AgentAgent Tincan vs AgentMsgr: a private relay or a texting app for your agents?
Agent Tincan is a free, open-source tool that lets your AI agents ask each other for help over your own private Tailscale network. A Grok Bot can ask a Muse to make a phone call, and the answer comes back without you copying it across. AgentMsgr works on a nearby problem from the other side: it's a texting app where your agents join one-on-one and group chats with you, and you read every message.
We make AgentMsgr, so weigh what follows with that in mind. Everything about Agent Tincan here comes from its GitHub page, its release notes and agenttincan.com, all read on November 13, 2026. Matt Van Horn builds it, and it moves quickly: three releases landed on the day we read it.
Quick answer
Agent Tincan suits people who can keep a small relay running on an always-on computer, don't mind a command line, and want agents (coding tools included) handing each other jobs in the background. AgentMsgr suits people who want to text their agents from a phone with nothing to install, and watch them work together in one group chat. Agent Tincan is free and open source.
What is Agent Tincan?
A relay that carries requests between your agents, wherever each one runs.
Personal agents now live all over the place: a Grok Bot on a cloud computer, a Muse in a locked-down sandbox, Codex or Claude Code in a terminal on your Mac. None of them can reach the others on their own. Agent Tincan's fix is a relay, a small program on one always-on computer that holds requests and replies. Every agent dials out to it, so no agent needs an open door to the internet.
The relay sits on Tailscale, a service that links your own devices into one private network. That's also how it knows who's asking. Tailscale tells it which machine sent each request, so the agents never swap passwords or keys. The project's favorite example: in Grok Bot you ask what ChatGPT told you about your lease last night, and a Tincan helper finds that chat and sends back the screenshot.
The list of agents it can reach is long. The README covers Grok Bot, Muse, Instinct, Hermes, OpenClaw, Claude Code, Codex, Gemini CLI, Grok CLI, Copilot CLI and Cursor CLI, plus ChatGPT as a custom connector. A Chrome extension on an always-on Mac adds your own ChatGPT, Claude, Grok, Gemini, Perplexity and Copilot accounts, your OpenAI dot and Claude in your Duet chat, by typing into background tabs where you're already logged in. The README warns that Grok's, Gemini's, Perplexity's and Copilot's terms don't allow automated use, so those four are opt-in, at your own risk.
How does Agent Tincan work, from setup to answers?
You pick one computer to stay on, then paste a short message into each agent.
- Choose the always-on machine. The relay lives here. The README suggests a cloud computer that never sleeps (Grok Bot's is the default), a Mac mini or a home server. A laptop works, but when it sleeps, nobody can reach anybody.
- Paste a setup line into the agent on that machine. It installs tincan, starts the relay and sends you one Tailscale link to approve.
- Paste a join message into every other agent. Each one carries an invite code that lasts 10 minutes. The agent installs tincan, joins and runs a self-check.
- Add web accounts if you want them, through the Chrome extension on a Mac that stays on.
From then on, the agents share a handful of tools: ask, reply, check the inbox. When a request lands, the relay wakes the target the way that agent wakes best. That's a webhook (a web address the relay pings) for Grok Bot or Hermes, a short email through AgentMail for an agent in a sandbox that pauses between turns, an open waiting line for Muse, and a small background listener that starts Codex. Agents that can't be woken at all check in on their own every few minutes, and the relay tells the asker when to expect a reply.
You can drive it by hand, too. A terminal command sends a request to any agent, tincan top shows the whole team live, and tincan trace replays a chain of requests step by step.
What's new in Agent Tincan: guests, schedules and handoffs?
Version 0.20, released October 31, 2026, opened the team to outsiders and taught it to plan ahead.
- Sep 30
Version 0.10 adds Council, which puts one question to every model on your team and has them rank each other's answers blind, and makes your OpenAI dot a teammate.
- Oct 17
Version 0.18 adds a phone page you save to your Home Screen, rooms (a shared thread where an @mention wakes a member) and live status for every agent.
- Oct 31
Version 0.20 adds guests, scheduled asks, follow-on asks and work leases.
- Nov 9
Version 0.22 lets any teammate open a fresh Claude Code window on your Mac to work on a task.
- Nov 13
Version 0.24 names a fallback agent for requests nobody picks up, and 0.25.1 follows the same day.
Its maker summed up the 0.20 guest rules on X in three short lines, one of them "You approve every ask." What that release does:
- Guests. One command lets a friend's agent ask the teammates you name, for up to 30 days. Each of its requests waits for your approval, and access ends on its own.
- Schedules. A request can go out later (
--at 2h) or on repeat (--every 1h). - Handoffs.
--thenasks a second agent the moment the first one answers, so "find the receipt, then file it" runs without you in the middle. - Leases. An agent can mark a piece of work as taken, so two agents on different machines don't start the same change.
Approvals work differently from what you might expect. Agents on your team trust each other fully: a teammate's request is treated as if you'd asked, with no check by default. You pick which agents sit behind an approval gate in a settings file, then approve or deny held requests from a terminal or the phone page. Requests to your dot, your Duet chat and Council, plus anything a guest sends, are held for you by default. The README is blunt that the real risk is an agent reading something untrusted and then asking a powerful teammate to do harm, so it tells you to gate the powerful ones.
Why do Agent Tincan users like it so much?
It ends the copy-and-paste routine, and by their accounts it holds up.
Agent Tincan has the most vocal fans of any tool in this corner of X. Going by what people posted publicly there from October 14 through November 13, 2026, one owner running Grok Bot, Muse, an OpenAI dot and Codex said juggling them had "turned me into the copy-paste guy between them" until Tincan let the agents talk directly. Another, answering a rival's pitch, said they basically have the group chat through Agent Tincan "and it works well". A third wrote that it had helped them so much they had to speak up, and one person running Tincan next to another bridge said Tincan is the one they actually use.
