Signs in to Antigravity with a Google account (Google AI Pro, Ultra or
the free tier) and calls the models Antigravity serves: Gemini, and the
Claude and GPT-OSS models Google hosts for it. Requests go through Google's
Code Assist API (daily-cloudcode-pa.googleapis.com) the way the
Antigravity app sends them. Provider id: antigravity. Ported from
magpie's built-in Antigravity account.
There are two ways to sign in:
-
Sign in with Google (browser) is Antigravity's own Google sign-in:
- The plugin listens on a port of
127.0.0.1(and::1). - It opens Google's consent page with PKCE and Antigravity's client and scopes.
- Google sends the browser back to
/oauth-callbackwith a code, which the plugin trades for a refresh token.
- The plugin listens on a port of
-
Import an account takes either a refresh token or the JSON an account manager exports:
- one account (
{"email", "token": {"refresh_token", "project_id"}}); - a list of accounts, or an
{"accounts": […]}object; - plain lines of refresh tokens, where
#starts a comment.
The import signs in the first account it finds. An export of another subscription (a
typeother than Antigravity) is refused. - one account (
The sign-in is kept as an oauth entry:
refreshis<refresh token>|<project>.accessis the access token, minted again from the refresh token when less than two minutes of it are left.accountIdis the account's email.
When Google refuses the refresh token (invalid_grant), magpie asks you to
sign in again.
Code Assist serves an account under a project:
loadCodeAssistnames the project.- An account without one is set up for its default tier through
onboardUser, as the app does on first sign-in. - The tiers are read again hourly, so a plan bought or lapsed shows up.
- The plan shown is the paid tier's name, else the current tier's.
fetchAvailableModels lists the account's models:
- The plugin names the account's paid and current tiers, as the app does.
- A tier Google refuses with a 400 is left out for a day.
- The list is kept for 10 minutes.
- Antigravity's internal and retired models are left out.
- The list keeps the order of the app's picker.
A model Antigravity lists at several levels (gemini-3.7-flash-low,
-medium, -high) is offered as one model, with a variant per level.
A request picks the level in this order:
- magpie's
x-magpie-effortheader; - the variant's
thinkingLevel; - the model's default level, which is high.
Asking for a level the model lacks picks the nearest one, and a tie goes up.
The plugin sends every request to v1internal:streamGenerateContent,
wrapped in the app's envelope ({model, project, userAgent: "antigravity", requestType: "agent", requestId, request}):
- The User-Agent is
antigravity/hub/<version> <os>/<arch>. The version is read from the app's update manifest every 6 hours. - The session id is a number made from the host's session header when named,
else the conversation's first real user prompt (skipping injected
AGENTS.mdand environment preambles), so unrelated conversations that share a preamble don't collide on one session. - Each reply chunk (
{response: …}) is unwrapped back into Gemini's own. - A
generateContentrequest gets the stream read whole.
The plugin also changes the request itself, as the built-in did:
- System prompt: lines Antigravity turns away (Claude Code's billing
header, Codex's and Hermes Agent's own prompts, omp's RFC 2119 line) are broken up with a
zero-width space.
DeepSeek Harness's paragraph on
@paths ("Tokens prefixed with @ are paths…") goes reworded with the same meaning: word for word, it got mostgemini-4-argonanswers fast, without thinking and shallow (#1028). - Tool calls:
- Tool call ids are reduced to
[A-Za-z0-9_-]. - A tool's
{output}or{error}response becomes{result}.
- Tool call ids are reduced to
- Schemas: a tool's schema is flattened to the subset Antigravity takes.
- Unions keep their own shape.
- Every array gets
items. nulltypes becomenullable.$refis resolved.- Blank enum values are dropped.
- Claude:
- A tool's images follow every function response.
AUTOfunction calling becomesVALIDATED.- A model without "thinking" in its name loses
thinkingConfig.
- Gemini:
maxOutputTokensis dropped.- Gemini 3 and later think at a
thinkingLevel. Earlier models lose a level a variant put on them.
| What Antigravity answers | What the plugin answers |
|---|---|
| 429 "Resource has been exhausted" to a system prompt with a turned-away line | 400 "turned this request away as from an unapproved channel", with what to change (it isn't a quota, so nothing is rested) |
| 429 from a plan's quota, the account having Google AI credits | with the Use AI credits switch on (off by default), the request is asked again once with enabled_credit_types: ["GOOGLE_ONE_AI"], as the app's Use AI Credits does; with it off, the 429 says the switch would spend them |
| 401 | 401, sign-in kept |
| 404 | 404 "model … not found" |
| Refresh refused | 401, x-magpie-sign-in: expired |
Each model has a window from its quotaInfo:
- The window reads 100% used when Antigravity leaves the fraction out.
- It says the model's family (Gemini, Claude, GPT-OSS).
- For a level, it also names the model magpie offers.
Then retrieveUserQuotaSummary's groups give the pools that models share,
each with a 5-hour and a 7-day window (aside). Each model is put in the
pool it reads from:
- a group whose versioned name is the model's own (Gemini 4 Argon);
- else the group of its family, when the model reads like one of that group's windows;
- else a pool of its own, named for the model, when it reads otherwise.
- Resting an account that Antigravity turned away for 6 hours, and asking its other accounts first. The 400 above makes magpie try the next account at once.
- Google Search grounding, and image models (drawing).
countTokensanswers 404, so magpie estimates the count.- Sending the system prompt as written: the plugin always breaks up the turned-away lines.
- Importing many accounts at once: an import signs in one account.
- Model names from models.dev: a model Antigravity leaves unnamed is named for its id.