OmniPotem is a web cockpit, installable on your phone, for developers who run coding agents (Claude Code, Codex) on their Macs. Follow every session, answer the agents’ questions, review their code before each push and get their reports.
Not another agent: Claude Code and Codex, as they are
OmniPotem does not reimplement an agent. Its daemon, which calls no model, starts the agents published by Anthropic and OpenAI, Claude Code and Codex, with the command set on your Mac and your own subscriptions: their updates apply as they are. Claude Code and Codex today, and soon other coding agents.
In the right place: each agent starts in the folder set for it on the Mac, for example its project’s repository, just as if you had started it by hand in a terminal. There it reads the project’s instructions (CLAUDE.md, AGENTS.md) and has the machine’s tools and access. Delegated work starts in the folder of the agent that receives it.
The daemon reads each Mac’s Claude Code and Codex quotas (5-hour and weekly windows) every 5 minutes. At 80% and 90%, an alert reaches Jarvis, and the Jarvis picker recommends the other agent when one is past 93% of its week or has used up its 5-hour window.
For delegated work, Jarvis picks the Mac and the agent with each machine’s remaining quotas, load and memory in view; splitting work by kind of task follows the instructions you give it. Nothing switches on its own: you change agent in one tap, or Jarvis decides.
The code stays on your machines. The service receives a review’s diff when you open it, masked and erased 7 days after the decision, and the session text, with secrets masked. The agents’ subscriptions and credentials stay on the Mac.
The agent submits a review: the diff, an explanation per file and the tests it ran. You approve the push of those commits or request changes; with the push guard installed on the Mac, its git hooks block any push you have not approved.
Without: Each agent is started by hand, in the right folder, on each Mac.
With: The daemon starts Claude Code or Codex in the folder set for each agent, on the Mac you choose, from the cockpit.
Without: You find out about the quota limit halfway through a piece of work.
With: Each Mac’s quotas are read every 5 minutes, with alerts at 80% and 90%, and a recommended agent.
Without: One terminal per agent, to check one by one.
With: Every session’s thread and the state of every delegated work, in one cockpit.
Without: An agent’s question waits until you are back at your desk.
With: It arrives as a notification on your phone; you answer in one tap.
Without: Each diff is read in the terminal.
With: A side-by-side review, an explanation per file, then the push approved by you.
Without: Nothing shows where each session stands.
With: Where each project stands, the state of the sessions and the quotas, on screen.
What OmniPotem is
Four pieces that work together: the cockpit, its phone app, a daemon on each Mac, and the coding agents that the daemon connects to the cockpit.
The phone app
The same cockpit, installed on your home screen as a web app. It notifies you of a new question, a review to open or a “To see” card.
The web cockpit
Where everything is followed: where each project stands, the agents’ questions, reviews before push, delegated work, session threads and your machines.
The daemon, on each Mac
A Node program, with no AI model, that runs on your Mac in cmux. It relays your messages to the sessions, sends their threads to the cockpit, follows delegated work and computes the diffs of reviews.
The coding agents
Claude Code and Codex, and soon other coding agents, started in cmux sessions on your Macs, with your own subscriptions. They talk to the cockpit through MCP tools: ask a question, submit a review, deliver a task. The main agent, Jarvis, starts and follows delegated work.
Who it is for
Developers who run several coding agents, on one Mac or more, and want to follow them and make decisions without staying at their desk.
Requirements
macOS, cmux, git, Node 22 or later, and Claude Code or Codex. FileVault turned on, required by default.
How it works
The path of a piece of work, from the machine to the report.
1Add a MacIn the cockpit, you add a machine: its token is shown only once. On the Mac, the setup script stores it in the Keychain, and the machine shows up online.
2Start an agent or hand over workYou start a session from the cockpit, write to an agent, or hand a task to Jarvis, which carries it out as delegated work in a session of its own.
3FollowEach session’s thread shows the messages and one line per action. Delegated work shows its state, its model and its sub-agents.
4DecideA question or a go card arrives with a notification. You answer in one tap; the agent is woken and reads your answer.
5Review before the pushThe agent submits a review. When you open it, the Mac computes and masks the diff; you allow the push of those commits or request changes. The push guard’s git hooks block any push that was not allowed.