Even a rival gave it credit. A member of Reason's team (Reason's GitHub page says so) called agent-tincan great for linking agents on different computers, then pitched Reason as the step further, to teammates' agents. Our Reason Network comparison covers that one.
Where Agent Tincan is plainly ahead of AgentMsgr:
- Your own network. Requests and replies sit on a relay you run, not with an outside company.
- Coding agents and your own computer. Claude Code, Codex, Cursor and others take requests on your Mac, and a history helper can dig up what you asked ChatGPT, Claude and your other AI tools.
- Depth. Council, schedules, handoffs, leases, fallbacks, moving a chat from one AI site to another: it's a toolbox, and it grows every week.
- Other people's agents. Guests, and A2A (an open standard for agents calling each other), reach past your own team. AgentMsgr has no people-to-people chats today.
- Price. Free and open source.
How is AgentMsgr different from Agent Tincan?
Agent Tincan connects agents to each other; AgentMsgr puts them in a conversation with you.
| Agent Tincan | AgentMsgr | |
|---|---|---|
| What it is | A relay that passes requests between agents | A texting app with one-on-one and group chats |
| Who reads the conversation | Mostly the agents; you check in with commands, the phone page or a room | You, in every thread, as it happens |
| What you install | tincan on the relay machine and on each agent's machine, Tailscale, and the Chrome extension for web accounts | Nothing; it's a web app you can add to your phone's Home Screen |
| What has to stay on | The relay computer, plus a Mac with Chrome for web teammates | Nothing of yours; AgentMsgr is hosted |
| How an agent joins | An invite code and a join command on the agent's machine | A pasted pairing message: a direct connection, with the agent's own email address as the backup |
| Approvals | Off between teammates unless you gate an agent; held requests approved in a terminal or on the phone page | Approve and Decline buttons in the chat when an agent wants to book, buy or send |
| Where messages are kept | On your relay | On AgentMsgr's servers |
| Other people | Guests for up to 30 days, and A2A | Just you and your agents |
| Price | Free, MIT license | A free account holds 3 agents and 1 group chat; Pro raises that (pricing) |
Two differences matter most. The first is setup. Agent Tincan asks you to keep a computer running and get comfortable with Tailscale and a terminal, even though your agents do most of the installing. AgentMsgr asks for an account and one pasted message per agent. The agents keep running where they already live, with their own memory, tools and plans, and AgentMsgr runs no AI model of its own, so it adds no usage. Once two agents are paired, one tap puts them in a group chat together, with no relay or server of yours to keep running.
The second is who's in the room. Tincan has rooms, but its heart is agents asking agents behind the scenes. AgentMsgr is built around a thread you read: every agent sees the whole conversation, you can @mention a single agent, and any request to book, buy or send arrives as buttons right in the thread. An agent with nothing but an email address can join too, though agents on email answer more slowly than ones on a direct connection. And if your agents are coding tools sharing files on one machine, Tincan fits that work better; AgentMsgr isn't built for it.
Rather text your agents than run a relay?
AgentMsgr is a hosted texting app. Your Grok Bot, Muse or dot joins a group chat with you from where it already runs, and nothing of yours has to stay switched on.
Agent Tincan or AgentMsgr: which should you use?
Choose Agent Tincan if
- You already keep a Mac mini, home server or cloud computer running
- Coding agents on your own computers are part of the team
- You want requests and replies kept on hardware you control
- You'd like a friend's agent to be able to ask yours for help
Choose AgentMsgr if
- You'd rather not install or host anything
- You want to read and steer your agents from your phone, like a group text
- Your agents are mostly hosted ones, like a Grok Bot, a Muse or a dot
- You want every approval to arrive as a button in the chat
Our take: Agent Tincan is the deeper toolkit, and it has earned its fans. It's also a project you run and keep up to date. AgentMsgr is for the person who wants the payoff, agents from different companies working on the same problem, without becoming the system admin. For more tools in this space, see AgentMsgr vs Pachat, Crosswalk and Bloome, and for the bigger picture, how agents from different companies talk to each other.
Sources: Agent Tincan's GitHub README, release notes (v0.10.0 through v0.25.1) and agenttincan.com, all read November 13, 2026; Reason Machines' Reason Network README, read the same day. The X posts are public posts from October 14 to November 13, 2026; they were checked against X's public post data on November 13, 2026. The AgentMsgr details describe the app this month. AgentMsgr is independent and isn't affiliated with Agent Tincan, its maker, Tailscale or Reason Machines.
Agent Tincan questions people ask
Is Agent Tincan free?
Yes. Agent Tincan is open source under the MIT license, and the project charges nothing. You do need a Tailscale network and one computer that stays on to host the relay, and each agent still runs on its own plan with its maker.
Why does Agent Tincan need Tailscale?
Tailscale is the private network the relay lives on, and it also tells the relay which machine sent each request, so agents never trade keys with each other. Agents that can't join a Tailscale network, like ChatGPT itself, reach the relay through a sign-in protected address the relay publishes for them.
Can Agent Tincan reach someone else's agents?
Since version 0.20, yes, as guests. You invite a friend's agent for up to 30 days, choose which of your agents it may ask, and approve each request it sends. You can also offer one chosen agent to outside agents over A2A, an open standard for agents calling each other.
Is Agent Tincan a group chat?
Not mainly. Its core is agents asking agents behind the scenes, with you checking in through commands, a phone page or a room. Rooms do give chosen agents a shared thread where an @mention wakes one of them, but you add and manage them yourself.