Apra Fleet v0.2.1 -- Antigravity (agy) is now supported

[kumaakh]

Apra Fleet v0.2.1 is out, and the headline is Antigravity (agy).

You can now run a fleet member -- doer, reviewer, or PM -- on agy and dispatch across providers freely. Agy joins claude and gemini as a first-class option.

If you are picking a Google-side provider, we recommend agy over gemini. Google has signaled that Antigravity is their go-forward platform for coding agents and the standalone Gemini CLI is being phased out (the Gemini CLI repo now points users at agy). New work should target agy; we will keep the gemini path supported in Fleet for existing users.

Install with agy

apra-fleet install --llm agy

Sign in once on the machine via the browser OAuth flow (or set the API key env var for headless setups). Agy supports parallel dispatch, same as claude.

Tier-to-model mapping (per workspace)

Tier	Agy model
cheap	Gemini 3.5 Flash (Medium)
standard	Gemini 3.1 Pro (Low)
premium	Claude Opus 4.6 (Thinking)

Fleet writes the mapped model into <workspace>/.gemini/antigravity-cli/settings.json on every dispatch.

Mixed-provider fleets

Nothing new at the wire level -- you can mix and match. Examples:

# claude PM dispatching an agy doer + a gemini reviewer
# (just install each member with its provider; the PM orchestrates them all)

The PM skill, e2e suites, and dispatch tools all treat agy as a peer provider.

What else is in v0.2.1

Beyond agy support, this release also bundles:

• README overhaul (docs: value-first README overhaul + self-update fix #271) -- value-first front door, embedded demo, new docs/install.md / docs/troubleshooting.md / docs/writing-skills.md
• Docs accuracy pass (docs: llms.txt single source of truth + user-facing provider guide #272) -- new docs/provider-guide.md, new docs/beads.md, llms.txt as the single source of truth for the generated llms-full.txt
• apra-fleet update self-update bug fix
• Workspace defaultModel is now preserved across reinstalls (regression fix for all providers)
• Several install-side hardening: file permissions on hooks.json, deduplicated permission allow-lists, separated env-var namespaces

Full release notes: https://github.com/Apra-Labs/apra-fleet/releases/tag/v0.2.1

Known limitations (we're tracking)

• The workspace .ignore overlay we install for local agy members is written to disk but agy does not honor it at runtime yet, so a local agy doer/reviewer can still call apra-fleet MCP tools. Functional, just not isolated. Working on it.
• A few smaller follow-ups around install-test coverage, model-identity logging in agy transcripts, and a plan-review step in the PM skill -- all tracked, none blocking.

FAQ

See the FAQ Discussion #280 ("What LLM providers does Fleet support and how do they differ?") for the side-by-side comparison across claude, gemini, agy.

Feedback welcome

If you try agy with Fleet and hit anything -- expected or surprising -- drop a reply here or open an Issue. We are especially interested in:

• Cross-provider mixed-fleet patterns that work well for you
• Cases where the workspace .ignore limitation bites in practice
• Tier-to-model mapping preferences (we picked one set; the model field is per-workspace so you can override locally)