harnsy

Claude Code in Telegram: your AI lead’s questions, approvals and the weekly result in one chat

Your agents work until they need you. Then they wait. You are at lunch, in a meeting, on a train, and a question with two possible answers sits on a screen you are not looking at. When you come back, an hour of work has not happened.

8 min readPavel Buchnev
In this article
  1. What you will have
  2. What you need
  3. How it works
  4. Steps
  5. What you still check
  6. Next step

This guide puts that question in your pocket.

What you will have

A Telegram chat with your project’s bot. When your AI lead puts a question to you with choices, it arrives there as buttons, and you press one. You can write to the lead from the same chat, by text or by voice. And once a week, the summary of a scheduled job (for example, the week’s post drafts) lands in the chat when the run ends.

What you need

  • harnsy Max. Telegram bots are part of harnsy Max; the free edition does not have them. See the prices.
  • A Telegram account, and a bot you create yourself with @BotFather. It takes one token; you paste it into harnsy; by design the token stays in harnsy and is not handed to your agents.
  • Claude Code, with a Claude plan that includes Claude Code, paid separately to Anthropic. Use your plan under Anthropic’s own terms. harnsy is not an AI model: it connects the AI programs you install and sign in to yourself.
  • A Claude Code lead, with the Claude Code hook that harnsy’s installer offers to add. Without the hook the lead’s questions stay in the dashboard.
  • Someone who can open a terminal to install harnsy and Claude Code. After that, you work in the browser and in Telegram.
  • For voice messages only: a speech-to-text provider you set up yourself (OpenAI, or your own Whisper server); your audio goes to that provider (not to harnsy). Voice is optional; text works without it.
  • Your computer on and harnsy running. The bot lives on your computer, not in a cloud: while harnsy is not running, it does not answer, and a message you write then is not delivered later.

How it works

Your AI lead puts a question to you with choices → harnsy sends it to your private chat with the bot, as buttons → you press one → the answer goes back to the lead and the work goes on. The first answer wins, whether it comes from Telegram or from the dashboard.

  1. AI leadPuts a question to you with choices.
  2. harnsySends it to your private chat with the bot, as buttons.
  3. YouPress one button in Telegram.
  4. AI leadGets your answer, and the work goes on.
One question, from your AI lead to your phone and back. The first answer wins: from Telegram or from the dashboard.

A schedule gives your copywriter a task once a week → the copywriter drafts → when the copywriter hands the work in, harnsy posts its summary to your private chat with the bot.

harnsy does the bot, the chat link, the schedule and the task. Claude Code does the thinking and the writing. Telegram carries the messages: what you and your team write in the chat passes through Telegram’s servers. Bot chats are not end-to-end encrypted, and Telegram’s own terms apply to them. If your questions or drafts hold customers’ data, send it there only if you may.

Steps

  1. Create the bot. In Telegram, open @BotFather, send /newbot, give the bot a name, and copy the token it gives you. Treat the token like a password.

  2. Add the bot to harnsy. Open Settings › Bots and press «Add a bot» › Telegram. Under «The bot is for» choose «A project», pick the project, paste the token into «Bot token» and press «Create the bot». harnsy checks the token with Telegram and stores it sealed; by design the dashboard does not show it again. One bot serves one project: through it you talk with that project’s lead. More about bots: Telegram and Discord bots.

  3. Make the chat yours. On the bot’s card press «Bind a person»: harnsy shows a code that works for a short time. Open your private chat with the bot in Telegram and send /start followed by that code. Send it in the private chat only: in a group the code is spent. Then, on the card, press «Make owner» and switch on «Receiving messages». The card lists four setup steps, «Bot works», «You are bound», «You are the owner» and «Receiving messages», and ticks each one as it is done. When all four are done, it folds them into one line: «Set up: you can write to agents». Until you are the owner and receiving is on, the bot answers nothing in your private chat. From then on, text you write there goes to your project’s AI lead.

    Shop botTelegram botExampleWorks

    Project: Pottery stall

    1. Bot workstoken checked
    2. You are boundyour name in Telegram
    3. Receiving messagesnot yet

    Almost ready. Receiving is off and you are not the owner: the bot answers nothing in your private chat.

    The bot’s card after you sent the code: two steps are ticked, two are yours. Press «Make owner», then «Turn on» beside «Receiving messages».
  4. Try a question. Ask your lead in the dashboard to do something that needs your choice. For example:

    Draft two headlines for this week’s post and ask me which one to use.

    When the lead puts this to you as a question with choices, it arrives in Telegram: «The lead of «your project» asks:», the question, and a button for each choice. Press one, and the message changes to «Answered from the chat:» with your answer. The same question waits in “Needs you” in the dashboard until one of the two is answered. This path needs a Claude Code lead and the Claude Code hook that harnsy’s installer offers to add. A question the lead simply writes in the dashboard chat stays there and is not sent to Telegram.

