More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    agy helpers: parallel Gemini sub-agents for Claude Code

    1

    Claude splits off independent, checkable side tasks.

    Free

    2 usesSecurity scanned
    agy helpers: parallel Gemini sub-agents for Claude Code

    agy helpers: parallel Gemini sub-agents for Claude Code

    Example session with this skill installed

    Use agy helpers to find the latest stable versions of React, Vue, Angular, Svelte and Next.js, with release dates and official sources.

    • Read your context and instructions
    • Compiled the agy helpers

    React 19.3.1, Vue 3.5.44, Angular 22.2.4, Svelte 5.57.3 and Next.js 16.4.3 are the latest stable releases as of 2026-11-01, all published in October 2026.

    agy run: 5 tasks (5 ok, 0 failed), 213 s (the helpers ran in parallel; Angular was slowest), 804,024 agy tokens. Full results: C:\Users\User\AppData\Local\Temp\agy-runs\20261101-225230-2972\results.json

    Caveats and next steps

    How these were checked: For each framework, the helper compared GitHub's latest release with npm's latest tag. The two matched for all five, and every helper reported high confidence. As you asked, I didn't open the pages myself. The dates are publish dates in UTC.

    Background: All five are patch releases. Angular 22.0.0 came out on 2026-06-03, and Next.js 16.4 was announced on 2026-10-06 at nextjs.org/blog/next-16-4. React's CHANGELOG.md only goes up to 19.3.0, so the 19.3.1 notes are only on the GitHub release page.

    • Pre-releases (nobody double-checked these):
      • Vue 3.6.0-rc.14 came out on 2026-10-30.
      • Next.js has 16.5.0-canary.36.
      • Angular has 22.3.0-next.3.
      • Svelte has no pre-release.

    Two details look wrong

    • React's newest canary is reported as 19.3.0-canary-…, which has a lower version number than 19.3.1.
    • The Angular helper found no v23 pre-release, even though Angular usually ships a new major version in November.

    To confirm on your machine: run npm view <pkg> version for react, vue, @angular/core, svelte and next. I can also ask the Angular helper one more question to check its npm next tag for a v23 pre-release.

    React's CHANGELOG.md

    Connects securely to your tools. The creator never sees your data.

    About this skill

    What it does

    agy helpers turns Google's agy CLI (the Antigravity CLI, running Gemini 3.8 Flash High) into a pool of parallel helper agents for Claude Code. Claude splits off independent, checkable side tasks. agy runs one headless helper per task, and Claude reads only a compact summary: each helper's answer with its sources and a confidence level.

    It's built for people who have a Google AI plan with agy and keep running into Claude usage limits. The heavy reading happens on Gemini, while Claude keeps the planning, judging and final answers.

    Measured results

    The same tasks, both ways, in Claude Code with Opus 5.5 (single runs, Claude cost at API list prices):

    TaskClaude aloneClaude + 5 agy helpers
    Summarize 5 long docs (290 KB)1:34, $0.822:08, $0.21 (−74%) + 226k Gemini tokens
    5 quick version lookups1:31, $0.443:38, $0.22 (−50%) + 715k Gemini tokens

    It cuts Claude usage most on reading-heavy work, and it is slower. Use it for legwork, not for one-step lookups.

    Good jobs for helpers

    • Web research on several topics at once
    • Checking a list of claims, versions or links
    • Summarizing or reviewing many files
    • A second opinion on a plan or a diff
    • Running approved tests or linters in several folders

    How it works

    • One command runs a whole batch from a JSON task list: up to 20 helpers at once, 6 by default.
    • Three helper agents with fixed tool lists (read-only, commands, edit). No browser and no sub-agents.
    • Per-task permissions through agy's --settings: web pages, MCP tools, exact commands, or edits in the task's own folder. It never uses --dangerously-skip-permissions.
    • Every run syncs your Claude Code MCP servers (user scope, project scope and approved .mcp.json servers) and plugins into agy, so helpers can use the same tools.
    • Structured answers (status, answer, sources, confidence, notes) through --json-schema, follow-ups to any helper, and a results.json with everything.

    Requirements

    • Claude Code
    • agy (Google Antigravity CLI), installed and signed in
    • PowerShell 7 (pwsh)

    Install

    Unzip into ~/.claude/skills/ so you get ~/.claude/skills/agy/SKILL.md, then run the health check:

    pwsh ~/.claude/skills/agy/scripts/agy.ps1 doctor -Smoke
    

    Or clone the GitHub repository and run pwsh ./install.ps1.

    Tested

    32 offline unit tests (CI on Windows, macOS and Linux) and a 12-check live end-to-end suite (Windows 10, agy 1.4.12). Open source under the MIT license: https://github.com/Mhrnqaruni/claude-code-agy-helpers

    • Tested on Windows 10 with agy 1.4.12. macOS and Linux pass the unit tests, but the live tests have not been run there.
    • It saves Claude usage, not time: in our tests it was 36% slower on reading-heavy work and more than twice as slow on quick lookups.
    • Gemini 3.8 Flash can be wrong (in testing a helper once called a two-year-old article the latest news), so check what matters.
    • agy does not cache prompts, so a research helper can use 100k+ Gemini tokens.
    • Commands must be listed exactly; agy documents no wildcards for command rules. agy's sandbox and auto mode were not available on Windows.
    • MCP servers that sign in with OAuth, legacy SSE servers and the claude.ai connectors can't be shared with agy.
    • Google's terms for scripting agy on consumer plans are unclear; check your plan.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    2 uses

    Times used to date

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 15 days ago

    • Free with an Agensi account

    Requires

    Google AI
    Antigravity
    Claude Code
    agy (Google Antigravity CLI), signed in
    PowerShell 7 (pwsh)

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Network Access
    Environment Variables

    Allowed Hosts

    antigravity.google
    generativelanguage.googleapis.com

    File Scopes

    ~/.claude/skills/agy/**
    ~/.gemini/config/**
    ~/.claude.json
    ~/.claude/plugins/installed_plugins.json
    .mcp.json
    .claude/settings*.json
    %TEMP%/agy-runs/**

    Runs headless agy processes from PowerShell, each with its own allow list: web pages, MCP tools, exact commands, or file edits in the task's own folder. It never uses --dangerously-skip-permissions. It reads Claude Code's MCP and plugin settings and copies approved MCP servers into ~/.gemini/config/mcp_config.json, installs three helper agents into ~/.gemini/config/agents, and writes run folders to the system temp directory. Environment variables fill ${VAR} placeholders in MCP settings and AGY_HELPER_MODEL can change the model.

    Listed15 days ago

    What's inside

    Frequently Asked Questions

    No. Claude plans, delegates and checks the answers. agy helpers only do well-scoped legwork such as research, reading and checks, and the final answer stays with Claude.

    agy runs on your Google plan's agy allowance, not your Claude usage. Claude only reads a short summary, about 1,100 tokens for six helpers in our tests.

    Usually not. In our measurements it was 36% slower on reading-heavy work and more than twice as slow on quick lookups. The win is Claude usage: 74% and 50% less in those same tests.

    Yes. Every run copies your Claude Code MCP servers (user scope, project scope and approved .mcp.json servers) into agy, and removes ones you deleted. Servers that sign in with OAuth and legacy SSE servers can't be shared.

    Helpers can only run the exact commands you list for a task, and edits are limited to the task's own folder. The skill never uses --dangerously-skip-permissions, and helpers have no browser or sub-agent tools.

    Gemini 3.8 Flash High by default. You can pick another agy model with the AGY_HELPER_MODEL environment variable.

    New in Agents & Orchestration