Summary
Build an experimental version of pyclaudir whose agent backend is Google's Antigravity CLI instead of Claude Code. Goal: run the same Telegram harness on a Gemini-powered agent and compare behavior, cost, and quality.
Motivation
- Antigravity CLI was announced at Google I/O 2026 (and is the successor to Gemini CLI, which sunsets June 18). It's an agent-first CLI — conceptually close to what pyclaudir already wraps.
- A side-by-side variant lets us compare Claude Code vs Antigravity CLI as the harness backend: speed, output quality, cost, tool/agent behavior.
- Pure experiment for now — "see how it goes," per the owner.
Scope
pyclaudir currently wraps the Claude Code CLI as a subprocess (stdin/stdout streaming). An Antigravity-backed variant needs an equivalent integration layer:
- Subprocess wrapper for the Antigravity CLI invocation.
- Streaming / message protocol adapter — map Antigravity CLI's I/O format to the harness's existing message loop.
- Auth model — Google account / API credentials instead of the Claude subscription.
- Tool surface — confirm whether the harness's MCP tools (
send_message, query_db, memory, etc.) work through Antigravity CLI.
Keep it a separate variant / branch / config, not a replacement — Claude Code stays the default.
Open questions — investigate first
Antigravity CLI is brand-new (announced today); its interface needs hands-on investigation before any implementation estimate. Step 1 is a spike to answer:
- Does Antigravity CLI support a non-interactive / headless /
-p-style mode the harness can drive?
- Does it support MCP servers? (I/O announcements mentioned MCP integration for Gemini Spark — needs confirming for the CLI specifically.)
- What's its streaming output format (JSON events?) and how close is it to Claude Code's?
- Auth + rate-limit / cost model.
- Tool-permission model — equivalent of
--allowedTools?
Suggested first step
A throwaway spike: get Antigravity CLI installed, drive it non-interactively from a script, and confirm the streaming + tool story. That spike decides whether a full variant is feasible and what it costs.
Requested by the owner.
Summary
Build an experimental version of pyclaudir whose agent backend is Google's Antigravity CLI instead of Claude Code. Goal: run the same Telegram harness on a Gemini-powered agent and compare behavior, cost, and quality.
Motivation
Scope
pyclaudir currently wraps the Claude Code CLI as a subprocess (stdin/stdout streaming). An Antigravity-backed variant needs an equivalent integration layer:
send_message,query_db, memory, etc.) work through Antigravity CLI.Keep it a separate variant / branch / config, not a replacement — Claude Code stays the default.
Open questions — investigate first
Antigravity CLI is brand-new (announced today); its interface needs hands-on investigation before any implementation estimate. Step 1 is a spike to answer:
-p-style mode the harness can drive?--allowedTools?Suggested first step
A throwaway spike: get Antigravity CLI installed, drive it non-interactively from a script, and confirm the streaming + tool story. That spike decides whether a full variant is feasible and what it costs.
Requested by the owner.