Skip to content

Mob can be reached from Telegram, WeChat iLink and email. A channel is a plugin that receives messages and answers them through one shared Bridge (apps/mob/src/channel.ts), so commands, approvals, delivery records and replies behave the same everywhere.

  • answer runs a message as a command in the channel’s client or as a turn in the sender’s conversation, carries out the actions commands return (open a conversation, decide an approval), and keeps only what the sender should read.
  • deliver records each message under its ID in the deliveries table, so a message is answered and sent at most once and an interrupted send is reported, not repeated. /deliveries lists them. A channel that answers while it keeps receiving calls accept as each message arrives and finish once it answers; started hands it the turn’s stop as soon as the turn is admitted, so it can keep later messages in order.
  • A message for a conversation whose turn is still running is answered with Texts.busy instead of waiting, and a turn stopped by the sender ends with what it wrote so far and Texts.stopped. A decision answered that way is not taken; settled names an approval only once it no longer waits.
  • Texts is what a channel says in its own words: the help greeting, the approval prompt, the line before a decision’s reply, and the answers for stale approvals, failures and empty turns. Telegram speaks English and WeChat iLink Chinese.
  • A plugin keeps its transport and decides who is writing. There is no channel registry: a channel registers like any plugin, with an extension for its background run (see Extensions).
  • /help, /new, /model, /models, /compact, /usage, /approve, /deny, /grants, /revoke, /memories, /forget and /signin are offered to every channel (/list, /open, /stop, /schedules, /unschedule and /tasks are also offered to Telegram); owner-only commands refuse other senders.
  • The bridge replies with the text of the turn’s last step that wrote any, which concludes the turn. Text written between tool calls, reasoning and tool calls stay in the local conversation, where the web UI and TUI show the whole turn.
  • Texts.used, when set, adds one line naming the tools the turn called and how often, so the sender knows work happened and where to read it.
  • A channel’s turn sets ChatRequest.channel, and the model is told that the reader sees only its last text, so it ends with an answer that stands on its own. No second model call writes a summary.
  • Failures and approval requests are added after the conclusion. A turn whose last step writes nothing falls back to the last text written before it, such as the line before a tool call that asked for approval. A failed turn answers the sender with what went wrong.
  • Setup: enter the BotFather token and your numeric allowed user ID, then enable. Mob long-polls the Bot API and answers only that user’s private chat; no public listener is needed, and this machine must stay running.
  • The bot must not have an active webhook or another polling process. Mob reports the conflict and does not remove an existing webhook. /test telegram checks the bot’s identity and webhook state.
  • An update is acknowledged only after its receipt commits, and polling offsets survive restart. Restart marks in-flight processing failed and in-flight sends uncertain; neither is replayed. There is no resend: a new message is a new attempt.
  • Polling continues while a reply is written: each message is answered on its own, in the order it arrived. A message sent while the conversation’s turn runs gets a busy reply with a Stop button, so it is never queued behind the turn.
  • After a second, a reply in progress shows as a “Thinking…” draft (sendMessageDraft) with Telegram’s stop button, renewed until the reply is sent; where drafts are unavailable the bot shows typing instead. The stop button and /stop (Telegram only) stop the running turn, which ends with what it wrote and “Stopped.”
  • Replies are Markdown rendered to Telegram’s HTML subset: headings become bold lines, lists bulleted lines and tables aligned preformatted text, and only http, https and tg links are kept. Long replies are split between blocks (inside a block only when it alone is too long) to fit Telegram’s 4096 UTF-16 unit limit. Link previews are off. If Telegram rejects the HTML, the same text is sent plain.
  • Captions are read as the message’s text. Photos, files, voice and other media are not read; a message with only media is answered with a note saying so.
  • Approvals show Approve, Deny, Allow in conversation and Always allow buttons; a tap is handled as the /approve or /deny message it stands for, and its row is removed once that approval is decided or found stale. A tap while that conversation’s turn runs leaves the row in place. Approval buttons use the conversation saved with the approval, even after /new, /open or a restart, without changing the bot’s currently selected conversation; so does /approve ID or /deny ID typed with an approval’s ID. A conversation-scoped grant applies to that original conversation.
  • Notices: while the bot polls, it sends the owner the notices Mob raises, such as a coding agent’s task that waits for a decision, is done or failed, with approval buttons for what waits. A tap decides in the task’s conversation and the reply is how the agent went on, so the turns a decision from the bot started send no notice of their own. Notices raised while the bot is not polling are not sent later; the conversation keeps them.
  • Buttons stand in for typing commands. A tap on any button is handled as the command its data holds, as if the owner had typed it; a command longer than Telegram’s 64-byte limit rides on a short token kept in memory, and a tap on a token from before a restart answers that the button expired, with the menu.
    • /start and /help offer New conversation, Conversations, Model, Compact, Memories, Remembered approvals, Usage and More. More lists Schedules and Tasks when those plugins provide their commands.
    • Conversation, model, memory, approval, schedule and task lists have Previous and Next buttons, eight entries per page. They read full service data rather than the command completer’s twenty suggestions. The current conversation and effective model are marked ✓; models also offer following the default and show provider errors alongside the choices.
    • Memory and approval entries show their scope. Deleting a memory, cancelling a schedule, or clearing a whole list requires a confirmation tap. Completed mutations offer Back to list. Typed commands keep their existing behavior.
    • Task buttons open the task’s conversation. Notices also offer Open conversation and, when available, Schedules, alongside approval buttons.
    • A tap on a list’s choice removes the old choice buttons. Long callback commands still use expiring tokens. List menus query their data once; paging, confirmations and More stay local even when their internal commands are typed instead of tapped. They do not send prompts to the model. Multiline messages remain chat prompts even when they begin with a menu command.
  • The bot’s / menu comes from the command registry.
  • Connects natively through Tencent’s WeChat iLink HTTP API (ilinkai.weixin.qq.com) in TypeScript (apps/mob/src/wechat).
  • Sign-in is a QR code shown in the web UI (get_bot_qrcode, then get_qrcode_status until confirmed). After confirmation, click Save configuration to store the bot token and ID locally and enable polling; confirmation alone leaves them in the form, unsaved.
  • Messages arrive by long polling getupdates with its sync cursor; replies use sendmessage bound to the incoming context_token, with typing indicators.
  • Refuses senders not on its allowlist and treats only its first allowed user as the owner. Reads 批准 and 拒绝 as /approve and /deny, and accepts a bare reply when one request is pending.
  • Deliveries are recorded as wx:<bot>:<message>.
  • Messages with IDs must have saved receipts before the polling cursor advances. A receipt or cursor storage failure stops polling; resolve the storage problem and restart the plugin to resume. A recorded failed or uncertain send is not retried when the same message arrives again.
  • A nonzero API ret is an error even with HTTP 200. Polling errors use the normal retry delay; failed sends appear as failed-or-unknown in deliveries.

