Product at Muti AI
Noah works on muti, the room where the coding agents on your machine meet the people you trust. He writes about sharing agents without sharing laptops, the approvals a human should never delegate, and what agents do beyond code.
Posts by Noah
6 min readComparison
Of every agent product launched in 2026, Alibaba's QoderWake is the closest in shape to muti. Both run on your own machine, and both let a group send it work, through a chat that has to be paired first. They split on whose agent it is, what it runs, and who approves when a colleague's request reaches a delete.
9 min readComparison
Tencent's WorkBuddy lets you drive the agent on your own computer from WeChat and WeCom. Alibaba's QwenWork merges desktop, cloud and DingTalk agents into one product. QoderWake runs a team of digital employees on your machine behind a single bot the whole group can mention. All three let colleagues use the agent. They answer how in very different ways.
5 min readPersonal agents
Manus Cue gives each personal agent an email address, a phone number, a wallet and a computer. That makes the agent reachable by anyone who has the number, which changes the permission question from "what may my agent do" to "what may a stranger make it do". What Manus 2.0 shipped, and what reachability asks of an agent.
6 min readComparison
Grok Bot gives every account one cloud computer and a team of bots that share it, running routines while your laptop is closed. muti gives the agent already on your laptop a room so other people can use it. Two answers to "teammates you can give real work to", and the security boundary each one draws.
7 min readComparison
Three weeks apart, Meta and OpenAI shipped always-on personal agents with a computer of their own, a memory and a cartoon face. Muse is built around isolation and shopping, dots around ChatGPT, Codex and the channels you already work in. The differences that matter, and the one thing neither does.
6 min readTrust
Muse, dots, Grok Bot, WorkBuddy and QoderWake have settled on almost the same three buttons for asking permission. They differ on where the approval is enforced, how long a grant lasts and what it covers. And every one of them assumes the person approving is the person who asked.
9 min readComparison
Meta Muse, OpenAI dots, Manus Cue, Grok Bot, Tencent WorkBuddy, Alibaba's QwenWork and QoderWake all shipped by September. Where each agent runs, what it can reach, how it asks permission, and the question almost none of them answer: who else can use it.
4 min readRuntimes
Nous Research shipped three ways to put Hermes Agent into Block's Buzz. There is a fourth shape worth knowing: leave Hermes on your own machine and give it a room, so the requests, approvals and audit line are the same as for Claude Code or Codex.
5 min readComparison
Coze 3.0 puts several people and several agents in one project. Doubao's work partner gives the agent an identity inside Feishu, where the company already works. Buzz and Kylon build a new workspace. Four routes to the same destination, and what each one says about who approves the agent's riskiest step.
5 min readComparison
Kylon is a furnished AI-native workspace: rooms, AI members with memory, spreadsheet apps, three thousand integrations and workflows, hosted for you. muti is one room around the agent already running on your machine. What each is for, and the one question that decides it.
5 min readComparison
Buzz is a self-hosted relay where humans and agents are equal members, each with their own keys. muti is a hosted room around an agent that never leaves its owner's machine. Both keep work out of a vendor's cloud. They disagree about what you should own and who has to approve.
7 min readComparison
A chat box, a workspace, or the machine you already work on. Three answers to the same question, with the products that picked each one: Block's Buzz, Kylon, Coze, Doubao in Feishu, Alook, Tutti and muti. What each shape gives you and what it costs.
9 min readCost
AI seat sprawl is the new unused SaaS licence. Sort everyone into producers, askers and ghosts, downgrade the askers, and give them a producer's agent to ask instead of a seat. A one-afternoon audit with a worksheet, before-and-after numbers, and the structural change that keeps the seats from coming back.
5 min readCost
Codex now courts marketers and ops people, but a Business seat, a laptop install and per-person permissions stand in the way. The cheaper shape: one developer's Codex on their machine, teammates ask from Slack or the web, the developer approves. Cost table, what the developer keeps, and where the ToS line is.
6 min readCost
Claude Team is 0 to 5 a seat with a two-seat minimum. Most small teams have one or two people who produce with Claude and several who only need its output. Count the producers, buy them seats, and give everyone else a room to ask in. With the arithmetic, the ToS line, and when Team is still the right call.
8 min readScenarios
GPT-6 Astra is built to operate a computer for hours, and it lives in Codex. The obvious way to use it is not one account per person but one machine running Astra, shared through a room with the family, friends or team you trust: they ask, it works there, you approve the risky steps and keep the record.
3 min readComparison
Claude Code Agent Teams let one person run a lead session that assigns tasks to teammate sessions. An agent room lets several people send work to one owner’s agent. One is about parallelism inside your work; the other is about who is allowed to ask. They stack.
4 min readTrust
One rule sits under every decision in muti: nothing is shared unless the owner shares it. The room boundary, the fact that a mention grants nothing, the credential wall, the four always-pause actions and instant revocation are all consequences of it. Here is the rule and what it forbids.
3 min readComparison
Dispatch relays your Telegram messages to the Claude Code on your computer. A room does the same job for a group of people with permissions, approvals and a record. Both keep the agent local. They differ on who is allowed to talk and what happens when the message says “delete”.
3 min readData
What people actually ask other people’s agents to do, from the first thirty days of muti’s private beta. The numbers are small and we say so. Roughly a third of requests came from someone other than the owner, most approvals were answered in under a minute, and the tool that asked most was Bash.
3 min readScenarios
The icon set changed again. Someone has to export it at three sizes, optimise it, and drop it where the site build reads from. An agent pinned to the assets folder does that in a minute, and a room lets the designer ask for it without a ticket.
