Use a Google Gemini subscription in OpenCode V2 — through the Antigravity or Gemini CLI client protocol.
Full tool calling, streaming, model discovery, a live quota sidebar and quota-aware auto-continue.
Warning
Unofficial and not affiliated with Google. This plugin emulates Google's own clients
(Antigravity / Gemini CLI) to use their subscription entitlements from OpenCode. Google's
terms may treat this as a policy-violating use case and may restrict or terminate accounts.
Use it only with your own account, at your own risk, and read LEGAL.md first.
Antigravity access depends on your plan and account. The Gemini CLI profile is rejected for
the individual tier (UNSUPPORTED_CLIENT); it remains useful for paid Code Assist
Standard/Enterprise subscriptions only. If you need the lowest-risk setup, use the official
Gemini API key or Vertex AI provider instead.
No warranty and no liability. No guarantee is given for the information in
LEGAL.md. To the maximum extent permitted by law, the authors accept no
liability for damages, data loss, account restrictions, suspensions, terminations, charges
or any other loss arising from the use of this software. Use at your own risk.
OpenCode ships first-class support for ChatGPT and Copilot subscriptions, but not for Google's consumer plans. This plugin closes that gap without an API key:
- the Antigravity profile speaks the current consumer protocol and gives you Gemini, Claude and GPT-OSS models from a Google AI Pro/Ultra plan;
- the Gemini CLI profile speaks the Code Assist protocol for paid Standard/Enterprise organizations;
- both keep OpenCode's native Gemini request format, so function calling, thought signatures and streaming work unchanged.
OpenCode ──Gemini REST──▶ loopback proxy ──Cloud Code Assist──▶ Google
│
├─ OAuth tokens (per profile)
├─ quota gate + auto-continue
└─ model discovery
| Feature | What it does |
|---|---|
| Zero-loss tool calling | OpenCode's Gemini payload passes through untouched: tools, parametersJsonSchema, functionCall/functionResponse, thoughtSignature |
| Streaming | Native SSE, unwrapped from the Code Assist envelope by an incremental parser |
| Thinking | Requests ask for Gemini thought summaries (includeThoughts). Whether text arrives is up to the gateway; see Thinking |
| Model discovery | fetchAvailableModels for Antigravity; ids, names, context/output limits and image support come from your account |
| Two account limits | 5-hour window and weekly limit per pool (retrieveUserQuotaSummary), not per-model noise |
| Auto-continue | On quota errors the request stays open, the gate wakes at the reset timestamp, then the same request resumes |
| Sidebar | Live usage panel: 5h/1w for Antigravity, API usage for Gemini CLI |
| Quota tool | Agents can inspect account, project, tier, limits and gate state with gemini_quota |
| Local-only credentials | Tokens live in a 0600 JSON file; no telemetry, no analytics |
- OpenCode 2.x (
opencode --version). - Node 22.18+ for the helper CLI (the plugin itself runs inside OpenCode's Bun runtime).
- A Google account with Antigravity access, or a paid Code Assist Standard/Enterprise subscription for the Gemini CLI profile.
This repository ships no Google credentials. Login needs the installed-app OAuth client pair of the client whose entitlement you use. Provide it per profile via environment:
export OPENCODE_ANTIGRAVITY_CLIENT_ID="..."
export OPENCODE_ANTIGRAVITY_CLIENT_SECRET="..."or in the plugin options:
The Gemini CLI profile uses OPENCODE_GEMINI_SUB_CLIENT_ID /
OPENCODE_GEMINI_SUB_CLIENT_SECRET (or options.oauth["gemini-cli"]). Without these values
login fails with a clear message. Obtaining a compatible identity is your responsibility; see
LEGAL.md.
Clone the repository somewhere stable:
git clone https://github.com/officialPlocki/opencode-gemini-sub.git ~/.local/share/opencode-gemini-sub-src
cd ~/.local/share/opencode-gemini-sub-src
npm install --omit=devAdd it to the global OpenCode config (~/.config/opencode/opencode.json):
{
"$schema": "https://opencode.ai/config.json",
"plugins": [
{
"package": "/home/you/.local/share/opencode-gemini-sub-src",
"options": { "profiles": ["antigravity"] }
}
]
}Reload OpenCode (/reload), then run /models and pick an antigravity/... model.
Note
OpenCode caches plugin modules by path. After updating the code, bump the version and
install into a versioned directory (npm run install:local does this) or restart the
service, otherwise the old module stays loaded.
/antigravity-login
A browser opens (the URL is also posted into the session). Use a browser on the same machine: the OAuth callback is a loopback port.