Answer the owner’s mail in the email plugin’s settings (email.channel.enabled) turns on answering mail. It names the owner’s address and the authserv-id the provider writes first in the Authentication-Results header of mail it receives (mx.google.com for Gmail; a received message’s original source shows yours). A background extension named email holds one IMAP connection, waits with IDLE (or polls where the server has none), and answers new INBOX mail as the email client.

  • Who is answered. Mail is answered only when its one sender is the owner and the topmost Authentication-Results header with the configured authserv-id says the owner’s domain passed DMARC, or signed it with DKIM. A From header alone is never trusted. This depends on the provider stripping forged Authentication-Results headers that claim its authserv-id, as providers following RFC 8601 do; a wrong authserv-id means no mail is answered rather than any mail being trusted.
  • No loops. Mail marked as sent by a program (Auto-Submitted, Precedence, List-Id, an empty Return-Path) is not answered, and Mob’s replies carry Auto-Submitted: auto-replied.
  • Threads. A thread is a conversation: the key is the first of its References, or its In-Reply-To, or the message’s own ID. Replies set In-Reply-To and References and keep the subject with Re:. Only what a reply adds is sent to the model: the text above the quoted message and the signature. A command is the first line.
  • Cursor. UIDVALIDITY and the last UID handled are saved per mailbox. The first run, or a new UIDVALIDITY, starts from the newest message, so old mail is never answered. New mail is handled oldest first, read with BODY.PEEK, and stays unread.
  • Not supported. Attachments are not read, and mail over 2 MiB is skipped without being downloaded.