# Roma Bridge Roma Bridge is a small app on your computer. It starts your agent when you hand it a task in Roma. This page says what it is, what it needs, how work reaches it, and what to watch for. ## What it is One Roma Bridge per computer. It keeps your local agents on the line with Roma: Claude Code, Codex, or a command of your own. It runs while your computer is awake, even with Roma closed. Your agent sets it up. Send your agent the sentence from Settings → Agents. At the end your agent runs one command, `npx -y roma-bridge pair`. Your browser opens a Roma page with a code. Check that your agent shows the same code. Press Approve. The first time, your computer asks once more whether it may take tasks from your account. Press Connect. The page must be open on the same computer. Approving on a phone or on another computer connects nothing. There is nothing to open afterwards. On a Mac, Roma Bridge starts by itself when you log in. The agent's page under Settings → Agents says whether Roma can reach it. Roma for Mac carries its own copy of Roma Bridge. It can install that copy for you from the Roma Bridge page under Settings → Agents. ## What you need A Roma account. Node 20.3 or newer. A browser on the same computer. Claude Code or Codex, installed on that computer and signed in to its own account. Or a command of your own. Roma Bridge runs the one you add; it does not sign in for you. ## How work reaches it Delegate an item to the agent, or @ the agent in the item's chat. Roma Bridge picks that up within a second and starts a session in the agent's folder. Without a choice, that is the folder Roma in your home folder. You change it on the agent's page. The agent reads the item, does the work, writes the result into the item's body and posts in its chat. Press Reply on its post to continue the same session. It remembers what it did there. ## Permissions Each agent's page has a setting, "Skip all permissions". It is on by default. On, the agent runs commands as you on this computer without asking. Claude Code calls this bypass permissions; Codex calls it full access. The folder is where a session starts, not a fence. A command can reach anything your account can. The other setting is "Edits only": file edits and Roma, no commands. For Codex this keeps writes inside the folder and runs commands in its own sandbox. The sandbox can reach the network. Your own Claude Code settings files still apply in both modes. A rule in `~/.claude/settings.json` or in the folder's `.claude/settings.json` holds here too. A short deny list is always on for Claude Code: no wiping a disk, no force push. ## Questions There is no permission prompt. A run has nobody to click one. The agent cannot use a question tool either. When it needs something from you, it asks in the item's chat. Your Reply there resumes the session with your answer. ## Traps - Your computer asleep means the agent is offline until it wakes. The agent reads Offline after one minute without a check-in. A task you hand it then waits until the computer is back. - Claude Code not signed in, or at its usage limit, means runs fail. The task's chat says so. - A run cut short is reported in the item's chat. - A run is stopped after 30 minutes. - Two agents on one folder can edit the same files at the same time. - The agent's Roma rights are your full rights. It acts as you on your whole workspace. - Mail and teammates' text in an item reach the agent as material, never as instructions. The agent is told so. Still, read what lands in a shared item before you delegate it. [Events](/developers/events) has the poll Roma Bridge rides on.