Agents

Every bot runs one coding agent on your computer. Codync drives agents through the Agent Client Protocol (ACP), so it can run any agent that speaks it: the ones it knows by name, everything in the official ACP registry, and any command you give it. Each bot picks its own agent, and different bots can use different ones.

Codync uses the agents with the accounts they're already signed in to on your computer, such as Claude Code with your Claude plan or Codex with your ChatGPT plan. Cursor's usage checks are the exception to reading usage locally; see Privacy.

If an agent's login stops working (for example, signing in somewhere else revoked it), the bot first tries your message again on a fresh start of the agent, which picks up a newer login on the computer. If that fails too, the chat shows the agent's sign-in card: Sign in opens the agent's sign-in options, and your message goes out again once you've signed in.

Usage limits

On desktop, open Settings > Usage to see the limits reported by your connected computers. On the phones it's State > Usage limits, and in the terminal UI press U. The host refreshes usage every five minutes, and GitHub Copilot's every 30.

Usage shows the agents your bots run. The button beside the computer's name opens a list of every agent that can report on that computer: use its switches to turn on one no bot uses yet or hide one you don't want to see, and drag an agent by its handle to change the order Usage lists them in. In the terminal UI, press s on the usage sheet to open that list, and J or K to move an agent. The choices belong to the computer: your other devices follow them. An agent that's on but has nothing to read says No reading yet, which usually means it's signed out.

An agent signed in to more than one account lists each login profile under its name, the default one first, with that account's own limits. Claude Code and Codex accounts report theirs; other agents' accounts say No reading yet. In the menu bar or tray menu, an agent with more than one account is a single row that opens every account's limits, and with more than one computer each computer opens its own agents. The notch island and the widgets show the account nearest its limit.

These agents can report their limits: Claude Code, Codex, Google Antigravity, Cursor, GitHub Copilot, Grok Build, Hermes Agent, Amp, Auggie, Junie, Kilo, Kimi Code, Kiro CLI, Mistral Vibe, oh-my-pi and OpenCode. Unified Cursor plans, including Super, show one Weekly allowance with its next reset. Older Cursor plans retain their monthly usage readings.

Cursor usage works with the CLI's saved sign-in file on Linux and Windows, or the Cursor desktop app's sign-in. On macOS, a CLI login stored only in Keychain needs the desktop app signed in too; CLI file storage also works. You do not need to install the Cursor desktop app on a Linux or Windows host just to report usage.

Each limit here shows how much of it is already used, such as 62% used, and its bar fills up as you use it, turning amber and then red near the limit. Some agents' own screens show what's left instead, so 99% used in Codync is 1% left there. Antigravity's Gemini limits are labeled Gemini 5h and Gemini 7d, next to its Claude ones.

Codex needs its CLI installed and signed in on the computer. The host reads current account limits without starting a conversation or using a model. If the login stops working or the current limits are unavailable, old session-file readings are not shown. Sign in again from Extensions or run codex login --device-auth on that computer.

If Claude Code runs on a token from claude setup-token (CLAUDE_CODE_OAUTH_TOKEN), Claude's limits show after a bot's first turn, and only the ones Claude reports with its answers.

Agents Codync knows by name

For these agents, Codync knows the command it looks for, how to install it and how to sign in. The agent must be installed on the computer before a bot can use it.

AgentCommand Codync looks forSign in from a terminal
Claude Codeclaudeclaude auth login
Codexcodexcodex login --device-auth
Cursorcursor-agent or agentcursor-agent login
Pipirun pi, then type /login
oh-my-piompomp login
OpenCodeopencodeopencode auth login
Grok Buildgrokgrok login --device-auth
Gemini CLIgeminirun gemini and sign in with Google
GitHub Copilotcopilotcopilot login
Qwen Codeqwenrun qwen, then type /auth
Hermes Agenthermeshermes acp --setup to choose a provider and model
goosegoosegoose configure
Kimi Codekimikimi login
Factory Droiddroidrun droid
Ampampamp login
Kilokilokilo auth login
Clineclinecline auth
Auggieauggieauggie login
Mistral Vibevibe-acp or vibevibe --setup
Kiro CLIkiro-clikiro-cli login
Devindevindevin auth login
Qoder CLIqodercliqodercli login
CodeBuddy Codecodebuddy or cbcrun codebuddy
WorkBuddythe CLI inside the WorkBuddy appsign in from Codync
MiniMax Codemcodemcode login
Juniejunierun junie
Google Antigravityagysign in from Codync with Log in with Google
Cortex Codecortexrun cortex and set up a connection
Poolsidepoolpool login
DeepSeek Harnessdshrun dsh web and save your DeepSeek API key under Settings → Models

A Pi bot gets Codync's tools (messages, memory, routines, connectors, managing the team) through pi's own MCP support, so it needs pi 0.99 or later. With an older pi the bot still runs, but without those tools until you update pi. If your pi mcp.json has a server with the same name as one of Codync's, such as memory, Codync's is called codync-memory for that bot, so the bot has both.

