› cavi-ai/
GitHub ↗

Models

Use agy models for the live catalog available to the current Antigravity session. Reference the OpenClaw id as antigravity-cli/<model>. Effort suffixes (-high, -medium, -low) on agy models rows are not part of that id.

Discovery asks agy models and serves one catalog row per model. Reconnect writes those rows into the provider config. The plugin also keeps a static fallback snapshot and aliases such as pro, flash, sonnet, opus, and gpt-oss. Aliases expand before the CLI run. A non-empty model id not yet present in the static catalog is passed through to agy --model instead of being rejected by the plugin. The legacy Claude Opus 4.6 transport uses claude-opus-4-6-thinking; current Opus models use their catalog id.

Thinking is the OpenClaw thinking param, not part of the model id. It is passed as agy --effort, limited to the levels agy models lists for the model: low, medium, high for Gemini Flash and Claude 5.5; medium for GPT-OSS. A level outside that snapshot is passed through when the live agy models listing includes it. agy requires the flag on those models, so off sends the lowest listed level. An unlisted level sends the nearest listed level, and a tie sends the lower one. When no level is selected, the default is medium if that model lists it, otherwise the lowest listed level. Legacy Claude 4.6 models reject --effort, so the flag is omitted for them.

Catalog context-window values are conservative budgeting floors because the CLI does not publish per-model limits through this plugin boundary. Reported per-token cost stays zero unless the Google provider catalog has a positive rate for that same model id. agy models does not supply a rate. A zero cost is not a claim that the subscription itself is free.