Get LoomWatchGet it
Free and open source · runs on Mac and Windows · no code needed

Put AI agents to work as one team. They improve themselves.

A team of AI agents on your Mac or Windows PC, each thinking with the model you pick: Claude, GPT, Gemini or a free one, on the ChatGPT or Claude plan you already have, or one OpenRouter key. They pass work to each other, pause for your OK wherever you choose, keep a record of every step, and learn from a checker and from you.

Six friendly velvet figures holding one gold line.

Point at one and it waves. Click it to build one like it.

Step 1Agents

Build an agent once. Watch it work.

Give it a job, pick the model it thinks with, and choose what it may do without asking. Try it on a real request before any team relies on it. On the bench below, each key is a permission: drag it onto a hook to allow it, or keep it, and the agent asks you first.

One agent, every team
It's kept once, in My agents. Change it, and every team it's in uses it from that team's next run.
Your plan, or one key
Your ChatGPT or Claude plan, or any model with one OpenRouter key. Every run's receipt shows what it cost.
My agentsNew agentNot saved yet

Scouta rose square figure

Ready

Model
Its job pins its skills on the board
Its figure how you tell it apart
What it may do without asking the keys on its hooks

    Try it plays a recorded run in your browser. Nothing calls a model, and nothing you type leaves this page.

    Step 2Workflow

    Put your agents in a team. Check the plan in one sentence.

    Connect them in order, and put yourself wherever you want to check the work. LoomWatch reads the whole team back as one plain sentence before anything runs. Then drag in the folders, files and skills each agent needs.

    Or just describe it
    Describe the job, and Ask LoomWatch proposes a team on the canvas. Nothing changes until you press Apply.
    Step by step, or with a lead
    A pipeline runs its steps in order, with stops where you want them. Or a lead hands out work and asks its teammates questions.
    In a project
    Every team works in a project. What you write about it, its folders and its notes reach every agent in it.
    Research and write

    
              
    Step 3Chat

    Talk to the team. Nothing moves on until you decide.

    @team starts the work, @Writer asks one agent, and anything else is a note. Watch who has the work, approve it or send it back, and read the answer with its receipt.

    Talk while it works
    A note reaches the agent at work in its next turn; Send now hands it over at once.
    Every step on a timeline
    Blue while an agent works, gold where it worked. Drag back to replay any moment.
    Ask about any part
    In anything an agent made, mark some words, part of a picture, some cells, a slide, a page or a moment in a video, and ask an agent about just that.
    From any app, on a Mac
    LoomWatch waits in a black notch at the top middle of your screen. Point at it to see who is working, or press Control-Option-Space to message an agent.
    Research and writethe team's chatReady
      Try asking

      Starts the team · 3 steps

      This page plays a recorded run for each request, so a request can't be free text. A note can, and it never leaves this page.

      ThenIt improves

      Each agent improves itself. You see what it learned.

      When a checker in the team sends work back and the fix passes, the agent can keep a one-line lesson. What you say about its answers, and answers you love, teach it too. Its lessons are kept in My agents and go with it to every team.

      Only real feedback teaches it
      Work a checker sent back, what you say about an answer, or an answer you love. It doesn't grade itself.
      Take a lesson back
      Each one says where it came from. Undo it in the chat, or Forget it on the agent.
      Writerimproves as it works

      Support team

      You

      @team Sum up this week's customer emails.

      Wr
      Writer
      • 46 emails, most about late deliveries.no source
      • 12 people asked for a refund.no source
      Ch
      Checker

      Not met Each number links to its source.

      Wr
      Writer
      • 46 emails, most about late deliveries.inbox export
      • 12 people asked for a refund.refund log
      Ch
      Checker

      All met Each number links to its source.

      Writer learned: Put a link to the source next to each number.From the check

      Later, in the Launch team

      You

      @team Sum up the launch notes.

      Wr
      Writer
      • 12 November is the new launch day.launch plan
      • Still to do: the payment test and the French text.task board
      • Mai writes the press note, due Friday.meeting notes
      Ch
      Checker

      All met Each number links to its source.

      Writer learned: Start each bullet with its key words in bold.From your Love it

      A recorded example. Nothing calls a model.

      What it does

      Everything LoomWatch does, in the order you use it.

      Step 1

      Agents

      • Build an agent onceEvery team uses it as you last saved it.
      • A face you pickSix figures, six colours, two more to earn.
      • Makes real filesSlides, Word, Excel, charts and dashboards. PDFs and videos need a program it names.
      • Try it on its ownA real run before a team relies on it.
      • Asks before it actsWeb, files and commands wait for your OK.
      Step 2

      Workflow

      • Hire, or describe itPick saved agents, or describe the job.
      • Stops where you decideYou review before the work goes on.
      • Give each agent what it needsFolders, files, skills and Notion pages.
      • On a scheduleRuns by itself while LoomWatch is open.
      • Set up a projectAsk LoomWatch proposes it with its first team.
      • A room for each projectHome shows who works where, and who waits for you.
      Step 3

      Chat

      • One chat per team@team starts it, @Writer asks one agent.
      • Talk while it worksIt says what it's doing; your notes reach it.
      • Approve, or send it backYour note travels with the work.
      • Answers with receiptsFiles, sources and every step it took.
      • Ask about any partWords, pictures, cells, slides, pages or a moment in a video.
      Then

      It improves

      • Learns from a checkerWork a checker sends back, once fixed, can become a lesson.
      • Learns from youWhat you say about its answers, and answers you love.
      • Takes lessons to every teamIt's the same agent in each one.
      • Undo any lessonEach one says where it came from.
      Underneath
      • On your computerNo account, no cloud of its own.
      • Your plan, or one keyChatGPT, Claude, or any model on OpenRouter.
      • Every step recordedCalls, handovers, answers.
      • Free and open sourceUpdates in one click.

      Your teams and their history stay on your computer. Agents send their work to the company whose model you picked (OpenAI, Anthropic, or through OpenRouter), and nothing else leaves unless you turn on Send to Notion.

      Start

      One download, and your first agent is already waiting.

      It opens with your first agent asking what to make first. It asks which AI should answer only when you send: your ChatGPT plan, your Claude plan's credit or an OpenRouter key. Until then, look around as much as you like.

      Download for Mac macOS 11+
      Download for Windows Windows 10 or 11, 64-bit

      If Windows says it protected your PC, click More info, then Run anyway.

      This page is a sketch. This is the app.