Get early access

Agents in orbit.

Cosmo is a native Mac app where Claude Code, Codex and the models you have keys for work as one team, on the subscriptions you already pay for.

You are on the list. We will write to when early access opens.

We only email you about early access. Privacy Policy

See how it works
Tasks6 open tasks, 2 being worked on, 2 need you
All Tasks Launch week New Goal Start All
In Progress2
T-11High

Build the launch page

VVega

Vega is working…

T-13

Check copy against the brief

OOrion

Orion is checking the work…

Blocked1
T-10Needs you

Pick the hero image

OOrion
In Review1
T-12Needs you

Draft the launch post

LLyra
Done2
T-9

Outline the page

OOrion
T-8

Research the audience

LLyra

Try it: drag a card to another column.

Works with what you already use

  • Claude Code
  • Codex
  • OpenAI
  • Anthropic
  • Google Gemini
  • Mistral
  • OpenRouter
  • Ollama
  • LM Studio

How it works

From a goal to shipped work. Planned with AI, done by agents, approved by you.

  1. 01

    Describe the goal

    Say what should be achieved, by when, and what the agents must know. Plan with AI turns it into work for your agents, teams and workflows.

  2. 02

    Check the plan

    Every task gets an owner and the tasks it waits for. Nothing reaches the board until you create it.

  3. 03

    Agents pick it up

    Started tasks are picked up as soon as nothing blocks them. A lead hands out subtasks, and every step lands in the task's activity.

  4. 04

    You approve what matters

    Commands and connectors ask first. The sheet has no default button, so Return never approves by accident.

  5. 05

    They learn, with your yes

    After each task the agent reflects and proposes a lesson. It joins its instructions only when you approve it.

Features

Everything a team of agents needs. In one native Mac app.

Chat

Every engine, one conversation.

Pick Claude Code, Codex or any model you have a key for, message by message. When you switch, the new engine gets what it has not seen yet, and the conversation carries on.

  • Second opinion. Type /second and another engine reviews the last answer, right in the thread.
  • Files as cards. Whatever an answer made or changed appears as a card, marked New or Changed, with a preview.
  • Always know where words go. The window says it: sends to Anthropic, sends to OpenAI, or on this Mac.
OrionPlans and assembles, in Launch, Claude Code, opus, high effort, sends to Anthropic

Draft the launch plan for Friday and check the build still passes.

Claude Code

ReadPackage.swift
Bashscripts/test.sh
Writedocs/launch-plan.md

Tests pass (48 of 48). The plan is in launch-plan.md: three phases, owners, and a go/no-go check at 10:00.

Newlaunch-plan.mdMarkdown · 3 KB
ChangedCHANGELOG.mdMarkdown · 12 KB

Second opinion, Codex

Two gaps: no rollback step, and the 10:00 check has no owner.

Add both, give the check to Vega.

Claude Code

Editdocs/launch-plan.md

Added a rollback step after the deploy, and Vega now owns the 10:00 check. The plan is ready for review.

Try it: type a message, or / for commands. Here the app’s built-in demo engine answers; type tool, approve or fail.

Workflows

Describe a workflow. Watch it build itself.

Tell the Workflow Builder what you need, and it creates the team, its agents, the steps and the wiring. Or draw it yourself: fan out to several engines, then bring their answers together.

  • On a schedule, every few hours or every day, even with the window closed.
  • It asks first. A step that needs approval pauses until you answer.
Launch Brief3 agent steps, Every day at 09:00

Brief

Cosmo 2.0 ships Friday: the canvas builder, several accounts, skills for every engine.

V

Release notes

Vega · GPT-6 Sol

Write release notes from the changes merged since 1.9.

L

Announcement

Lyra · Mistral Large

Write the announcement for the blog, in our voice.

O

Final review

Orion · Claude Opus 5.5

Check both drafts against the brief and each other.

Launch kit

Collects the final result

70 %
Designing the team and the steps…Claude Code · 4 s

Try it: drag the steps, click one to inspect it, press Run. Arrow keys move a selected step.

Teams

A team that keeps working.

Give each agent a role, an engine and its own instructions. Members share a folder and see each other's work; the lead splits big tasks and puts the results together.

  • Heartbeats. Agents pick up their tasks on their own, wake on a timer within working hours, and propose work when there is none.
  • A usage guard. Near the plan's limit, agents hold back and say when it resets.
  • Review across vendors. Codex checks Claude Code's work and the other way round, before it reaches you.

Governance

Agents learn. You decide what.

Agents cannot retrain their model. They get better through lessons you approve, and every change to their instructions is kept, with the reason.

