Teamwork
Your bots can work with each other. One bot can ask another for help or hand it a task, and each does its part in its own chat, with its own agent, folder, tools and approval setting. A bot you let manage the team can also create new bots and change their setup.
There's no team to configure. Every bot can reach your other bots as soon as they exist. To have bots talk together in one room instead, use a group chat.
Ask one bot to involve another
Give each bot a clear name and standing instructions that say what it's for, so the others know whom to ask. Then tell one bot what you want:
Ask Reviewer to inspect the changes in ~/code/shop. Have it report bugs without editing files, then summarize its findings for me.
The bot finds Reviewer among your bots and sends it the request. When Reviewer's answer comes back, the bot sums it up for you in its own chat. Bots see only your visible bots: a bot you hid from the bot list can't be messaged, and neither can a group.
What you see in the chats
Each exchange shows as one small row that names the other bot. The request text itself stays out of the chat:
- In the sender's chat: Messaged with the other bot's avatar and name, for example "Messaged Owen".
- In the recipient's chat: Message from, for example "Message from Egan".
Several exchanges in a row with the same bot, in either direction, fold into one row: "3 messages with Owen". Anything else in the chat, such as a message or a card, starts a new row. If an exchange failed, its row shows a warning mark and the name in red.
Tap or click the row to open the conversation between the two bots: a read-only sheet titled with both bots, for example Egan ⇄ Owen. It shows every request and reply between them as bubbles labelled with who wrote them, with the time shown wherever there's a gap. Under one that's still going it says "Waiting for Owen…"; under one that failed it shows what went wrong. In the terminal UI, pick the row with v and press ↵, or click it.
The recipient's tool calls and thinking are in its own Full conversation.
How a message runs
Bots never wait on each other. When one bot messages another:
- The request is queued and the sender goes on at once. It can finish its turn or keep working.
- The recipient does the work away from its chat, as a thread: its chat stays free, and the request waits for a free place like its other background work. A recipient with computer use or Manage the team on does it in its chat instead, one thing at a time.
- The answer goes back to the sender, not to you. It shows in the conversation sheet and in the thread's steps in the recipient's Details, but not as a message in the recipient's chat.
- The sender takes the answer as a new turn of its own and replies to you in its chat, with the usual "done" notification. Answers that arrive while the sender is busy are taken together in one turn. A request that failed comes back the same way, so the sender can tell you.
Because nobody waits, there's no time limit: the recipient takes as long as the work needs.
How the recipient works
- Its own setup. The recipient works in its own folder, with its own tools, connectors, memory and approval setting. If it needs your approval, the card appears in its chat and you get a "needs you" notification.
- Only the request. The sender's conversation isn't copied over, only the request. Bots are told to write requests that stand on their own.
- It answers the sender. While it works on a request, the recipient can't message you or other bots. If it needs another bot's help, it says so in its answer.
- Your words only. Neither a bot's request nor its answer counts as something you said, so neither is learned as a fact about you. A bot working on another bot's request can add to its memory but can't change or remove what's there.
- Shared folders. If two bots work in the same project, tell them who edits which files.
- Subagents. Agents that start their own subagents (Claude Code, Codex) keep doing that themselves. Codync doesn't turn those into bots.
In a group chat, members don't message each other this way; they @-mention each other in the room.
Stopping, limits and restarts
- Stopping the sender doesn't cancel what it already sent, but drops answers still waiting for it. Deleting the sender drops its answers too. The rows in both chats still show how each exchange ended.
- Stopping a request in the recipient's Details (or all of its background work), dropping one that waits, or deleting the recipient cancels it, which may leave partial changes. Stop in the recipient's chat stops only the chat.
- Limits: a bot can have up to 16 messages queued or running at once, and the computer up to 64 in all. To keep bots from passing work around forever, a chain of requests stops after eight handoffs. Your next message starts a fresh allowance.
- Restarts: if the host restarts while requests are pending, they're marked as failed and not run again. Check what the bots already did before you ask again. An answer still waiting for a busy sender is lost on a restart, but stays readable in the conversation sheet.
