What LLM providers does Fleet support and how do they differ?

[kumaakh]

apra-fleet v0.2.1 added Antigravity (agy) as a third supported LLM provider alongside Claude and Gemini. What are the differences between them -- install command, auth setup, model tiers, parallel dispatch, and any known limitations?

[kumaakh]

Fleet supports three providers today: Claude, Gemini, and Antigravity (agy). The table below covers the key differences across install, auth, model tiers, parallel dispatch, and known limitations.

---

Provider comparison

	Claude	Gemini	Antigravity (agy)
Install flag	--llm claude (default)	--llm gemini	--llm agy
Auth mechanism	Claude Code OAuth -- provision_llm_auth automates it	Google account OAuth -- provision_llm_auth automates it	Google account OAuth -- run agy interactively once on the machine and sign in via browser; login persists. For headless/CI set ANTIGRAVITY_API_KEY. Auth is per-machine and set up out of band; provision_llm_auth does not apply to agy.
Tier-to-model mapping	Fixed: cheap/standard/premium map to Haiku/Sonnet/Opus variants	Fixed: maps to Gemini Flash/Pro/Pro tiers	Configurable per workspace via settings.json. Defaults: cheap = Gemini 3.5 Flash (Medium), standard = Gemini 3.1 Pro (Low), premium = Claude Opus 4.6 (Thinking)
Parallel dispatch	Yes -- Agent(run_in_background=True) enables true parallel member calls	No -- dispatches are sequential; each call blocks until complete	Yes -- parallel dispatch supported, same as Claude
Recommended for PM role	Yes -- most thoroughly tested for orchestration workflows	Caution -- sequential dispatch, output truncation (~8 KB), no --max-turns	Yes -- first-class PM option, equal peer to Claude
Recommended for member role	Yes	Yes -- 1M context window is an advantage for large codebases	Yes
Local .ignore MCP block	Honored	Honored	Not yet honored for local members (issue nx4) -- a local agy doer or reviewer can still call apra-fleet MCP tools. Remote agy members are not affected.

---

Install examples

# macOS/Linux
./apra-fleet-installer install --llm claude
./apra-fleet-installer install --llm gemini
./apra-fleet-installer install --llm agy

# Windows (PowerShell)
.\apra-fleet-installer.exe install --llm claude
.\apra-fleet-installer.exe install --llm gemini
.\apra-fleet-installer.exe install --llm agy

---

Auth setup by provider

Claude and Gemini: Run the installer, then call provision_llm_auth from the PM if a member's auth expires. The fleet handles the OAuth dance.

Antigravity (agy): Auth is per-machine and set up out of band. Run agy interactively once on the machine and sign in via Google OAuth in a browser; the login persists. For headless/CI environments, set ANTIGRAVITY_API_KEY as an environment variable instead (obtain from Google AI Studio). provision_llm_auth does not apply to agy.

---

Customizing agy model tiers

agy's tier mapping is controlled by a settings.json in the member's workspace:

{
  "llm": {
    "cheap":    { "model": "gemini-3.5-flash", "thinking": false },
    "standard": { "model": "gemini-3.1-pro",   "thinking": false },
    "premium":  { "model": "claude-opus-4-6",   "thinking": true  }
  }
}

This lets you tune cost vs capability per project, independently of what other members or workspaces use.

---

Quick decision guide

• Orchestrating a multi-member sprint? Use Claude or agy as the PM -- both are first-class options. For new Google-side setups, prefer agy over the standalone Gemini CLI.
• Want a different model perspective for review? Use Gemini or agy as the reviewer.
• Need workspace-level control over which model each tier hits? Use agy.
• Running unattended on a CI runner? Set ANTIGRAVITY_API_KEY for agy; for Claude set ANTHROPIC_API_KEY; for Gemini use provision_llm_auth.