Headless machines:
node bin/gemini-sub.js login --profile antigravityThe CLI prints the URL and waits for the redirect. Credentials are shared with the plugin.
Other commands:
node bin/gemini-sub.js status --profile antigravity --quota
node bin/gemini-sub.js quota --profile antigravity
node bin/gemini-sub.js models --profile antigravity
node bin/gemini-sub.js logout --profile antigravityFor the paid Code Assist profile use --profile gemini-cli and /gemini-login.
- Run
/modelsand choose a model, e.g.antigravity/gemini-3.8-flash-low. - Work as usual. Tool calls, subagents and streaming are handled by OpenCode.
- Open the sidebar to see the remaining 5-hour and weekly limits of the active pool:
GEMINI USAGE
Gemini Models
▶ ready
5h ██████ 98% frei · 4h22
1w ██████ 100% frei · 6d23h
While a request is parked on a quota window:
⏸ waiting 1m · 18:07
The panel is visible whenever one of the plugin's models is selected — idle or running. It disappears when you switch to a model from another provider.
| Command / tool | Purpose |
|---|---|
/antigravity-login · /antigravity-logout |
Sign in/out of the Antigravity profile |
/gemini-login · /gemini-logout |
Same for the Gemini CLI profile (if enabled) |
/gemini-status |
Accounts, projects, tiers, limits, gate state, unavailable models |
/gemini-models |
Models exposed by each profile |
/gemini-autocontinue on|off |
Toggle waiting for a quota reset at runtime |
gemini_quota (tool) |
Machine-readable status for agents |
Enabled by default. On a quota or rate-limit error the proxy does not forward the 429:
- classifies the error (Google RPC
RetryInfo,QuotaFailure,Retry-After, prose hints); - resolves the reset from
retrieveUserQuotaSummary(5h/weekly) or the error itself; - parks all in-flight requests on a shared gate until reset + a small buffer;
- keeps the SSE connection alive with comment frames;
- retries the identical envelope, so the agent loop and tool state survive.
Terminal cases are reported immediately: spending caps, limit: 0, a reset beyond
maxQuotaWaitMs, or an ineligible account.
Runtime toggle (not persisted):
/gemini-autocontinue off
/gemini-autocontinue on
Persistent default:
{ "options": { "autoContinue": false } }{
"plugins": [
{
"package": "/home/you/.local/share/opencode-gemini-sub-src",
"options": {
"profiles": ["antigravity"],
"autoContinue": true,
"maxQuotaWaitMs": 86400000,
"keepAliveMs": 15000,
"concurrency": 2,
"antigravityVersion": "2.24.0",
"models": [{ "id": "gemini-3-flash", "context": 1048576 }]
}
}
]
}| Option | Default | Meaning |
|---|---|---|
profiles |
["antigravity","gemini-cli"] |
Enabled connection profiles |
autoContinue |
true |
Wait for the quota reset instead of returning the 429 |
maxQuotaWaitMs |
86400000 (24 h) |
Longest accepted wait; 0 = unlimited |
keepAliveMs |
15000 |
SSE comment interval while parked |
concurrency |
2 |
Max parallel upstream requests per profile |
antigravityVersion |
2.24.0 |
Hub version reported in the User-Agent |
antigravityUseLegacyHeaders |
false |
Add older X-Goog-Api-Client/Client-Metadata headers |
endpoint |
per profile | Override the backend host (tests/proxies) |
projectId |
resolved | Force a Cloud project |
models |
built-in + discovery | Add or override model metadata |
dataDir |
~/.local/share/opencode-gemini-sub |
Credential directory |
includeThoughts |
true |
Ask Gemini models for thought summaries (Claude/GPT-OSS never get this field) |
telemetry |
false |
Send usage metrics like the original client |
mock |
false |
Canned responses, development only |
debug |
false |
Log proxy decisions |
Client identity is configured through the environment variables or oauth options described
in OAuth client identity. No defaults are shipped.
Antigravity models are discovered per account (fetchAvailableModels). Typical ids:
gemini-3.8-flash-high | -medium | -low | -tiered
gemini-3.5-flash-lite | gemini-3.1-flash-lite | gemini-3-flash
claude-sonnet-4-6 | claude-opus-4-6-thinking | gpt-oss-120b-medium
Interna (chat_*, tab_*), aliases (gemini-2.5-*), image and disabled entries are filtered.
Availability is account-specific: Google serves different ids to different plans, and models
such as gemini-4-argon-disabled stay hidden until the backend enables them.
