Antigravity Model Bridge

A lightweight VS Code Language Model Provider for the Google Antigravity backend.

Pick Gemini 3.8 Flash, Gemini 3.1 Pro, Claude Sonnet 5.5, Claude Opus 5.5 or GPT-OSS 120B from the standard model picker and keep using native chat and Agent mode. Requests run on your Antigravity account rather than Copilot premium requests.

Unofficial. Not affiliated with, endorsed by, or supported by Google. This talks to undocumented v1internal endpoints, so it can break when Antigravity updates. Use a Google account you are willing to risk.

Highlights

• Native integration — the models appear in the standard model picker, in Chat and Agent mode.
• Automatic model discovery with a curated default, so the picker shows what Antigravity offers rather than every internal alias — new generations appear on their own.
• Full tool calling — built-in, extension and MCP tools are forwarded to the backend, with schemas rewritten to survive its strict validator, including recursive stripping of MCP x-* vendor metadata and unique ids across multi-step Agent turns.
• Antigravity Web Search — optionally attach #antigravityWebSearch in Agent mode to run the installed agy CLI's native search_web tool and return grounded source URLs.
• Antigravity Image Generation — attach #antigravityGenerateImage in Agent mode to run the installed agy CLI's native generate_image tool and return the raster image inline, with optional aspect-ratio control.
• Native reasoning display — the model's thinking appears in Copilot Chat's own collapsible block.
• Context usage reporting — completed gateway token usage feeds VS Code's Context Window widget, including its prompt breakdown when the installed VS Code build supports it.
• Multi-turn reasoning continuity — thought signatures are replayed so tool-using conversations survive past the first turn.
• Quota groups in the status bar: all Gemini models (including Pro and Flash) share one displayed group; Claude/GPT remains separate, matching the native Antigravity usage display. The percentage is the lowest reported remaining fraction within each group.
• Thinking Effort control in the picker, next to the model, offering only the tiers the backend can serve.
• Multiple accounts, switched manually.

Tool execution, confirmation, workspace trust and permissions remain handled by VS Code. The Antigravity Web Search and Image Generation tools are the exceptions in transport: their registered runners start the local agy CLI, while Copilot still owns tool approval and the conversation loop.

The extension does not silently expose Gaussian/Codex's codexForCopilot_searchWeb marker to Antigravity models. That marker belongs to the Codex Responses backend. Use #antigravityWebSearch when you want the Antigravity-native search bridge.

Get started

1. Install the extension.
2. Run Antigravity: Sign in with Google from the Command Palette. A browser opens and the callback returns to http://localhost:51121/oauth-callback.
   • The port is not configurable — it is the redirect URI registered with the OAuth client, and Google rejects any other. If it is busy, close whatever holds it and retry.
3. Open Chat and pick an Antigravity model from the model picker.

Credentials live in VS Code SecretStorage and refresh automatically, including refresh-token rotation. Signing out removes only this extension's copy.

Model selection

Discovery returns every model the backend can route to, including inline-autocomplete models (Tab_*), internal numeric aliases (Chat_20706), server-side routing aliases (*tiered*), image models, duplicate display names, and past generations — dozens of entries in practice.

The default latest narrows that to Antigravity own published list: the three newest Gemini generations of each line, plus Claude and GPT-OSS. That list is a rolling window rather than a fixed set — it showed 3.5/3.6/3.7 Flash, then 3.6/3.7/3.8 once 3.8 shipped — so the window is encoded instead of the members, and a new generation appears while an obsolete one leaves on its own. Older Gemini and Flash Lite are excluded.

Set antigravity.modelSelection to all to see everything the backend can route to, or list ids in antigravity.hiddenModels to remove specific entries.

Effort tiers

Antigravity addresses each effort tier as its own model id, so discovery reports "Gemini 3.8 Flash (High)", "(Medium)" and "(Low)" as three separate models.

By default they are folded into one entry with a Thinking Effort control beside the model in the picker, the way Copilot's own models work. Only the tiers the backend actually routes are offered — Gemini 3.1 Pro shows High and Low but no Medium, because no such model id exists.

antigravity.reasoningEffort sets the starting value; whatever is chosen in the picker wins for that request. Set antigravity.effortSelection to tiers to go back to one picker row per tier.

Requirements

• VS Code 1.108.0 or later
• A Google account with Antigravity access (the free tier is enough)

Runs in the local UI extension host so it can reach the loopback OAuth port. With Remote-SSH, keep the extension installed locally rather than installing a second copy on the remote host.

Common commands

• Antigravity: Manage — accounts, quota, logs and settings in one place
• Antigravity: Sign in with Google
• Antigravity: Show Auth Status
• Antigravity: Refresh Models and Quota
• Antigravity: Open Logs

Diagnostics and privacy

Structured logging follows VS Code's log level, adjustable through Developer: Set Log Level…. Every request, auth flow and discovery call carries an operationId so retries and endpoint fallbacks can be traced.

Logs exclude prompts, reasoning, tool arguments and results, credentials and tokens — keeping only counts, sizes, transport decisions and redacted error text.