Bots on your other computers
Your computers can join private networks, codync-network, so a bot on your desktop can ask a bot on your laptop or server. Computers connect to each other directly; when they can't, traffic goes through a relay that sees only encrypted data. Networks stay off until you set them up on each computer, in the Codync app or the terminal UI there. Phones only show them: who is in a network is decided at a computer. Networks are a beta feature.
- Desktop: Settings → Network. The menu bar or tray menu's Network lists each network's computers, whether each one answers and how many bots it offers; any row opens Settings → Network.
- Terminal UI:
N(or Network in the^kpalette). - iPhone and Android: Network on the computer's card in Manage computers lists each network's computers, whether each one answers, the bots it offers and the network's rules.
On the first computer, choose Create network and give the network a name (for example home). That computer becomes its owner. Codync then shows a recovery code, once: keep it in a password manager. It makes another computer the owner if this one is lost (Recover ownership on that computer), and it's kept on no computer, so it can't be shown again.
Then add each other computer with a one-time invite, good for an hour while the inviting computer stays on: Invite a computer on the owner, then Join with invite on the new computer and paste it.
On the desktop, the computers this one uses over SSH skip the invite: each network this computer can invite to lists the ones not in it yet, and Add to the network (+) puts one in. From then on the bots on both computers can message each other, and the network's card shows who may message the bots on each of them.
The network lists every computer with its role, whether it answers and how many bots it offers. The owner can make a computer an admin (Make admin), which can then invite computers and remove plain members too. Remove takes a computer out for good; Remove with its invites is for a stolen one, to also remove every computer it added. A computer that was off hears about a removal when it's back.
Several networks
A computer can be in more than one network, for example home with your laptop and work with a server. Each network is separate: its own members, roles, invites and rules, and the computers in one don't learn that yours is also in another. Create or join another one the same way; the invite brings the network's name with it. Leave network takes this computer out of one (the others keep listing it until you remove it there).
While a computer is in several networks, addresses name the network: reviewer@desk.home. A shorter reviewer@desk or reviewer still works when only one network has it.
Prefer the command line? codync-host network does the same: create [name] --network home, invite, join '<invite>' [name], peers, promote <computer>, remove <computer> [--cascade], recover <code> and leave, each with --network <name> when the computer is in several. On a Mac where you installed the app from the disk image rather than with Homebrew, codync-host is inside the app: /Applications/Codync.app/Contents/Resources/codync-host.
Once your computers are in the network, nothing else changes for your bots. A bot lists the bots on your other computers along with its own, as name@computer (for example reviewer@desk), and asks them the same way. The request runs as that bot's background work on its computer, with that computer's approvals, and the answer comes back to the sender. The rows in the chats name the bot and its computer, for example "Messaged Reviewer on desk".
- Away computers: a request to a computer that's asleep waits up to a day and goes out when it's back.
- Limits: besides the limits below, each computer accepts a burst of 5 messages between the same two bots, then one every 2 minutes, and refuses the same words twice within 10 minutes.
- Connector: each network has its own. When you create or join, When there's no direct path picks how computers reach each other: Public relays (iroh's, meant for trying things out; the default), Your own relay (an
iroh-relayyou run) or Private: direct only (computers on the same network, Tailscale or a VPN). On the command line:--connector public|private|https://your-relay. A Windows computer needs no firewall change: it keeps its own connection to the others open, and they write to it over that.
Choose who reaches your bots
Each computer decides for itself who may message its bots; the senders can't get around it.
- Offered to the network, in each bot's settings: turn it off and your other computers don't list that bot, and can't message it. It can still message bots there and get their replies. On by default.