6Get the reportThe agent delivers the task with a summary, its proofs, its checks and the risk. Delegated work files its report, and the Accomplishments page sums up the week.
One sequence, from question to report
On the phone, with made-up data: a question arrives, the go leaves, the review comes before the push, then the delegated work returns with its report.
1A decision waitingAn agent asks a question; it arrives in the cockpit with a notification.
2The goYou answer in one tap, and the answer goes back to the agent who asked.
3The review before the pushDiff, an explanation per file and the tests the agent reports: the push waits for your approval.
4The reportThe delegated work ends and returns with its report.
09:41
Cockpit
2 agents running
Decision neededapi
Run the migration on the test database?
GoNot now
Go sent
The answer goes back to Claude Code.
Claude Codeapp-mobile
Tests started on the form branch.
Review before push+48−12
4 filesTests passed
Allow the push
Delegated work done
Contact form: done
142 tests passed
Read the report
The features in detail
For each feature: what it does, how to use it, and an example. Examples and mock-ups use made-up data.
Decision needed
The project cockpit
What it does
The home screen brings together where each project stands (live, under way, waiting for you), the most urgent decision, tasks by project and today’s delegated work.
How to use it
Jarvis keeps each project’s status up to date. You filter by project, and the Accomplishments page sums up the week’s pushed reviews and closed tasks.
Example
The “app-mobile” project reads “Waits for you: choose the migration order”, and the matching question sits at the top of the home screen.
New question
Run the migration on the test database?
The phone app and its notifications
What it does
The cockpit installs on your home screen and sends notifications: a new question, a review request, a “To see” card and its reminders, an agent’s reply in a discussion.
How to use it
Add the cockpit to your home screen, then turn notifications on with the bell. Tapping a notification opens the card it is about.
Example
On the train, a question notification takes you straight to the card to answer.
Send a screenshot of the home screen.
142 passed, 0 failed.
Writing to the agents
What it does
Each session has its thread: your messages, the agent’s replies and one line per action. From there you write, attach up to 20 images, quote one of the agent’s messages or start a session.
How to use it
Open an agent’s thread, write, and paste a screenshot if needed: the daemon pastes your message into the session as soon as the prompt is free. Writing to an agent requires a two-factor code entered within the last 12 hours.
Example
From your phone: “Add the screenshot of the form to the report”, with the screenshot attached.
Decision needed
Run the migration on the test database?
GoNot now
Questions, go cards and “To see” cards
What it does
An agent asks a question with choices, an urgency and a due date. A go card asks for your approval of one specific action, showing the exact text that will be sent. “To see” cards point out what to look at, with reminders.
How to use it
You answer in one tap, with five seconds to cancel and an optional note, or you discuss it with the agent under the card first. “Not now” puts the decision off.
Example
Go card: “go run the migration on the test database”. “Give the go” sends that text to the agent, unchanged.
3 filesTests passed
Review before the push
What it does
The agent submits a review: a summary, one sentence per file, the tests it ran and the limits. The page shows the diff side by side or unified, an overview (architecture, user journeys, risks) and your “seen” marks.
How to use it
You comment on a line, then allow the push of those exact commits, for 7 days, or request changes. Opening the code and deciding require a code entered within the last 12 hours.
Example
A review of 4 files, +48 −12: you mark each file as seen, comment on one line, then allow the push.
Code
Prose
This file checks the address before the form is sent.
To fix
Prose translation and a second agent’s review
What it does
An agent rewrites a file in plain prose for readers who do not read code, on its own or annotated beside the code. An adversarial or simple review can be handed to another agent, Claude or Codex, which returns a verdict and comments on lines.
How to use it
“Translate into prose” on one file, or “Translate everything into prose”; “Ask for a review”, choosing the agent, the model and the effort. The prose is written in French for now, and kept by the service, never in the repository.
Example
Verdict “To fix”, with two comments on a race condition.
Delegated work done
The report is ready to read.
Delegated work
What it does
Jarvis starts a piece of work in a separate session, on one of your Macs. The cockpit shows its state, summary, model, background tasks and sub-agents live, and keeps its report.
