{
  "$schema": "https://raw.githubusercontent.com/JoshRob297/opencode-antigravity-auth/main/assets/antigravity.schema.json"
}

Most users don't need to configure anything — defaults work well.

Model Behavior

Option	Default	What it does
`keep_thinking`	`false`	Preserve Claude's thinking across turns. Warning: enabling may degrade model stability.
`session_recovery`	`true`	Auto-recover from tool errors

Account Rotation

Your Setup	Recommended Config
1 account	`"account_selection_strategy": "sticky"`
2-5 accounts	Default (`"hybrid"`) works great
5+ accounts	`"account_selection_strategy": "round-robin"`
Parallel agents	Add `"pid_offset_enabled": true`

Every strategy respects entitlements: an account whose measured catalog lacks the requested model is skipped for that model, regardless of how healthy or free it is.

Quota Protection & Transport Resilience

Option	Default	What it does
`soft_quota_threshold_percent`	`100`	Skip account when quota usage exceeds this percentage. Defaults to `100` (disabled) to avoid premature account exhaustion lockouts on low quotas. Set to `90` for conservative protection against Google penalties.
`quota_refresh_interval_minutes`	`15`	Background quota refresh interval. After successful API requests, refreshes quota cache if older than this interval. Set to `0` to disable.
`soft_quota_cache_ttl_minutes`	`"auto"`	How long quota cache is considered fresh. `"auto"` = max(2 × refresh interval, 10 minutes). Set a number (1-120) for fixed TTL.

How it works: Quota cache is refreshed automatically after API requests (when older than quota_refresh_interval_minutes) and manually via "Check quotas" in opencode auth login. The threshold check uses soft_quota_cache_ttl_minutes to determine cache freshness - if cache is older, the account is considered "unknown" and allowed (fail-open). When ALL accounts exceed the threshold, the plugin waits for the earliest quota reset time (like rate limit behavior). If wait time exceeds max_rate_limit_wait_seconds, it errors immediately.

Transport Circuit Breaker & Tool Pruning (v2.3.0+): To prevent 60-second gateway TCP resets (ECONNRESET) caused by massive command/tool dumps, outputs exceeding 30,000 characters in historical turns (older than the last 2) are automatically pruned with a clear summary snippet. Furthermore, a circuit breaker halts runaway retries after 3 consecutive transport dropouts on the same session, prompting context compaction.

Rate Limit Scheduling

Control how the plugin handles rate limits:

Option	Default	What it does
`scheduling_mode`	`"cache_first"`	`"cache_first"` = wait for same account (preserves prompt cache), `"balance"` = switch immediately, `"performance_first"` = round-robin
`max_cache_first_wait_seconds`	`60`	Max seconds to wait in cache_first mode before switching accounts
`failure_ttl_seconds`	`3600`	Reset failure count after this many seconds (prevents old failures from permanently penalizing accounts)

When to use each mode:

cache_first (default): Best for long conversations. Waits for the same account to recover, preserving your prompt cache.
balance: Best for quick tasks. Switches accounts immediately when rate-limited for maximum availability.
performance_first: Best for many short requests. Distributes load evenly across all accounts.

App Behavior

Option	Default	What it does
`quiet_mode`	`false`	Hide toast notifications
`debug`	`false`	Enable debug file logging (`~/.config/opencode/antigravity-logs/`)
`debug_tui`	`false`	Show debug logs in the TUI log panel (independent from `debug`)
`auto_update`	`true`	Auto-update plugin

For all options, see docs/CONFIGURATION.md.