    Buttons come with one question of up to five choices. A longer one arrives as text, and you answer by replying to that message with a number or your own words. Up to three questions can be open in the chat at once; when a fourth comes, the oldest leaves the chat and waits for you in the dashboard.

    Shop botyour private chat with the botExample

    The question arrives

    The lead of «Pottery stall» asks:

    Which headline should this week’s post use?

    1. 1. New mugs, fresh from the kiln
    2. 2. Why one mug takes three days
    • New mugs, fresh from the kiln
    • Why one mug takes three days

    After you press a button

    The lead of «Pottery stall» asks:

    Which headline should this week’s post use?

    1. 1. New mugs, fresh from the kiln
    2. 2. Why one mug takes three days

    — Answered from the chat: Why one mug takes three days

    The lead’s question in your chat with the bot. One press answers it: the buttons go, and the message keeps your answer.
  5. Decide about permission requests. By default they stay in the dashboard, in “Needs you”. If you want them in the chat too, switch on «Permission asks to the owner’s chat» in Settings › Assistant. You get an «Allow» button only when the whole command is shown in the message; otherwise you can only deny it there and decide in the dashboard. There is no “always allow” in the chat.

  6. Send the weekly result to the chat. Ask your AI lead to set this one up. The schedule form in the dashboard cannot send a result to your private chat: its «Send the result» list holds only group chats you have allowed for the project. Your lead can set it. Write to it, in the dashboard or in your chat with the bot:

    Make a weekly schedule for the copywriter: every Monday, draft this week’s three posts in the owner’s voice and attach them to the task. Send the result of each run to my private chat with the bot.

    When the lead has made the schedule, you find it under «Schedules», on the copywriter role. Check that it is there before you go on; if it is not, ask the lead again. If your settings do not let leads set schedules, it arrives as a proposal, and it runs only after you turn it on. A schedule you made yourself in the form cannot be changed by the lead: let the lead make this one.

  7. Run it once now with the ▶ button on the schedule’s row («Run now»), and watch the chat. If no message comes, ask your lead where the schedule sends its result. When the copywriter hands the work in, harnsy sends a message that starts «harnsy: the result of schedule #…, item #…», then the schedule’s name and the summary, up to 40 lines in all; a longer one ends with «… the rest on item #…». There is no link in it: the full drafts stay attached to the task, as your instruction asked, and you open them in the dashboard. The summary comes before your AI lead has checked the work: treat it as a first look. One result of a schedule reaches your private chat at most once in 30 minutes: if you run it again sooner, the second result stays on the task only.

    Shop botyour private chat with the botExample

    harnsy: the result of schedule #3, item #24

    «Weekly posts»

    Three post drafts for this week are attached to the task:

    1. 1. New mugs, fresh from the kiln
    2. 2. Why one mug takes three days
    3. 3. Market day this Saturday
    The weekly result in your chat with the bot: harnsy’s line, the schedule’s name, then the copywriter’s summary. No link and no file: the drafts are on the task.
  8. Optional: talk instead of typing. With a speech-to-text provider set up, hold the microphone in the bot’s chat and say what you need. The lead gets the transcript with a note that it was transcribed from a voice message. By default a voice message can be up to 5 minutes long and 10 MB.

What you still check

  • The transcript of a voice message may have recognition mistakes. If a name or a number matters, type it.
  • By default publishing is yours. Your AI team does not publish for you: the summary arrives in your chat with the bot, the drafts wait on the task, and you post them yourself, to your Telegram channel or anywhere else.
  • Anyone who can use your Telegram account can answer in this chat. Protect the account (a passcode and two-step verification), and keep permission requests in the dashboard if you are not sure.
  • Who can write to the bot. By design only people you link can; a group needs its own link, and in a group the bot hears only a mention of it or a reply to its message.
  • A photo, a file or a sticker you send is not passed on to the lead; only its caption is.
  • A chat with a bot is not a private vault: do not send passwords or keys there. harnsy cuts secrets it recognizes out of an agent’s text before it is sent, and that is a safety net, not a guarantee.
  • harnsy adds no sandbox of its own: agents act with your user’s permissions. harnsy asks you once whether to start agents without Claude Code’s permission checks; if you agree, or if Claude Code on this computer already runs that way, your agents act without asking you first. The shield in the top bar shows which it is; only you can change it.

Next step

Want a job worth sending to the chat? Read Competitor monitoring for small business: a weekly report on new products and prices.

Telegram, BotFather, Claude Code, Claude, Anthropic and OpenAI are names of their owners’ products; harnsy is not affiliated with or endorsed by them.

Join the people who run it

Ask questions, see what others built.

Join the Discord
site-3flead · Claude Codeliveview only
❯ Plan #42 with the team.

Waiting for the breakdown from analyst-7a…

from analyst-7a through harnsy❯ #42 broken down: three acceptance criteria, including a retry after 24 h.

I’ll hand #42 to site-9a.

❯