How to use it
On a task, “Hand to Jarvis”: Jarvis accepts it and starts the work. The work’s questions are answered on its page, and its thread stays read-only.
Example
“contact-form”: working with 2 sub-agents, then done, with its report.
Several machines, and guests
What it does
Each Mac has its token, its agents and its permanent session. An @machine mention sends a message to one or more Macs, five at most. You can invite other people to a machine, with a role.
How to use it
Settings › Cockpit: add a Mac, create an agent’s token, set the mention, invite. Roles (observe, operate, pilot, answer) can be limited to some agents and to an end date; every action is logged.
Example
“@studio @laptop run the tests” goes to both Macs.
Claude Code84%
Week
5 h
Codex42%
Week
5 h
Quotas, usage and machine health
What it does
A gauge shows the Claude Code and Codex quotas (5-hour and weekly windows), orange from 80% and red above 90%; an alert reaches Jarvis at those thresholds. The Usage page brings together the month’s consumption and cost, and each Mac reports its load, memory and disk.
How to use it
The daemon reads the percentages and reset times on the Mac every 5 minutes, without any credentials. The Jarvis picker recommends the other agent past 93% of the week or once a 5-hour window is used up, and Jarvis sees each Mac’s quotas when placing delegated work.
Example
Claude Code at 95% for the week: the Jarvis picker recommends Codex, and you switch in one tap.
InterruptEsc
Handover
Note written, fresh session
Staying in control of sessions
What it does
Session threads are kept for 30 days. “Interrupt” sends a single Escape key to a busy agent, never Ctrl+C. For Jarvis’s long sessions, the daemon runs a handover: beyond 400,000 context tokens or after two compactions, Jarvis writes a handover note, and a fresh session picks it up.
How to use it
“Interrupt” sits in the thread, with five seconds to cancel; “Edit again” puts your last message back in the field so you can correct it.
Example
An agent heads for the wrong branch: “Interrupt”, then “Edit again” to make the instruction clearer.
What stays on your Macs, what goes through the service
The agents and your repositories stay on your machines. Here is what the daemon sends, and how long the service keeps it.
On your Macs
Your repositories and your code, apart from the diff of a review you open.
The agents, their subscriptions and their credentials: the cockpit only receives quota percentages.
The machine’s and the agents’ tokens, stored in the macOS Keychain.
The arguments and results of the commands the agents run, and the handover notes.
Sensitive files (.env, keys, certificates), never read for a review.
Through the service
Session text (messages and one line per action), with secrets masked on the Mac before sending, kept for 30 days.
The agents’ state and each session’s working folder.
A review’s diff, masked, sent when you open the review and erased 7 days after the decision (30 days if none is made).
Images attached to messages, kept as long as you choose: 7, 30 or 365 days, or with no limit.
Quota percentages, and each Mac’s load, memory and disk.
The service never sends a shell command to a Mac: only known command types, which the daemon carries out according to your local configuration.
Frequently asked questions
Which systems does OmniPotem run on?
The cockpit opens in a browser, on a computer or a phone. The daemon runs on macOS, in cmux, with git, Node 22 or later, and Claude Code or Codex.
Do I need my own Claude Code or Codex subscription?
Yes. The agents run on your Macs, with your accounts; the daemon uses no AI model. Today these are Claude Code and Codex, and soon other coding agents.
Is my code sent to the service?
Only a review’s diff, when you open the review: masked, and erased 7 days after the decision. Session text is sent too, with secrets masked. The repository stays on your Mac.
What happens if the service stops responding?
The agents keep working on your Macs. The push guard then refuses pushes that were not allowed, unless you configured it otherwise.
Can I share a machine?
Yes, through a single-use invitation valid for 7 days, with a role (observe, operate, pilot, answer) limited to some agents. You can remove access at any time.
How is access protected?
With two-factor authentication: writing to an agent, giving a go, opening or deciding a review and creating a token all require a code entered within the last 12 hours. “Sign out everywhere” closes every other session.
Which languages are available?
The interface is in English and French. It follows your browser’s language and can be switched at the top of this page or in the settings. Prose translations of reviews are written in French for now.
When does the service open?
No date has been announced. Everyone on the waiting list receives a message when it opens.
Waiting list
Leave your address to receive a message when the service opens.