Context window

After a completed response, the extension passes the gateway's prompt, completion, cached, and reasoning token counts to VS Code through its internal usage response part. This changes the context indicator from 0 to the actual reported usage. VS Code 1.120 and newer can then render the same prompt breakdown used by native providers (system instructions, tool definitions, messages, and tool results); older builds may continue to show only the response stream.

Configuration

Under Settings → Extensions → Antigravity:

• antigravity.modelSelection — latest (default) or all
• antigravity.effortSelection — setting (default) or tiers, see above
• antigravity.hiddenModels — model ids to omit from the picker
• antigravity.reasoningEffort — starting value for the picker's Thinking Effort control
• antigravity.showThinking — show the model's reasoning (on by default)
• antigravity.toolCallIds — forward tool-call ids on the wire: claude (default), all, or off
• antigravity.projectId — override project discovery, if you see 403s naming a project
• antigravity.endpoint — pin generation traffic to one gateway host
• antigravity.showStatusBar — show remaining quota
• antigravity.cliPath — optional path to the agy executable used by Antigravity Web Search and Image Generation

Proxies come from http.proxy or HTTPS_PROXY/HTTP_PROXY, with NO_PROXY respected.

Reasoning

Reasoning is shown by default, rendered as Copilot Chat's own collapsible Thinking… block. The bridge sends the same completion marker used by Copilot's Agent host when thinking ends, so the answer stays a normal message instead of being attached to an unfinished reasoning card. LanguageModelThinkingPart is present in the VS Code runtime but absent from @types/vscode, so it is feature-detected: builds that lack it fall back to a markdown blockquote above the answer. Turn it off with antigravity.showThinking.

Not every model returns its reasoning. Gemini Pro and Claude send thought text and get a block; Gemini Flash reports thought tokens in usage but sends no thought parts, so nothing renders even though it is thinking. The log distinguishes the two — thoughtTokens>0 with thoughtParts=0 means the gateway withheld the text, which is a backend behaviour rather than a fault in this extension.

Web search

Antigravity Web Search is a VS Code tool backed by the locally installed agy CLI. Select it in the Agent tool picker or reference it as #antigravityWebSearch. The bridge starts a bounded headless agy run, instructs it to execute exactly one native search_web call, and returns the resulting markdown answer and source URLs. It does not reuse the Gaussian/Codex hosted search marker.

The CLI must already be authenticated. Set antigravity.cliPath if agy is not on PATH; on Windows the standard %LOCALAPPDATA%\agy\bin\agy.EXE installation is detected automatically. Search runs consume the CLI's own Antigravity quota and remain subject to its permissions and network policy.

Image generation

Antigravity Image Generation is a VS Code tool backed by the locally installed agy CLI. Select it in the Agent tool picker or reference it as #antigravityGenerateImage. The bridge starts a bounded headless run, instructs it to execute exactly one native generate_image call, and attaches the validated PNG, JPEG, GIF, or WebP result to the tool response. An optional aspectRatio hint (1:1, 16:9, 9:16, 4:3, or 3:4) can guide the native generator; final dimensions remain controlled by Agy's image backend.

The bridge accepts generated artifacts only from the matching Antigravity conversation's local brain directory and never reads arbitrary paths returned by the model. Newer Agy CLI builds may omit output_path from the headless event; the bridge then discovers only recent raster files in that same conversation directory. It does not expose the CLI's base64 payload or credentials to chat. Image generation uses the CLI's Antigravity quota, asks for VS Code confirmation before starting, and may take longer than a normal text response.

If the tool reports “generated no readable raster image artifact”, the Agy model call may still have succeeded but the artifact was not available or was written in an unsupported format/location. Retry once after the CLI finishes; the bridge rejects stale, non-raster, oversized, or outside-session files.

Develop locally

npm install
cp src/api/credentials.example.ts src/api/credentials.ts   # then fill in the values
npm run check          # typecheck + tests + bundle
npm run package:vsix

Press F5 for an Extension Development Host.

The OAuth client credentials are deliberately not in this repository — see src/api/credentials.example.ts for what to put there and where it comes from. The build refuses to run without it rather than producing an extension that cannot authenticate.

How it works

The gateway speaks a Gemini dialect for every model family, Claude included; Anthropic-style messages arrays are rejected outright. Most of the work lives in src/translate/:

• schema.ts — the backend validates tool schemas with protobuf JSON, which hard-fails on any unrecognised key. Since VS Code and MCP servers emit their own annotations (enumDescriptions, markdownDescription, …), schemas pass through an allowlist of the fields the Schema proto defines rather than a denylist of known-bad ones.
• toGemini.ts — a functionResponse is paired to its call by name, not by id, so the message walk records callId → toolName on the way past the call.
• thinking.ts — Gemini selects effort from the tier baked into the model id; only Claude takes a numeric budget, capped strictly below maxOutputTokens.
• api/agent-metadata.ts — requests carry the agent envelope: requestType, a five-segment requestId, telemetry labels and an FNV-1a session id, in the field order the real client uses.

License

MIT. Antigravity, Gemini and Google Cloud are trademarks of Google LLC.