Troubleshooting

Run node ~/.agentmingo/reporter.mjs status in a terminal. It says which AgentMingo the computer reports to, which repositories' threads report from it, how many reports are waiting to send, whether AgentMingo answered, the reporter's version, and the last few problems it noted.

A computer needs macOS or Linux, Node.js 18 or newer, and Claude Code. Check with node --version and claude --version. Native Windows isn't supported yet.

• “Node.js 18 or newer is needed”, or “node: command not found”. Install a current Node.js (from nodejs.org, or with a version manager such as nvm), open a new terminal, and run the command again. Claude Code's own installer doesn't include Node.js.
• “This setup command has expired … or was used already”. Each command works once, within 15 minutes. Create a new one on the Sessions page and paste it straight away.
• “curl: (6) Could not resolve host”, a timeout, or a proxy error. This computer can't reach AgentMingo. Try curl -fsS 'https://agentmingo.com/api/health?live=1': it should answer with "ok":true. Behind a company proxy or firewall, ask for agentmingo.com to be allowed.
• It doesn't run in PowerShell or the Command Prompt. The command is for macOS and Linux terminals. Native Windows isn't supported yet.
• “Viewers can't connect a computer”. A computer's threads move the tasks they work on, which needs the member role. Ask an admin of the organization.
• “You have … computers connected, as many as one person can”. Disconnect one you don't use any more on its organization's Sessions page, then run the command again.
• It finished, but says to run a claude mcp add command. Claude Code's command line wasn't found or didn't answer. Run the line it printed: it gives Claude the tools for workflow steps. Threads report without it.

• Only some repositories report. A thread reports only from a repository your organization connected (the GitHub App, or a Git remote under Settings → Integrations), or one you added yourself: run node ~/.agentmingo/reporter.mjs repos add in its folder. Threads anywhere else, and in folders that aren't git repositories with an origin remote, send nothing about themselves (a repository named like one of yours has just its name sent, once a day, to check it). node ~/.agentmingo/reporter.mjs status lists the repositories that report.
• Cursor's own agent. AgentMingo ignores Cursor's own agent (its chat and Agent panel) for now. Cursor runs Claude Code's hooks for it, and from reporter version 11 the reporter recognises those calls and sends nothing. An older reporter reports Cursor's chats as Claude Code threads until it updates (it updates itself unless you turned that off; then run a new setup command). node ~/.agentmingo/reporter.mjs status shows its version. Claude Code inside Cursor (its extension, or claude in Cursor's terminal) works and reports as usual.
• Sessions says it's waiting for your computer's first report. The setup command hasn't run there yet, or it stopped with an error: see above. A computer reports as soon as the command finishes.
• A session that was open before you ran the command. Open sessions pick up the hooks on their own when you send a message. If one doesn't, restart it.
• The repository was renamed or moved. If your clone's origin still has the old name, AgentMingo asks GitHub where it leads, and the thread reports once GitHub says it's your repository, shortly after its first message. To report at once, point the clone at the new name: git remote set-url origin <new address>.
• The computer reports to more than one organization. Each thread goes to the organization that lists its repository. Check that you're looking at that one.
• status says AgentMingo didn't answer, or asked the computer to wait. Reports wait on the computer and go when AgentMingo answers again. If it keeps saying so, AgentMingo may be down, or this computer can't reach it (see the curl check above).

• When you're told. Once a thread has waited 1 minute on a permission prompt or a question: in the app's bell, and wherever your Notifications page says. A shorter wait isn't told. The notice opens the thread's page, where you can answer a permission prompt if your computer turned that on (below).
• Where you're told. Your Notifications page in the app sets it, per kind: email, this browser, and chat (Slack, Discord, Google Chat, Microsoft Teams). Waiting threads don't send email unless you turn that on there.
• No browser notifications. Allow notifications for AgentMingo in the browser's settings, then turn them on again on the Notifications page. On an iPhone, add AgentMingo to the Home Screen first (Share → Add to Home Screen).
• No email. Look in spam. Confirm your email address first: an unconfirmed one gets no notice emails. Each email has a link that stops its kind; the Notifications page turns it back on.