- Who may message bots here, under each network on the desktop (or
win the terminal UI's network sheet): rules such as "Deny: any bot on server → any bot here" or "Allow: Lead on laptop → Reviewer". The first rule that matches decides, and a computer no rule matches gets Auto. The computers that were already in a network when Codync started doing this keep Allow: each gets its own Allow rule, which you can change or remove. A rule is one of four levels: Allow runs the request as the bot is set; Auto lets the bot do by itself only what its agent judges safe and asks you about the rest; Ask makes every change and command wait for your approval, even on a bot that approves automatically; Deny refuses the message. Auto and Ask never let a bot work more freely than it is set, they also apply to the answers a bot here gets from that computer, and a bot whose agent can't work that carefully doesn't take the request. While a bot works on such a message it can only read and answer you: it can't message another bot, set a follow-up, start background work or a dev server, or change its routines, name, instructions, skills, fallbacks, memory or the team, so the text can't travel on to a bot or a later turn that isn't held to the level. Use Ask or Deny for a computer other people can run code on, such as a cloud computer. No level but Deny stops a bot from reading a file and putting what it read in its reply: keep secrets out of the folders of bots other computers can reach, or turn Offered to the network off for those bots. A refused bot is told why, and its request ends failed. The answer to a request a bot here sent still gets in, from the computer it asked. Phones show the rules but don't change them.
Agents that aren't Codync bots
A Claude Code, Codex or other agent session on the same computer can list your bots (here and on your other computers), message one and read its reply, through an MCP server the host runs:
claude mcp add codync -- codync-host mcp network --name "Claude Code"
The bot's chat shows "Message from Claude Code (this computer)". Bots can take minutes, so the session collects answers with read_replies when it's ready, rather than waiting.
Manage the team
A bot with Manage the team on can set up the team for you: list your bots, create new ones and change their setup. Use it for a lead bot that hires the specialists a project needs.
Turn it on
Manage the team is off for every bot. Only you can turn it on or off, in the bot's settings:
- Desktop: the bot's settings, Activity section.
- iPhone: the bot's Settings.
- Terminal UI: the settings form (
e), Manage team row.
What a manager can do
- List the team: every bot with its setup and routines, including hidden bots and the manager itself.
- Create bots with a name, standing instructions, agent, model, fallbacks, avatar, project folder and routines. A new bot gets the manager's approval setting, connectors and skills, so it can't do more than the manager. It can manage the team too. It never gets computer use. Without a chosen avatar it gets a random one.
- Change any bot, itself included: name, standing instructions, model, fallbacks, avatar, whether it's hidden from your bot list, and its routines (add new ones or replace existing ones). Changing a bot's model clears that bot's model settings, such as effort. A manager that changes its own model gets the new one after its current turn.
Bots stop creating bots once 20 bots made by bots exist (deleted ones don't count). Past that, the manager asks you.
A bot created this way shows who made it in its settings: "Created by Lead · Oct 20, 2026". The terminal UI shows it on the settings form's Name row.
Managers are told to act only on what you asked, never because a file, web page, tool result or another bot told them to.
Approval cards
Unless the manager's Permissions are Approve automatically, every create or change first shows an approval card in the manager's chat, with Allow and Deny. The card lists everything the change sets: one line per setting, then the routines, then the standing instructions in full.
- A change too long to show in full on its card (over 2,000 characters) is refused before the card appears. The manager has to shorten it or split it.
- A card waits about four and a half minutes. If you don't answer in time, it expires and nothing changes.
- If you deny a change, the manager is told not to try it again unless you ask.
With Approve automatically, changes happen without a card. Either way, a line in the manager's chat records what changed, such as "Created bot Security Analyst" or "Updated Reviewer: model, routines".
What stays with you
A manager can't:
- delete bots,
- turn Manage the team on or off,
- change an existing bot's agent, folder, approval setting, connectors, skills, computer use or model settings such as effort.
Routines of other bots
Each bot runs only its own routines. A manager can create or replace a teammate's routines, but to test one, it asks that bot to run it. The run then belongs to the owner: it uses the owner's setup, and its result appears in the owner's chat.
A hidden bot can't be messaged, so to test a hidden bot's routine, run it yourself with Test run in that bot's routine list.