Signing in to Google Antigravity with agy in a terminal doesn't sign in the part Codync runs, so use the sign-in option in Codync's setup sheet for it.

On Windows, Antigravity unpacks itself into a _MEI… folder of about 130 MB in %TEMP% each time it starts, and removes it when it stops. Earlier versions of Codync ended it before it could, and a forced stop still does: updating Codync or restarting its host while Antigravity runs, signing out of Windows, or a crash. To get that space back, delete the _MEI… folders in %TEMP% that hold a google3 folder while no Antigravity bot is working.

WorkBuddy is separate from CodeBuddy Code. Codync runs the CLI bundled in the WorkBuddy app: WorkBuddy AI.app (from workbuddy.ai) or WorkBuddy.app (from codebuddy.cn) on macOS, or its install folder on Windows (%LOCALAPPDATA%\Programs\WorkBuddy AI or WorkBuddy). That CLI only ships WorkBuddy's lite build, so Codync starts it the way the app does. It runs on Node.js: if the computer has none, Codync offers to add its own. Codync does not install WorkBuddy, and there is no Linux build, so Codync on Linux doesn't list it. Sign in from Codync: the WorkBuddy app's login stays in the app, so this CLI asks again and opens a browser. A codebuddy command on your PATH stays CodeBuddy Code.

DeepSeek Harness runs its own dsh --profile acp mode; a dsh too old to have it (0.1.0) fails to start with the command that updates it. It has no sign-in command: it uses the DeepSeek API key saved in dsh web, or DEEPSEEK_API_KEY set in ~/.dsh/.env. After the host or the agent restarts, a DeepSeek Harness bot carries on in the same session.

Hermes Agent runs its own hermes acp mode, with the provider and model you pick in hermes acp --setup. On Windows, Codync also finds it in its install folders (%LOCALAPPDATA%\hermes\bin) without a PATH change. Hermes and Gemini CLI keep a memory of their own that every bot on them shares; see Memory.

Other agents from the ACP registry

Codync also lists every other agent in the ACP registry, about 40 more. These need no install: Codync fetches one the first time a bot uses it, through npx, uvx or a build for your platform, which it keeps in ~/.codync/agents. Codync refreshes its copy of the registry once a day.

Codync doesn't know these agents' own sign-in commands. If one needs an account, sign in to it on the computer first, or use a sign-in option the agent offers in its setup sheet.

A custom command

To run an agent that isn't listed, open the bot's settings, set Agent to Custom command and type the command that starts it in ACP mode, such as my-agent --acp. Enter a model ID by hand if you need one.

How Codync finds installed agents

The host looks for each agent's command on the same PATH your login shell uses, plus the usual install folders (Homebrew, ~/.local/bin, nvm, bun, volta, asdf, mise, pnpm and others). When an installed agent has its own ACP mode, Codync runs it directly. Otherwise it runs the agent's ACP adapter from the registry. OpenCode older than 2 is the exception: it keeps retrying a model that refused it, so a reply never comes, and Codync runs the registry's OpenCode instead, downloaded into ~/.codync/agents. That copy uses your OpenCode sign-in but keeps its sessions apart from yours. Update OpenCode to 2 or later and Codync runs yours again.

Opening Extensions checks again, so an agent you installed or signed in to outside Codync shows up without restarting anything.

Extensions: install and set up agents

Agents are set up from Extensions, which belongs to one computer:

  • Desktop: Extensions at the bottom of the sidebar. With several computers, each computer's heading in the sidebar has its own Extensions button instead.
  • iPhone: tap the connection status at the top of the Bots tab (such as 1 connected), choose Manage computers, and tap Extensions on the computer.
  • Terminal UI: press A.

The Agents section shows each agent's status: Installed, Not installed, Not signed in, Needs Node.js, or Sets up on first use (registry agents). Select an agent to open its setup sheet. It walks through up to three steps.

Install

For an agent Codync knows by name, Install runs the agent's official installer on the computer. Agents published only on npm (such as Gemini CLI, Cline, Auggie and Kilo) install with npm. The install runs in a terminal on the computer, and you watch it from the device you're on. Where Codync has no installer for your platform, the sheet tells you how to install the agent yourself.

Node.js

Some agents run through an adapter that needs Node.js, such as Claude Code, Codex and Pi. If the computer has no Node.js, Extensions shows Add Node.js and those agents read Needs Node.js. Adding it downloads an official Node.js into ~/.codync/node, used only by Codync's agents. Nothing outside ~/.codync changes, and a Node.js you install yourself always comes first. On Linux distributions that use musl, such as Alpine, install Node.js yourself.