• The ways to start one. In any thread of your coding agent (Claude Code today, in a terminal, VS Code or Cursor), say Work on WEB-12: the reporter's start_task tool makes the task's branch in a folder of its own next to your clone, and the agent asks you before it switches the thread there. Or run node ~/.agentmingo/reporter.mjs start WEB-12 in a terminal, from any folder: it makes the same folder and starts Claude Code in it.
• The agent made the folder but stayed where it was. Answer yes when it asks to switch: it then moves the thread with Claude Code's EnterWorktree, and the task's brief comes with your next message. If it can't switch, it gives you a link that opens a new terminal Claude Code session in that folder with the request filled in (press Enter), and the command to run instead. The tool needs reporter version 14 or newer: node ~/.agentmingo/reporter.mjs status shows yours.
• It says it doesn't know where your clone is. Run node ~/.agentmingo/reporter.mjs start WEB-12 once from inside your clone, or open the thread there: the reporter remembers each clone it sees.
• A .env file is missing in the task's folder. start copies only your clone's local .env files (.env, .env.local, .env.development, .env.development.local, .env.test and .env.test.local), and says which it left out: others, like .env.production, can hold production secrets that a thread in that folder could read. Copy one there yourself if a test needs it, or add --env-all to copy every one, in your own terminal (it won't run from Claude's own commands; --no-env copies none).
• The thread doesn't get its workflow's next step. From reporter version 14, a thread on a task's branch asks for its step before each message, so a workflow you start from its page (Follow it) begins with your next message.

• Turning it on. It's off unless you turn it on, on each computer: run the setup command with its box “Answer prompts and reply to threads from AgentMingo” ticked (it is by default), anywhere: a terminal, your editor, or a coding agent's thread you paste it into. Its one-time code turns it on, and install says so. A command made with the box unticked, even with --approvals added (its code carries the box), or an older command with --token instead of a code, connects the computer but leaves it off, and says so: run a new setup command from the Sessions page with the box ticked, or node ~/.agentmingo/reporter.mjs approvals on in a terminal of your own (it asks you to type yes, and won't run inside your coding agent; add --wait 2h for a longer wait than an hour). Only threads started after that wait; the ones open keep the hook they started with. node ~/.agentmingo/reporter.mjs status says whether it's on. What it trusts is on the Security page.
• “Answer this one in the thread”. The thread's page, its row on Home and Sessions, and the notice say why a prompt can't be answered from AgentMingo. Answering from AgentMingo is off on that computer (run a setup command with its box ticked there, then start a new session); its reporter is older than version 17 (it updates itself, or run the setup command again); a subagent asks; Claude Code didn't say the prompt was on screen within 20 seconds, so the computer left it to the terminal; the computer isn't checking; the session started before answering from AgentMingo was turned on (start a new one: open sessions keep the hook they started with); the computer couldn't reach AgentMingo in time (it tries for 15 seconds to open the prompt here, and says why in ~/.agentmingo/reporter.log); or it's a dialog Claude Code doesn't share with hooks (trusting a folder, say) or a headless run (claude -p). It shows only to the thread's own person, and only for a thread whose repository reports here.
• A question doesn't open on the thread's page. Questions and plans open here like any prompt, from reporter 17, when Claude Code shows them through its permission prompt. In a terminal the reporter waits only once Claude Code says the prompt is on screen: if it doesn't say so for a question, the page says to answer this one in the thread.
• No reply box on a waiting thread's page. The thread's page offers a reply (Continue, Yes, No, or your own words) to its own person, when the thread replied and waits for you, on a computer with this on and reporter 15 or newer, while its watcher listens: up to 8 hours after the thread's last reply (from reporter 18, while its task, pull request or workflow run is open, a week at most: below), while the computer is awake and Claude Code is running. For a few seconds after each turn, until its watcher first asks, the page says it isn't listening yet. The page says which of those is missing. An install from before reporter 15 listens under 20 minutes until its update raises the watch hook's timeout (node ~/.agentmingo/reporter.mjs status says), or until you run the setup command again. Sessions that were already open when the timeout was raised keep the one they started with (Claude Code may say their hooks changed): restart them to get the longer one.
• The thread's page says the thread isn't listening. After each turn, a thread listens for what you send from its page (a reply, more review rounds, its workflow's next step) for up to 8 hours, and only while its computer is on and awake. It stops when you type in the thread, and after the 8 hours (from reporter 18, a thread waiting on open work keeps listening past them: below); typing anything in the thread then makes it listen again once that turn ends. A thread started before its computer got 8-hour listening keeps the 20-minute limit it started with, even once the reporter updates (Claude Code loads a thread's hooks when the thread starts): its page says it stopped about 20 minutes after its last turn. Reopen it once (from your conversation history, or claude --resume) and it listens up to 8 hours after each turn from then on. Nothing can reach a thread that isn't listening until you type in it (a Claude Code limit), so type it in the thread.
• The thread's page says its computer's connection was removed. A thread listens through its computer's connection. Removing that connection stops it listening at once, and the page says when: Disconnect on the Sessions page, running the setup command again there (which replaces the connection), an admin revoking its token under Settings → API, or uninstalling the reporter. Connect that computer again (run a new setup command there: after a Disconnect its reporter has no connection, so typing alone does nothing), then type anything in the thread: once that turn ends it listens again, through the new connection.
• A thread waiting on open work stopped listening. From reporter 18, a thread that waits for you on work still open (its task isn't done or canceled, it has an open pull request, or its workflow run is going) keeps listening past its 8 hours: as they end, it takes one short turn (“Still listening for AgentMingo”), and that turn's end starts it listening again. That's about every 8 hours, for up to 7 days from its last turn, while its computer stays on and awake; the reply box stays open through that short turn. It needs replies on for that computer, and “Keep threads listening while their work is open” on under Notifications in AgentMingo (it is by default). Each short turn re-reads the whole conversation, so it uses tokens on your own Claude plan, more the longer the conversation; the thread's page lists them under “Kept listening”, with their tokens, which count in what the thread used. It stops at the end of the watch it's in once the work is done or closed, after the week, or when you turn the setting off, and the thread's page says so. It may also stop if its computer is asleep or off as a watch ends, restarts, or Claude Code is closed, or if the short turn fails (a usage limit, say): the page then says nothing listens, and can't tell which. It never keeps a thread listening that started before 8-hour listening (reopen it once). Typing anything in the thread makes it listen again once that turn ends.
• “Not picked up”. The computer stopped listening (it slept, went offline, or the 8 hours ran out), or the reply expired after 2 hours. Type it in the thread. If you typed in the thread before it was picked up, or the thread ended or went on without it, it isn't sent at all.
• “Your computer isn't checking right now”. It's asleep, offline, or Claude Code was closed, so an answer there would never arrive. Answer in the terminal.
• Only No is offered. The computer's reporter is older than version 17, which takes Yes only for some file edits. It updates itself with its next report; or run the setup command again there. Until then, answer the rest in the terminal.
• No “don't ask again” on a card. It's there when Claude Code suggested something to allow for the session (most commands and tools); when it didn't, Yes and No are the answers, as in the terminal. From AgentMingo it lasts the rest of that session only, never written to a settings file.
• The prompt left the terminal by itself. It was answered from AgentMingo. Claude Code says so in the terminal (“Allowed once from AgentMingo by you at …”), and the thread's page lists it under “Answered from AgentMingo”.
• Turning it off. node ~/.agentmingo/reporter.mjs approvals off. Prompts are then only reported, as before, and threads take no replies. A setup command made with its box unticked (--no-approvals) turns it off too.