  • Lessons wait for your yes. Add one to the instructions, save it as a skill, or reject it with a reason the agent sees.
  • Versioned instructions. See what changed and why; undo one change or restore any version.
  • Scorecards and thumbs. Tasks done, sent back, turns per task and lessons approved, per agent.
Inbox3 things need you
Needs YouActivityScorecards

OOrion: Reflecting on T-13…

Lessons to approve 1

VVega learnedafter a task, 1 hour ago

Run the page build before reporting done.

Sent back once for a broken link.

Work to review 1

T-12Draft the launch postLLyra

620 words: what is new, who it is for, how to try it. Checked by Vega: clear and on brief.

OpenSend Back…Approve

History of Vega’s instructions

  • Lesson approved: Run the page build before reporting done.Today at 10:02
  • Edited by youOct 1, 2026 at 18:20
  • Lesson approved: Keep components in the shared folder.Sep 29, 2026 at 09:15

Build from the outline and the brief.

Ask before changing the copy.

Learned:

- Keep components in the shared folder.

- Run the page build before reporting done.

Undo This ChangeRestore This Version

Try it: add the lesson to Vega’s instructions, or reject it with a reason.

And everything around it. Small things, done properly.

Type / for everything

Switch engine, ask for a second opinion, attach files or choose a working folder without leaving the keyboard.

Every model, one picker

Each engine's own list, read from the installed version. Newer models it cannot run yet say Needs update, with an Update button.

Effort, all the way up

Claude Code and Codex each list their own effort levels; Cosmo passes the one you choose, from Low to the top.

Both plans at a glance

A ring beside the model picker shows the 5-hour and weekly limits, as the engines report them. No model call, no guessing.

Several accounts

Add a second Claude Code or Codex account. When one reaches its limit, carry on with the other in one click.

One skills library

Import .skill files or write your own. Each skill reaches Claude Code, Codex and API agents in the way each engine understands.

Connectors and plugins

Give the engines MCP servers and Claude Code plugins. Secrets stay in the Keychain, never on a command line.

Any model with a key

OpenAI, Anthropic, Gemini, Mistral, OpenRouter, Groq, xAI, DeepSeek, Together, or any OpenAI-compatible server. Ollama and LM Studio stay on your Mac.

Projects that share context

Group chats in a project with notes every chat knows. Each one hears what the others decided, whichever engine answers.

Privacy

No server. Your keys stay in your Keychain.

Servers behind the app
There are none. Cosmo runs on your Mac and talks only to the engines and providers you set up. There is no Cosmo account.
Engine logins
Claude Code and Codex sign in with their own login, in Terminal. Cosmo never reads, stores or relays their tokens.
API keys and secrets
Kept in your login Keychain, read at most once per launch, and sent only to the provider they belong to. Never on a command line.
Where your words go
Claude Code sends to Anthropic and Codex to OpenAI, exactly as in Terminal. The window always says which, or on this Mac. OpenAI Responses requests are sent with store: false.
Local models
With Ollama or LM Studio, everything stays on your Mac.
What Cosmo keeps
Your chats, teams, tasks and workflows, as files in Application Support on your Mac.

FAQ

Questions, answered.

Cosmo is in development. Anything not here will be answered at launch.

Do I need a subscription?

Only for the engines you want to use. Claude Code and Codex run on your own plans, API models on your own keys, and local models on nothing at all. A built-in demo engine works with no account.

Which models can I use?

Whatever your engines offer. Cosmo reads Claude Code's and Codex's own model lists, and adds the models of every API provider you set up: OpenAI, Anthropic, Google Gemini, Mistral, OpenRouter, Groq, xAI, DeepSeek, Together, a local Ollama or LM Studio, or any OpenAI-compatible server.

Can the agents change my files?

In a chat, Claude Code asks before it edits a file or runs a command, and Codex only reads. On the task board you can let agents edit files in their team's folder without asking; commands and connectors still ask.

API models answer in text only: they cannot read files or use tools.

Do you see my code or my keys?

No. Cosmo has no server of its own. Your messages go from your Mac straight to the engine or provider you chose, and your keys stay in your Keychain.

Can I use more than one account?

Yes. Claude Code and Codex can each have several accounts, each with its own login and usage ring. When one reaches its limit, Cosmo offers to continue on another; it never switches on its own.

What do I need to run it?

A Mac with macOS 26 or later. For Claude Code or Codex, the engine installed and signed in with your own account.

How do I get early access?

Join the waitlist with your email address, at the top or the bottom of this page. We will write to you when early access opens. Cosmo needs macOS 26.

Put your agents in orbit.

Be among the first to try Cosmo on your Mac.

You are on the list. We will write to when early access opens.

We only email you about early access. Privacy Policy