3 min readConcepts
What happens to a request when the machine is asleep, the daemon is restarting, or the agent is busy. The inbox holds it, in order, with its id and its approval state, and the daemon works through it on return. Nothing is lost and nothing runs twice.
3 min readScenarios
Two runtimes, one room, one owner. Why you would put Codex and Claude Code side by side, how people address each, how a request can reference the other’s output, and what stays the same: permissions, approvals and the audit line.
3 min readScenarios
Read-only, time-boxed, revocable. How an agency or a freelancer lets a client ask the project agent questions, watch it work and receive deliverables, without a shared login, a screen share or a second subscription.
3 min readConcepts
Every task sent to a shared agent gets a public id like R20260906-07 and a small state machine: pending, approved, running, done, failed, cancelled, expired. The id is what people quote in chat, and resume means continuing under the same id after an interruption, not starting over.
3 min readComparison
Happy and Omnara put your Claude Code session on your phone. muti puts your agent in a room other people can enter. The first two mirror a terminal; the third changes who is allowed to talk to it. A comparison, and when you would want each.
3 min readScenarios
Five people, one founder’s Claude Code, one room. Monday’s deck, Tuesday’s campaign variants, Wednesday’s landing-page copy, Thursday’s report from the CRM export, Friday’s screenshots turned into bug reports. What they asked, what ran, what paused.
5 min readComparison
Three products put people and local coding agents in one room, and they answer three different questions. Alook gives agents identities so one person can direct many. Tutti makes agents collaborate across people’s sessions. muti lets you lend your agent to people you trust and approve what runs. A fair comparison.
3 min readTrust
Claude Code’s bypassPermissions mode removes every prompt. On your own throwaway sandbox that is a reasonable trade. The moment another person can send the request, it hands them your full local authority with nobody asking. Here is the line, and how to keep the speed without crossing it.
4 min readTrust
The day something goes wrong is the wrong day to start logging. Six fields every request line needs, five every approval line needs, what must never appear in the log, and why the owner, not the requester, holds it.
3 min readScenarios
A partner with a folder of videos to convert. A parent with ten years of photos to rename. A teenager who wants a spreadsheet turned into a chart. The computer in the study can do all of it, if the agent on it has a room the family can message.
3 min readComparison
Claude Code on the web runs in Anthropic’s cloud on a copy of your repository. A local agent runs on your machine and a room lets others reach it. Which one fits depends on whether you want the agent always on or always yours.
4 min readTrust
A precise list of the events the daemon on your machine sends to the room, field by field, and the categories it never reads or transmits. If you evaluate agent-sharing tools for a living, this is the page to check.
6 min readConcepts
Twenty terms that come up when people share the agent on their machine, from room, inbox, request, timeline, standing grant and takeover to relay, AI member, personal agent and agent computer. Each with a one-paragraph definition you can quote.
4 min readIM
Give an agent a Slack bot with its own name. DM it or mention it in a channel and the message is the task. Requests from other people wait for the owner, who approves with buttons in their own DM; pushes, deletes, spending and credential access never get an “always”. How muti’s Slack connector works and what it does differently from a relay.
5 min readIM
Let people in a Feishu (Lark) group send tasks to the Claude Code or Codex agent on your machine. Group-to-room binding, per-person identity codes, one task per thread, and approval cards for dangerous actions. The design of muti’s Feishu connector, and how to prepare now.
4 min readConcepts
To borrow an agent is to use someone else’s coding agent, on their machine and their subscription, from your own context, for as long as they allow. Here is what the word means, what the borrower gets and never gets, and how the owner stays in control.
4 min readTrust
plan, default, acceptEdits and bypassPermissions decide what Claude Code does without asking. Here is what each one allows, when each is appropriate, and how the picture changes the moment someone other than you is sending the requests.
4 min readComparison
Six ways to reach a Claude Code or Codex session from a phone, compared on who can use them, where the agent runs, what happens when the machine sleeps, and whether anyone approves the risky steps. Includes Anthropic’s Remote Control, Happy, Omnara, Dispatch and rooms.
5 min readComparison
Remote Control lets you continue your own Claude Code session from your phone. An agent room lets other people send work to that agent while you approve what runs. They solve different problems, and you may want both.
5 min readScenarios
The most ordinary use of a shared agent is sharing it with yourself. Send a screenshot from the train, let the machine on your desk do the work, and read the result when you arrive. How it works, what to set up, and what a good request from a phone looks like.
5 min readArchitecture
Where an agent runs decides who holds your code, your keys and your bill. A comparison of running agents on your own machine versus in someone else’s cloud, what each is good for, and why muti hosts the room but never the runtime.
5 min readScenarios
A coding agent is an operator for everything reachable from your machine’s terminal, which is most of your work. Meeting notes, decks, design exports, campaign copy, weekly reports. Here is how non-engineers can ask for that work from one room, and what to set up first.
5 min readTrust
Allowlists are not enough. Pushing to a shared branch, deleting, spending money and touching credentials should pause for a person every time, even when the agent is otherwise fully trusted. Here is why, and how to design approvals people will actually keep on.
5 min readHow-to
Your best agent runs on one machine that only you can reach. Here is how to let a co-founder, a teammate or a partner send it work, watch it run, and get the result, while your files, keys and account never leave your computer.
5 min readConcepts
An agent room is a shared space where a coding agent that runs on one person’s machine can be reached by a few people they trust. Here is what it is, what it isn’t, and why the room is the boundary.