Carries the load so Claude can focus on the climb.
Claude Code plugin. Routes tasks to the right model: Antigravity CLI for research, search, large-file analysis, and brainstorming — Antigravity or Codex for write mode and context handoff (your choice). Claude stays master orchestrator, focused on decisions, architecture, and final writes.
| Mode | Trigger | Result |
|---|---|---|
| Delegation | large file, log, codebase | Delegate reads/searches, Claude gets summary |
| Web Search | search / look up / docs for | Antigravity searches the web — faster, no Claude WebFetch tokens |
| Onboarding | "understand this project" | Q (quick Antigravity) · GC (Antigravity + Codex parallel, gap-fill loop, critical analysis) · X (Codex only) |
| Brainstorm | "brainstorm alternatives for X" | Antigravity always · Codex optional (user chooses) · Claude synthesizes |
| Plan Review | before complex implementation | Delegate flags edge cases and failure points |
| Write Delegation | "use codex/antigravity to write this" | Named CLI codes with context package, Claude reviews. No CLI named → Claude asks which |
| Rate Limit Recovery | Claude hits limit mid-task | Antigravity/Codex resume from handoff file or session log |
| Handoff | context heavy / "save tokens" | Claude packages context, hands off to Codex or Antigravity (your choice) |
| Prompt Optimizer (beta) | /sherpa:prompt-optimizer (no prompt = paste in browser) | Hook-intercepted — ~0 token overhead · Antigravity/Claude/Codex rewrites · browser editor · model picker · Codex/Antigravity models auto-discovered · click Use It |
| Optimize Mode (beta) | /sherpa:optimize-mode on|off | Persistent auto-optimize every prompt · choose backend · toggle off anytime |
1. Install CLI dependencies
# Required — Antigravity CLI (agy)
# Windows PowerShell:
irm https://antigravity.google/cli/install.ps1 | iex
# macOS/Linux:
curl -fsSL https://antigravity.google/cli/install.sh | bash
# Then sign in once, interactively:
agy
# Optional (enables write mode and context handoff)
npm install -g @openai/codex2. Install plugin
/plugin marketplace add JoshuaAdes/sherpa
/plugin install sherpa@SherpaTo update
/plugin marketplace update JoshuaAdes/sherpa
/plugin update sherpa@Sherpa
/reload-pluginsAuto-updates:
/plugin→ Marketplaces → JoshuaAdes/sherpa → Enable auto-update
To reinstall from scratch
/plugin marketplace remove Sherpa
/plugin marketplace add JoshuaAdes/sherpa
/plugin install sherpa@Sherpa- Claude spots a delegatable task
- Asks:
Y once · A always · C Claude · N skip - Delegate runs, output pipes inline to Claude
- Claude acts on the compressed result
Rolling session log in .sherpa/. Sherpa routes reads to Antigravity by default — Claude reads direct when delegation is skipped, failed, or user chooses C.
Explicit slash commands — complement automatic delegation, don't replace it.
| Command | What it does |
|---|---|
/sherpa:handoff |
Package session → sherpa-handoff.md, hand off to Codex or Antigravity |
/sherpa:brainstorm [topic] |
Antigravity always · Codex optional (G or GC prompt) · Claude synthesizes |
/sherpa:search [query] |
Antigravity web search |
/sherpa:onboard |
Quick or deep project map via Antigravity |
/sherpa:prompt-optimizer [prompt] (beta) |
Antigravity/Claude/Codex rewrites prompt · browser editor · backend + model picker · Codex/Antigravity models auto-discovered per account · re-optimize active card · Use It to submit |
Claude Code excels at coding decisions, architecture, and final writes. A lot of real engineering work isn't that: research, docs, web search, file reads, log scans, codebase exploration.
Antigravity's large context window handles token-expensive analysis without burning Claude tokens. Antigravity brainstorming brings genuinely different intelligence — not a budget substitute. Antigravity's built-in search covers what Claude Code can't do natively. Codex or Antigravity take over for write-heavy tasks and context handoffs — pick whichever fits the task, on explicit request.
Sherpa intercepts each task, routes it to the right model, and pipes back a compressed result. Claude spends tokens only where its reasoning actually matters. When context gets heavy, you can trigger a handoff — Sherpa packages the session into a handoff file and your chosen CLI drives to completion. If a rate limit hits mid-task, the rolling session log means work resumes without losing a step.
MIT + Commons Clause Free to use, modify, and deploy internally. Selling this plugin or products built primarily on it is not permitted. See LICENSE.