Skip to content

One-click install for agentic tooling (Claude Code plugin + npx-runnable CLI) #8

Description

@oveddan

#2 makes the skill exist. This makes it installable in one step.

The current path

Per docs/agent-skill.md on agent/claude-codex-skill (#4), a new user has to:

  1. clone the repo
  2. npm install
  3. npm run build
  4. cp -R .claude/skills/mft-configurator ~/.claude/skills/
  5. cp -R .claude/skills/mft-configurator ~/.codex/skills/

Five steps and a local toolchain, before touching a knob. The README (#7) claims "you describe what you want and it gets done" — that claim is undercut by a five-step manual install.

What one-click should mean

Claude Code — a plugin, not a copied directory. Ship .claude-plugin/plugin.json with the skill bundled, plus a marketplace manifest in this repo, so installation is /plugin marketplace add oveddan/mft-api followed by /plugin install. Verify the exact manifest shape and commands against the current Claude Code plugin docs before implementing — do not copy the shape from this issue.

The CLI — runnable without a clone. npx @oveddan/mft-config export should work. That removes steps 1–3 entirely and means the skill can reference a stable command instead of node dist/cli.js.

Codex — a documented one-liner. There is no marketplace equivalent to target, so the honest goal is a single copy-pasteable command rather than pretending parity exists.

Acceptance criteria

  • A user with neither the repo nor Node experience can install and invoke the skill in Claude Code without cloning
  • npx runs the CLI against a connected Twister with no local build
  • Codex install is a single documented command
  • The skill references the published command, not node dist/cli.js
  • README install section replaced with the one-click path; the from-source path moves to a "develop" section

Depends on


Alternative to explore later

Dan, 2026-08-07: "it can also be like a command that you run or program you install or something, that Claude talks to."

Rather than a skill that shells out to a CLI per invocation, the tool could be a long-running local process the agent connects to — an MCP server, or a small daemon with a documented endpoint. Install it once, and every agent that speaks the protocol can talk to it.

Why it might be better than the skill-plus-CLI shape:

  • The Twister connection stays open instead of being rediscovered on every command.
  • One install serves Claude Code, Codex, and anything else that speaks MCP — no per-tool skill directory to copy and keep in sync.
  • Typed tool definitions instead of the agent composing shell strings, which is a narrower and safer surface for a tool that writes firmware settings over USB.
  • It matches how chromatik-mcp already works, so the two would install and behave alike.

Open questions before committing to it:

  • Does it replace the skill in Add a Claude Code and Codex compatible MFT skill #2, or sit under it?
  • Does a persistent process holding the MIDI port conflict with the DAW or the stock Twister editor also wanting it?
  • Is the plan/apply flow still legible to a human when it's tool calls instead of copy-pasteable commands?

Not scoped yet — recorded as an option, not a decision.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions