Hopper plugins

Plugins for Hopper: a realtime inference API for open-source models in your agent (DeepSeek V4.1 Flash, Gemma 4 31B, Qwen 3 TTS, Nemotron ASR).

Plugin	What it does
hopper	Test DeepSeek V4.1 Flash and Gemma 4 31B on your own prompts, point your project's OpenAI-compatible client at them, speak text and transcribe recordings. Skills plus the Hopper connector (withhopper.com/mcp).
hopper-announce	Claude Code says out loud when a long task finishes. Claude Code only (hooks).

Install

Host	Command
Claude Code	claude plugin marketplace add https://withhopper.com/marketplace.json then claude plugin install hopper@hopper (and hopper-announce@hopper)
Codex	codex plugin marketplace add hopper-inc/plugins then codex plugin add hopper@hopper
Cursor, VS Code / Copilot	Add the marketplace hopper-inc/plugins
Any agent that reads Agent Skills	npx skills add hopper-inc/plugins

No plugin? Paste this into your coding agent:

Set up Hopper in this project using https://withhopper.com/skill.md

Layout

.claude-plugin/marketplace.json      Claude Code marketplace
.agents/plugins/marketplace.json     Codex marketplace
.cursor-plugin/marketplace.json      Cursor marketplace
hosted/marketplace.json              the Claude marketplace as a URL (served at withhopper.com/marketplace.json), generated
server.json                          MCP Registry listing (com.withhopper/hopper), for when the connector ships
hopper/
  .claude-plugin/plugin.json         Claude Code manifest
  plugin.json                        Agent Plugins manifest + OpenAI listing (Codex, ChatGPT, Cursor, Copilot)
  mcp.json · .mcp.json               the connector: portable (with OpenAI's auth block) and Claude Code formats
  skills/<skill>/                    hopper-{try,integrate,speak,transcribe}, each self-contained:
    scripts/ · references/           copies of https://withhopper.com/agents/* (hopper_try.py and hopper_voice.py originate here)
  .codex-plugin/plugin.json          Codex layout, generated from plugin.json by tools/sync.mjs
  evals/                             eval cases with fixture projects; triggers.json for skill collisions
hopper-announce/                     Claude Code plugin: two hooks + one settings skill
claude-directory/hopper/             the plugin as submitted to Claude's plugin directory, generated by tools/directory.mjs:
                                     no text to speech (directory policy), connector https://withhopper.com/mcp/claude
tools/
  check.mjs                          directory rules + manifest consistency + claude plugin validate
  sync.mjs                           refresh the skills' scripts/ and references/ from withhopper.com
  directory.mjs                      build claude-directory/hopper (Claude's plugin directory: no text to speech)
  run-evals.mjs                      run the eval cases in Claude Code or Codex under your own login
  package.mjs                        dist/hopper-<version>.zip for the OpenAI plugin portal

Develop

node tools/check.mjs                                  # before every commit
node tools/sync.mjs --check                           # bundled copies match the published ones
node tools/directory.mjs                              # after any change in hopper/: regenerate the Claude directory copy
claude --plugin-dir hopper                  # try it in Claude Code
claude plugin eval hopper --tag offline --scaffold --trust-plugin --allow-tools Bash Write Edit
node tools/run-evals.mjs --host codex --tag offline   # the same cases in Codex (or --host claude)
node tools/run-evals.mjs --host claude --triggers      # which skill loads for 35 prompts, including ones that should load none

--tag live cases register a real trial key and call the production API; run them before a release, not in a loop.

Bump version in the manifests and server.json for every release and add a CHANGELOG entry; tools/check.mjs fails if the versions differ. Pick one install path per machine: the plugin or npx skills, not both.

License

MIT