Environment variables:

OPENCODE_CONFIG_DIR=/path/to/config opencode  # Custom config directory
OPENCODE_ANTIGRAVITY_DEBUG=1 opencode         # Enable debug file logging
OPENCODE_ANTIGRAVITY_DEBUG=2 opencode         # Verbose debug file logging
OPENCODE_ANTIGRAVITY_DEBUG_TUI=1 opencode     # Enable TUI log panel debug output

Troubleshooting

See the full Troubleshooting Guide for solutions to common issues including:

Auth problems and token refresh
"Model not found" / 404 on a model your account is not granted
Session recovery
IAM and Google Cloud project permission errors
Safari OAuth issues
Plugin compatibility
Migration guides

Documentation

Configuration — All configuration options
Multi-Account — Load balancing, dual quota pools, account storage
Model Variants — Thinking budgets and variant system
Troubleshooting — Common issues and fixes
Architecture — How the plugin works
API Spec — Antigravity API reference

Translations:

Español (README.es.md) — Spanish translation of this README. English is the source of truth; the translation may lag behind.
Español (docs/CONFIGURATION.es.md) — Configuration
Español (docs/MULTI-ACCOUNT.es.md) — Multi-Account
Español (docs/TROUBLESHOOTING.es.md) — Troubleshooting

Credits & Acknowledgments

This project stands on the shoulders of the open-source community:

Original Author: NoeFabris — Created the original opencode-antigravity-auth plugin.
Upstream Contributors:
Revival Maintainer: JoshRob297 — v1.7.0+ maintenance, Gemini 3.7 Flash support & IAM 403 fixes.

License

MIT License. See LICENSE for details.

Legal

Intended Use

Personal / internal development only
Respect internal quotas and data handling policies
Not for production services or bypassing intended limits

Warning

By using this plugin, you acknowledge:

Terms of Service risk — This approach may violate ToS of AI model providers
Account risk — Providers may suspend or ban accounts
No guarantees — APIs may change without notice
Assumption of risk — You assume all legal, financial, and technical risks

Disclaimer

Not affiliated with Google. This is an independent open-source project.
"Antigravity", "Gemini", "Google Cloud", and "Google" are trademarks of Google LLC.

Name		Name	Last commit message	Last commit date
.github		.github
assets		assets
dist		dist
docs		docs
script		script
scripts		scripts
src		src
.gitignore		.gitignore
AGENTS.MD		AGENTS.MD
AGENTS.md		AGENTS.md
CHANGELOG.md		CHANGELOG.md
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.es.md		README.es.md
README.md		README.md
SECURITY.md		SECURITY.md
index.ts		index.ts
package-lock.json		package-lock.json
package.json		package.json
tsconfig.build.json		tsconfig.build.json
tsconfig.json		tsconfig.json
vitest.config.ts		vitest.config.ts

Latest commit

History

Folders and files

Repository files navigation

Antigravity OAuth Plugin for OpenCode (Revived & Extended)

What this fork changes

What You Get

OpenCode v1 & v2 Architecture Compatibility

Installation

Option A: Direct from GitHub (Recommended)

Option B: Local Directory (Development / Offline)

Quick Start (Zero-Config in OpenCode v2)

Models

Model Reference

Model Discovery

Hiding Foreign Google Models

Multi-Account Setup

Request Handling, Safety & Feedback

Safety Levels (safety_level)

OPSEC Safety Shield (safety_shield)

Concise Notification Handling

Troubleshooting

Configuration Path (All Platforms)

Multi-Account Auth Issues

403 Permission Denied (rising-fact-p41fc)

Gemini Model Not Found

Gemini 3 Models 400 Error ("Unknown name 'parameters'")

MCP Servers Causing Errors

"All Accounts Rate-Limited" (But Quota Available)

Session Recovery

Using with Oh-My-OpenCode

Infinite .tmp Files Created

OAuth Callback Issues

Configuration Key: plugins (v2) or plugin (v1)

Migrating Accounts Between Machines

Known Plugin Interactions

Plugin Compatibility

@tarquinen/opencode-dcp

oh-my-opencode

Plugins you don't need

Configuration

Model Behavior

Account Rotation

Quota Protection & Transport Resilience

Rate Limit Scheduling

App Behavior

Troubleshooting

Documentation

Credits & Acknowledgments

License

Intended Use

Warning

Disclaimer

About

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Safety Levels (`safety_level`)

OPSEC Safety Shield (`safety_shield`)

403 Permission Denied (`rising-fact-p41fc`)

Infinite `.tmp` Files Created

Configuration Key: `plugins` (v2) or `plugin` (v1)