Signing in

The Sign in step shows whether the agent is signed in on that computer, with Check again to ask again. Opening the sheet never downloads an agent: a registry agent that hasn't run on the computer yet shows Set up, which downloads it if needed and then checks its sign-in. If it isn't signed in, the sheet lists the ways to sign in:

  • Sign in (agents Codync knows by name): runs the agent's own sign-in command in a terminal on the computer. You type into it from your device, and links open on your device.
  • The agent's own options: each agent advertises its methods. Some run in a terminal, some open a browser on the computer, and some ask for API keys.
  • Browser sign-in: opens on the computer itself. If you're away from it, use Remote screen to finish there.
  • API keys: typed into a form and saved encrypted on that computer only. Codync gives them to the agent each time it starts, and the values never leave the computer. Get a key links to the provider's page, and Remove saved keys deletes them.

You can always sign in yourself in a terminal on the computer instead, using the commands in the table above. Sign in as the same user the host runs as.

For an agent Codync knows by name, the chat also tells you when it needs signing in: if the agent isn't signed in when you create the bot, or a turn fails because it signed out, the chat shows a card with Sign in. For a bot on another login profile, the card opens that login profile's sign-in, or its key form for an API key. Once you've signed in or saved a key, Codync sends your waiting message again. Other agents show their own error in the chat.

Login profiles

A login profile is one more login to an agent, so two bots can work at the same time on different ones. The agent's sheet lists them under Login profiles. The first is the default login profile: the one the computer was already signed in to, which bots use unless they pick another. There are two kinds:

  • An account: the agent's own sign-in, again. Claude Code, Codex, OpenCode, Pi, Google Antigravity, Devin and Gemini CLI can be signed in to more than one account. Choose Add account: the agent's sign-in follows, in a terminal on the computer for most agents and in a browser on the computer for Google Antigravity. The login profile then shows who it's signed in as, where the agent says. An account needs the agent installed on the computer.
  • An API key: for any agent. Choose Add API key and type the variables the agent reads its key from, such as ANTHROPIC_API_KEY, or a name of your own for a custom model (a config like OpenCode's "apiKey": "{env:MY_KEY}" reads MY_KEY). Import from login shell lists the variables your terminal sets, such as a key you export in ~/.zshrc; tick the ones to use and Codync copies their values then. Nothing from your shell reaches an agent unless you import it, and changing the variable in your shell later doesn't change the login profile: use Edit keys and import again.

What you can do with a login profile:

  • Name it: the pencil gives any login profile, the default one too, a name of your own, such as Work. An empty name goes back to who it's signed in as.
  • Remove it: the trash button signs an account out and removes it, or forgets an API key's variables. A login profile a bot still uses can't be removed, nor one whose agent is running or whose usage is being checked just then (try again in a moment), and the default one stays.
  • Pick a bot's login profile: in the bot's settings, under Agent, choose Login profile. Changing it restarts the bot's agent; the chat history stays.
  • Fall back to another one: a fallback can be the same agent on another of its login profiles, so the bot carries on there when the first one runs out of usage.
  • See its limits: Usage lists each account of an agent with its own limits.

In the terminal UI, the agent's sheet lists the login profiles under its other rows, and the keys in the sheet's footer add, sign in, rename, edit or remove the one under the cursor. In the bot's form, ←/→ on Login profile steps through the choices.

An account keeps only its sign-in to itself: your skills, rules, MCP servers and settings are shared with it, and with Claude Code, Codex, Pi, Antigravity and Gemini CLI a bot moved to another account carries on its conversation (OpenCode and Devin start a new one). Codync never stores an account's sign-in or uses it itself. An API key's values are saved encrypted on that computer only, given to the agent when it starts, and never shown again. On a Codync Cloud computer a provider's key is held outside the agent's reach, as its other saved keys are.

Models and options

In the bot's settings, under Agent:

  • Model: Default uses the agent's own default. The list shows the models the agent reports for the computer's sign-in; use the refresh button to load it again. If an agent doesn't report a model list, leave Default or type a model ID.
  • Reasoning effort: expand the Agent row to see the selected model’s supported levels. Default shows the agent’s reported level; selecting it removes this bot’s override. Each bot keeps its own choice across sessions, including when its model is Default, without changing the shared CLI settings. Available levels come from the agent, so they can differ by model.
  • Fast mode: a switch shown when the agent offers it for the model. Claude Code and Codex do. You can also type /fast in the chat.

On desktop and iPhone, the collapsed Agent row summarizes the harness, model, effort and Fast mode. Expand it to access all settings. Click the harness or model to search; selecting a model closes only the search list. The refresh control is beside the search field. In the terminal, use the Agent, Model and model-option fields with the keyboard.

Changing the model or its options starts a new agent session; the chat history stays. Switching the bot to a different agent starts it on that agent's default model.

The agent for a new bot

There's no separate default-agent setting. A new bot gets Claude Code when Claude Code can run on the computer. Otherwise it gets the first agent Codync knows by name that can. A registry agent is never picked on its own. Change the agent in the bot's settings after you create it.

When an agent runs out

A bot can have fallback agents that carry on the same turn when its own agent is out of usage, rate limited, signed out or won't start. See Agent fallbacks.