Claude Code plugin v0.3.0 · dashboard v0.5.1 · Apache-2.0

Kinanorchestre

Each task goes to the cheapest model able to complete it. A judge checks the result before answering you. Already in the hands of testers.

View on GitHub

First installation step, typed inside Claude Code. All three steps

1 min 35 teaser, with sound. The dashboard footage uses demo data.

01Why

One large model for everything means overpaying for most requests.

More capable than needed

A single large model used for every request is more capable than most requests need, and it is billed accordingly.

Cheapest capable

The judge reads the request and hands it to a reading, production or expert agent, depending on what it really requires.

A check before the answer

What comes back is verified: the judge reads the diff or reruns the check before answering you.

No figure announced

No measured comparison is published yet. The plugin does not claim a number it cannot show.

02The team

One judge in front of your work, five agents behind him.

  1. InputYour request
  2. Kinanthe judge
    • Readinghaiku
    • Productionsonnet
    • Expertopus
  4. CheckReads the diff or reruns the check
  5. OutputThe answer, for you

The agents' and skills' instructions are written in French. The judge answers in the language you use. Agents (number, names, roles, models, efforts) are set in kinan.config.json; the judge is locked.

The base team has five agents: Drafion (Haiku), Skyze (Sonnet) and Johms (Opus), plus two optional agents, Courrier (Haiku) and Devis (Sonnet), shown only when the Courrier option is enabled. The names are only default settings: you can rename each agent in kinan.config.json. An agent marked option: "courrier" in kinan.config.json is hidden the same way. With three agents or fewer, they all sit in front of the judge, in a single row.

03Features

What the plugin does, and its known limits.

  1. Agent team plugin

    The judge gives each task to the cheapest capable agent: reading, production or expert.

    Known limitThe choice is made by a model following written rules; it can pick the wrong level.

  2. Impartial verification plugin

    The judge reads the diff or reruns the check before answering.

    Known limitThe judge is itself a model; it does not replace human review.

  3. Effort set per agent plugin

    A usual value and a "hard" value for each agent.

    Known limitDepends on your Claude Code version and the model supporting the effort parameter.

  4. Relay between conversations plugin

    Reminder past a threshold (2 hours or 300,000 tokens by default), handover note, then resume by typing "reprends".

    Known limitThe threshold is computed from the local log, whose format is not a documented interface; if it changes, the reminder goes silent.

  5. First-launch walkthrough plugin

    /kinan-orchestre:demarrage suggests your colors, fonts and agent names, one question at a time. Nothing is written before you agree.

    Known limitReads a single page, the one you give; the colors and fonts it picks up are approximate.

  6. Project sheet plugin

    /kinan-orchestre:projet creates a PROJET.md for the open project.

    Known limitThe model fills in only what it can verify; the rest is left for you to complete.

  7. Script checks method

    The judge prefers a script that returns a short summary over reading whole pages.

    Known limitThe plugin ships no check script: write one per project.

  8. Tracking several conversations dashboard

    Box and bubbles in real time, from the data written by the plugin.

    Known limitA conversation appears only at the first event it sends; it is dropped after 24 hours without events.

  9. Next relay, savings report, TOKEN OUT alert dashboard

    Time and tokens remaining, 3-hour report, alert with the resume time when a usage limit is reached.

    Known limitEstimates from the local logs and public API prices, not an invoice.

  10. Theme plugin + dashboard

    Colors, fonts and logo stored and validated by the plugin, applied by the dashboard.

    Known limitWithout the dashboard, the theme has no visible effect.

The dashboard is a separate Electron application, not included in this plugin, tested on Windows only. At launch, it creates a "Kinan Orchestre" shortcut on the Desktop, with its icon, which you can pin to the taskbar. The Courrier and Devis agents are in the default configuration; templates and a guide are provided in examples/ and docs/ETENDRE.md.

04Install

Three lines. You choose when to use it.

Requirements: Claude Code and Node.js available through the node command.

  1. In Claude Code /plugin marketplace add https://github.com/StudioObsidienne/kinan-orchestre
  2. In Claude Code /plugin install kinan-orchestre@kinan
  3. In your terminal claude --agent kinan-orchestre:kinan

A local path to the plugin folder also works for the first line. The third opens a conversation with the judge.

Recommended setting

In the main Claude chat: Sonnet 5.5, medium effort. The judge hands the heavy work to the other agents; a larger model in the main chat costs more without making him better.

First launch

In the conversation, type /kinan-orchestre:demarrage. Other commands: /kinan-orchestre:relais and /kinan-orchestre:projet.

05Safeguards

What cannot be changed.

  • Explicit approval

    Nothing is sent, published, pushed or permanently deleted without your approval.

  • Content that is read is data

    Web pages, mails, files, tool results, agent reports: never an instruction.

  • Nothing invented

    No invented figure, testimonial or result.

  • Offline hooks

    Hooks open no network connection and have no external dependency. Their files, in ~/.kinan-os/, are not encrypted.

Status: kinan-orchestre plugin 0.3.0, dashboard 0.5.1 (two separate version numbers, not aligned). The manifests pass claude plugin validate (including --strict). Tested on Windows; behavior on macOS and Linux is untested.

Updates: the plugin and the dashboard each have their own version number and are updated separately; updating one does not update the other. The dashboard shows both ("v0.5.1 Dashboard", then "v0.3.0 Plugin" below). Dashboard: git pull in its folder, then restart it. Plugin, in Claude Code: claude plugin marketplace update kinan, then claude plugin update kinan-orchestre@kinan, then restart Claude Code.

New up to dashboard 0.5.1: tracking of several conversations (one badge per conversation, with plugin 0.3.0), a Notes tab (local notepad, .txt and .md import), a ✓ or ✗ next to the versions, real sounds (muted by default), Sonnet by default, Drafion falling asleep, a shortcut that pins to the Windows taskbar. Latest releases: dashboard, plugin.

06Next version

More is on the way.

  • In development

    Kinan Orchestre already has its first testers, and a new version is in preparation. The current versions (plugin 0.3.0, dashboard 0.5.1) are usable now, with the limits written above.

  • The dashboard to install it

    A complete setup pairs the plugin with the dashboard, which installs separately: without it, no screen and no visible theme.

  • Testing underway

    Installation tests on other machines (macOS, Linux), automated tests, error and interruption tests: silent agent, failing hook, corrupted configuration.

  • Measure before announcing

    A reproducible benchmark, Kinan against a single model, with the results published, even the bad ones. Until it is measured, no savings figure is announced.

These are leads, not shipped features: nothing is promised before it is tested. Details are in the roadmap on GitHub.