The button on a task's page works on a Mac with the reporter's helper on: tick “Start threads on this computer from AgentMingo” when you create the setup command, or run node ~/.agentmingo/reporter.mjs start-from-web on in a terminal of your own there (it asks you to type yes). See what that trusts.

• It says “Waiting for …” and then that it wasn't picked up. The helper checks about every 10 seconds, only while the Mac is on, awake and logged in (the page says when it last checked in), and a start waits 2 minutes. Wake the Mac and try again. On the Mac, node ~/.agentmingo/reporter.mjs start-from-web status says whether the helper is loaded and when it last checked; to see what launchd thinks of it, launchctl print gui/$(id -u)/com.agentmingo.start-from-web. What the helper does goes to ~/.agentmingo/reporter.log (its lines start with “start-from-web:”), and what launchd catches to ~/.agentmingo/helper.log. node ~/.agentmingo/reporter.mjs start-from-web helper --once runs one round by hand and prints each step; while the helper runs, it only says so (one runs at a time), so read the log instead.
• “The helper that starts threads on … has stopped.” That Mac's threads still report, but its helper stopped checking in: its LaunchAgent was removed or unloaded (by a cleanup app, say). Run node ~/.agentmingo/reporter.mjs start-from-web on in a terminal of your own there, or run the setup command there again with the box ticked. The button comes back at the helper's first check. node ~/.agentmingo/reporter.mjs start-from-web status says whether it's loaded.
• “It doesn't know where your clone is yet.” The helper makes the task's folder next to a clone the reporter already knows. Start one thread inside that clone by hand once (or run node ~/.agentmingo/reporter.mjs start WEB-12 there); after that, the button works.
• “macOS didn't let Terminal into your clone's folder.” A clone in Documents, Desktop or Downloads needs Terminal to have access: allow it in System Settings → Privacy & Security → Files and Folders (or Full Disk Access), then try again.
• Terminal opened, but the thread doesn't show up. Look at that Terminal window: Claude Code may be asking something only the computer can answer, such as whether to trust the folder (below), or git may have stopped on a key that needs a passphrase. Answer it there; the thread reports once it starts.
• Claude Code asked whether to trust the folder. From reporter 21, a start marks your main clone of the task's repository (say ~/code/web, not the task's folder ~/code/web-WEB-12) as trusted in Claude Code right before it starts it (one key, hasTrustDialogAccepted, in ~/.claude.json or the file Claude Code keeps its state in; what that trusts), so Claude Code doesn't ask. That clone then stays trusted for every Claude Code session in it and its worktrees, on any branch, until you set the entry back to false, remove it, or uninstall. Claude Code still asks when that Mac's reporter is older than 21 (node ~/.agentmingo/reporter.mjs status shows its version; it updates itself after its next thread, or run the setup command there again), when Claude Code was saving that file the whole time (it holds a lock on it while it does), or when the file couldn't be read: ~/.claude.json (or the file Claude Code keeps its state in when you set CLAUDE_CONFIG_DIR or CLAUDE_CODE_CUSTOM_OAUTH_URL, or still have its older ~/.claude/.config.json) isn't there, is empty or isn't JSON, as on a Mac where Claude Code has never run. The Terminal window says which, and so does ~/.agentmingo/reporter.log (its lines start with “start-from-web:”). Answer it in that window once; the next start finds the clone trusted.
• “Starting threads from here needs reporter 19.” The reporter updates itself after its next thread. Or run the setup command again on that computer.
• Turning it off. node ~/.agentmingo/reporter.mjs start-from-web off on the Mac stops the helper and removes ~/Library/LaunchAgents/com.agentmingo.start-from-web.plist, for every organization there. Or press Turn off on your Account page, one row (one organization) at a time: at its next check the helper stops taking that organization's starts, and it stops once no organization on that Mac is left on. Turning it on again needs the Mac. Turning it off leaves the clones its starts marked trusted in Claude Code as they are (start-from-web off lists the ones still trusted): uninstall sets them back (below), or set each entry's hasTrustDialogAccepted back to false yourself.

