Agy Chat: an open-source VS Code chat for Gemini that tries to feel like Claude. Looking for contributors
I spent weeks using Google's Antigravity chat in VS Code, and it never gave me the feeling Claude Code does. The answers were fine. The chat around them wasn't: slow starts, a reply that shows up all at once, few ways to

I spent weeks using Google's Antigravity chat in VS Code, and it never gave me the feeling Claude Code does. The answers were fine. The chat around them wasn't: slow starts, a reply that shows up all at once, few ways to undo or steer.
So I started to search for open-source solutions which were not bad but still did not click right, so i decided why not build a better community based alternative: Agy Chat for Gemini, and I'm opening it up for contributors.
👉 Marketplace: https://marketplace.visualstudio.com/items?itemName=maks-pavlovskyi.agy-chat
👉 Code: REPO_LINK
What it is
A VS Code chat panel for Gemini. It talks to the Antigravity CLI (agy acp) over the open Agent Client Protocol (ACP). It started as a fork of formulahendry/vscode-acp. No telemetry. Not affiliated with Google.
What's already built:
- A Claude-style chat panel: timeline of steps, thinking blocks, copy buttons
- Rewind: go back to an earlier message and roll the code back with it
- Question panel: when the agent asks you something, you answer with tabs, options and an "Other" box
- Permission cards for file edits and commands (secrets are never auto-approved)
- Several chats at once, one agy process shared between them
- Slash commands, editor context (file, selection, problems) and image attachments
- Tests against the real agy plus headless Chrome (~400 tests)
"But it depends on the Antigravity CLI. So does Google's own chat."
True. Both run on the same engine, so neither can make the model itself faster. I measured where the time actually goes (agy 1.5.9, macOS):
| Step | Time |
|---|---|
| Starting agy + handshake | ~0.1 s |
| First chat session, cold | 1.7–2.1 s |
| Every session after that | ~0.01 s |
| Prompt to first text (Gemini's side) | 2–7 s, depending on the model |
So the extension's own part is small. The room to win is in how fast it feels:
- Have the session ready before you type. That saves ~2 s on the first message of every chat. Not built yet.
- Show text as it arrives. Some models send the reply in pieces; the panel should show each piece the moment it comes.
- Send less context with each prompt, and keep the panel light.
And the parts where a Claude-like chat beats a basic one have nothing to do with the model: rewind, steering, clear permission prompts, a calm UI.
Where I need help
- Speed: pre-warm the session, render streaming text inside the panel, measure Agy Chat against Google's chat on the same prompts
- UX: close the remaining gaps with Claude Code's panel
- Open VSX release, so it works in VSCodium, Cursor and other forks
- More agents? ACP is an open protocol. Gemini only for now, but that's open for discussion.
- Testing on Windows and Linux. So far it's only checked on macOS. There is always room to grow - more features and improvments are open for discussion
How to start
- Install it from the Marketplace and try it with your Gemini account
- Open an issue with what feels slow or wrong (screenshots help a lot)
- Pick an issue labelled
good first issue, or comment on this post with what you'd like to work on
Node 22 and the Antigravity CLI (agy) are all you need to build and run the tests.
If you've ever thought "this chat could feel so much better", come build it with me. 🙌
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.