The plugin sets generationConfig.thinkingConfig.includeThoughts = true for Gemini models (never for Claude or GPT-OSS, which reject the field). Reviving the text in the transcript also needs the CLI setting:
{
"session": {
"thinking": "show",
"tps": true
}
}session.tps shows output tokens per second in assistant footers.
Gateway limitation: for tool-enabled requests the Antigravity backend often returns a thought part with empty text — a signature carrier, not a summary. I verified this against the API: the same request without tools returns a readable summary, with tools the only thought part has text: "". Some models (for example gemini-3-flash or gemini-3.1-flash-lite) are more generous than the 3.8 Flash family, but no client-side change can create text the backend does not send.
| Symptom | Cause / fix |
|---|---|
Not signed in |
/antigravity-login |
redirect_uri_mismatch |
Google rejected the loopback callback; check the exact error page |
UNSUPPORTED_CLIENT |
Gemini CLI profile: Google disabled the individual tier; use Antigravity or a paid project |
no Cloud project after onboarding |
Set GOOGLE_CLOUD_PROJECT to a project with the Gemini for Google Cloud API enabled |
| Model missing | Not in your account's inventory, or filtered as internal/disabled |
maxQuotaWaitMs error |
Raise the limit or retry later |
| Panel shows stale numbers | It refreshes via RPC every 2.5 s; /gemini-status forces a quota lookup |
Server logs contain [gemini-sub] ... lines. Set debug: true for per-request decisions.
- OAuth tokens are stored locally in
~/.local/share/opencode-gemini-sub/*.jsonwith mode0600. - No analytics, no telemetry by default, no third-party servers. Requests go to Google only.
- The loopback proxy binds
127.0.0.1and uses a random secret path segment. - Treat the credential files like passwords. Revoke access from your Google account if in doubt.
See SECURITY.md for reporting and details.
No warranty, guarantee or representation is made regarding the information in this repository
(including LEGAL.md). It is general information for users, not legal advice, and may be
incomplete or outdated. Obtain independent legal advice for your situation.
To the maximum extent permitted by applicable law, the authors and contributors accept no liability for any direct, indirect, incidental, special, consequential or punitive damages, loss of data, loss of profits, business interruption, account restrictions, suspensions, terminations, billing charges or other losses arising from the use of or inability to use this software — even if advised of the possibility of such damages. You use it at your own risk and are solely responsible for compliance with Google's terms and applicable law.
Haftungsausschluss (Deutsch)
Die in diesem Repository enthaltenen Informationen – insbesondere in LEGAL.md – dienen
ausschließlich der allgemeinen Orientierung und sind keine Rechtsberatung. Für Richtigkeit,
Vollständigkeit, Aktualität und Eignung wird keine Gewähr übernommen.
Soweit gesetzlich zulässig, wird keine Haftung für Schäden, Datenverluste, entgangenen Gewinn, Betriebsunterbrechungen, Kontosperren, Kontokündigungen, Abrechnungen oder sonstige Nachteile übernommen, die aus der Nutzung oder Nichtnutzung dieser Software entstehen – auch bei Kenntnis der Möglichkeit solcher Schäden. Die Nutzung erfolgt auf eigenes Risiko.
This project is an independent, clean-room implementation. No Google source code is copied or vendored. It is not affiliated with Google or the OpenCode project, and it comes with no warranty. Using it may violate Google's terms and can lead to account restrictions.
See LEGAL.md for what the plugin does with your account, what Google may do, and
the liability terms. Nothing in this repository is legal advice.
npm install
npm test # protocol, profiles, quota, proxy, envelopes
npm run typecheck
npm run install:local # sync into ~/.local/share/opencode-gemini-sub/<version>The test suite covers both envelope dialects, retrieveUserQuotaSummary parsing with a
per-model fallback, error classification, quota park/resume and the loopback proxy. End-to-end
validation in OpenCode uses mock: true:
opencode models | grep antigravity
opencode run --model antigravity/gemini-3-flash "hello"
opencode run --auto --model antigravity/gemini-3-flash '[[tool:read:{"path":"/etc/hostname"}]]'src/index.ts plugin: providers, commands, tool, catalogue, RPC
src/tui.tsx sidebar panel
src/proxy.ts loopback Gemini REST server + quota policy
src/codeassist.ts Cloud Code Assist client (both dialects)
src/profiles.ts OAuth clients, hosts, headers, envelopes, fallback models
src/oauth.ts Google OAuth loopback flow
src/credentials.ts per-profile credential store + refresh
src/errors.ts Google RPC error classification
src/sse.ts envelope <-> Gemini SSE translation
docs/PROTOCOL.md reverse-engineered protocol notes
MIT — see LICENSE.