• What it needs. The AgentMingo GitHub App on the repository, a review bot's comments or CI's check runs on the pull request, and review results on for the computer: tick the box when you create the setup command, or run node ~/.agentmingo/reporter.mjs reviews on. node ~/.agentmingo/reporter.mjs status says “review results on”. CI that reports only commit statuses, not check runs, isn't passed on as failures (its statuses only move the pull request's checklist along).
• Which reviewers' comments. Any review bot's inline comments (on a line of the change). CodeRabbit, Copilot code review, Greptile, Cursor Bugbot, Gemini Code Assist, Qodo, Sourcery are read in their own format: the title, the severity they give, and their suggested fix or prompt for agents. Another bot's inline comments are read as plain text, at medium severity. A reviewer's summary comment on the whole pull request isn't passed on, and neither are people's comments or replies.
• Which thread gets them. The thread working on the pull request's branch (or the one running its workflow). A thread on another branch doesn't, and neither does a pull request from a fork.
• When. With the thread's next message, or by waking it after it pushes, for up to 3 rounds a pull request by default. Each repository sets its own (0 to 10; 0 means results always wait for you) under Sessions → Review loop. After those, results wait for you: give the pull request more rounds (+1 round, +3 rounds on its thread's page), or say continue in its next message or a reply from its page. The watch that wakes it stops 18 minutes after the thread's last reply (8 hours on a workflow, or with replies on, from reporter 15); what arrives after that comes with the next message (from reporter version 12, that very message; before, the one after it). Waking uses Claude Code's asyncRewake hooks, and the woken turn runs on your own Claude plan.
• Main changes or ticket changes don't arrive. They ride along with review results, so they need review results on too. A main change also needs the GitHub App (it hears main's pushes) and the paths of files threads edit (“Let sessions know about each other”, on by default); a change to a ticket in Jira needs Jira connected.

The reporter updates itself when AgentMingo serves a new version: it checks the new one runs on your computer first, and keeps the one before. node ~/.agentmingo/reporter.mjs status shows the version and the last change.

To stop that, run node ~/.agentmingo/reporter.mjs auto-update off (and auto-update on to start again). Off, it stays at its version until you run a new setup command from the Sessions page, which also updates it by hand.

Run node ~/.agentmingo/reporter.mjs uninstall. It takes AgentMingo's hooks, permission rules and MCP server out of Claude Code, sets back to untrusted the clones its starts from AgentMingo marked trusted in Claude Code, and disconnects the computer. It leaves the ~/.agentmingo folder: delete it if you like. You can also disconnect a computer on the Sessions page; its reports are refused from then on.

In the app, use Help and feedback, at the bottom of the menu: a person reads every message and answers by email. Or write to hello@agentmingo.com. The output of node ~/.agentmingo/reporter.mjs status helps; read it first, since it lists your repositories' names.