orquestrar

Open plugin for AI agents

Ask your way. The agent brings the method.

orquestrar makes Claude, ChatGPT and Codex run any task like someone experienced: they work out what you want, test what the tool can do, check before delivering and stop to ask you before anything that can't be undone.

InstallView on GitHub
  • Claude
  • ChatGPT
  • Codex
  • Free and open
Example from a test roundCodex · with orquestrar
Youneed instagram posts for the whole month, i sell handmade candels
AgentI'll put together a month-long calendar with image ideas, copy and captions. I'll start with a sample so you can see if it fits your brand.
  1. How many posts a week? If unsure: three.
  2. Who do you sell to? If unsure: candles as gifts.
  3. What tone? If unsure: warm and simple.
Answer however you like, or just say “go ahead”.I don't have access to your Instagram: I won't post anything.
Translated summary of a real answer from the tests (case 05).

What changes

Same question, with and without a method.

A real test case: “my health plan denied a test my doctor ordered, I need to write them a letter or a petition. can you help?”

Without a method

  • Opens with five questions, none with a default answer.
  • Nothing starts until you answer them all.
  • Doesn't say how it will check the plan's rules.
  • Doesn't make clear the text needs review before it's sent.

With orquestrar

  • Starts drafting the letter and asks on top of the draft.
  • Three questions at most, each with a safe default.
  • Says it will check the applicable rules before claiming any right.
  • Delivers an editable document and reminds you that sending it is your call.

How it works

Five phases, and a triage first.

The agent goes through the phases on its own. You only come in when there's a decision that is yours.

  1. 0

    Triage

    A simple question gets a direct, complete answer. No ceremony.

  2. 1

    Intent

    Result, deadline and limits. Up to three questions, each with a default.

  3. 2

    Diagnosis

    Tests what the tool can do and picks the smallest way of working.

  4. 3

    Execution

    Prepares everything reversible. Stops before sending, publishing or deleting.

  5. 4

    Verification

    Checks numbers, sources and formats. Whoever produced it doesn't approve it.

  6. 5

    Delivery

    A document, spreadsheet or folder, with a short receipt of what was done.

Why use it

Less effort from you, more care from the agent.

No prompt writing needed

Ask as you'd ask a colleague. The agent finds what's missing with a few questions, each with a default answer.

Nothing irreversible without you

Emails, posts, submissions and deletions stay prepared and wait for your “send it”.

Checked before it reaches you

The agent says what it verified and what's still open. Whoever did the work isn't who approves it.

Something to use, not loose text

Results come as documents, spreadsheets or slides you can edit, plus a receipt of what was done.

Results

Tested blind, across three model families.

In each case, a judge gets two answers without knowing which used orquestrar: the same model with the plugin and the same model without it.

Claude

9/9

wins in 9 cases

with · 134 points

without · 94 points

Codex

8/9

wins in 9 cases

with · 101 points

without · 70 points

Gemini

9/9

wins in 9 cases

with · 126 points

without · 66 points

Version 0.4, with repetitions:16/1716 wins and 1 tie in 17 comparisons, no losses.
See the test reports

Install

Pick your app.

It takes a minute.

Claude

  1. Open Customize and go to the Plugins tab.
  2. Under Personal plugins, click + and choose Add marketplace.
  3. Choose to add from a repository and paste:Syntelix-AI/orquestrar
  4. Install orquestrar from the list.

Works on Claude web, desktop and Claude Code, which receives your account's plugins. If you had already uploaded the skill's .zip, remove that copy under Customize → Skills so you don't have two.

Packs by profession

The method is general. Every profession has its own way.

Packs teach the agent what belongs to each field: the right questions, typical documents, sources to check against, and where it must stop and ask. They're built from accounts of people who work in the field.

looking for professionals

Legal

Deadlines, filings, case-law research, and what is never filed without review.

looking for professionals

Academic research

Research questions, literature reviews, citations that actually exist.

looking for professionals

Studying

Study plans, notes and review that respect a student's real time.

looking for professionals

Personal organization

Doable weeks, lists that fit the calendar and priorities that don't get lost.

Tell us how you work with AIHow a pack is made

No coding needed. A short account already becomes a test case and shapes the pack for your field.

Questions

Before you install.

Is it safe?

The plugin is just text: instructions the agent reads. It doesn't run programs, install anything on your computer or ask for account access. The agent keeps the same permissions as before, and the skill tells it to stop and ask for your approval before sending, publishing, paying or deleting.

How much does it cost?

Nothing. orquestrar is open source under the MIT license. It uses the plan you already have in each app; plugin and skill availability depends on that plan.

Will my assistant get slow or ask too many questions?

It shouldn't. Simple questions still get direct answers, and the tests include a trap just for that. For bigger tasks, it asks three questions at most, each with a default: you can just say “go ahead”.

Does it work in other languages?

The skill is written in Portuguese, but the agent replies in the language you write in.

How do I remove it?

Uninstall orquestrar on the same plugins screen in your app. In a terminal: /plugin uninstall orquestrar@orquestrar (Claude Code) or codex plugin remove orquestrar@orquestrar (Codex).

Who makes it?

David Fiocchi, from Syntelix, with contributions open to anyone. Real usage reports beat any opinion: that's how the method improves.