DEV Community

Subhendu Das
Subhendu Das

Posted on • Originally published at bsky.app

Why Every Developer Needs an AI Assistant on Their Messaging Platform

Your AI assistant lives in a browser tab. Your actual life happens in a messaging app.

That gap is small enough to ignore most days and expensive on the days that matter. The idea arrives on a walk, the production question arrives at dinner, the "what was that flag called again" arrives in the ten seconds before your laptop wakes up. By the time you are at a keyboard with the right tab open, the thought has either been written down badly or lost.

Telechat closes the gap by putting Claude where you already are: Telegram, WhatsApp, Slack, Discord, and a local web chat — all at once, from a single Python process on your own machine.

One process, five surfaces, one memory

The multi-platform part is not a checkbox. It is the thing that makes the rest work.

Most bots are one bot per platform, which means one conversation history per platform, which means the context you built up on your phone is not there when you switch to the desktop. Telechat runs every adapter in one process against one SQLite file, so sessions, memories and history are genuinely shared. You can start a thought on Telegram at a bus stop, keep going in the local web chat when you sit down, and it is the same conversation — not a second bot that has never met you.

Set BOT_MODE=all and every surface comes up together. Set it to a comma-separated subset and only those do.

There is no webhook, no tunnel, no public URL, and no Telechat server anywhere in the path. Your conversations go from your messenger to your machine to Anthropic, and nowhere else. For a tool you will end up telling quite a lot about your work, "self-hosted" is a meaningfully different security posture from "we promise not to look".

Getting to a working bot

npm install -g telechat
telechat init
Enter fullscreen mode Exit fullscreen mode

telechat init is an interactive setup that uses the Claude CLI to walk you through each platform — it opens the right pages, grabs your tokens, validates them, and writes your config. There is also telechat setup for a plain step-by-step wizard with no Claude CLI required.

If you just want to see whether any of this is worth it:

telechat web
Enter fullscreen mode Exit fullscreen mode

That starts the local web chat on 127.0.0.1:8585. No bot token, no messenger account, nothing written to disk. It needs only a way to reach Claude — the Claude Code CLI, or an ANTHROPIC_API_KEY — and it tells you which one is missing if neither is there.

Two ways to reach Claude, and the choice matters for cost:

  • CLI mode uses the Claude Code CLI. If you already pay for a Claude subscription, there is no API key and no per-token bill.
  • API mode uses the Anthropic API directly. Needs a key, and works in Docker.

Group AI, without the group hating it

The failure mode of a bot in a shared room is that it will not stop talking. Every bot that has ever been added to a busy channel and removed a day later was removed for this reason.

Telechat's group behaviour defaults to restraint: it answers when mentioned or replied to, and stays quiet otherwise. /groupmode mention|all|off changes that per room, so a channel that genuinely wants an always-on participant can have one and everywhere else stays calm.

Each room gets its own conversation. The bot in your team's incident channel does not carry over context from your private chat, which is both the correct privacy behaviour and the correct usefulness behaviour — a group thread should read like a group thread, not like it is halfway through someone else's discussion.

Access is handled the same way, socially rather than administratively. /invite mints a one-tap link that grants a friend access — no user IDs to collect, no .env edit, no restart — and /invites shows who joined through yours.

Tools, MCP, and the part that stops it being a chat toy

An assistant you can only talk to is a novelty. The value shows up when it can do things, and Telechat ships a working set of tools behind slash commands:

/code, /project Start a coding task; set the working directory
/plan Two-agent multi-step planning with progress updates
/search, /fetch, /web Web search, URL summarize, page browse
/kb Knowledge base — upload documents, full-text and semantic search
/imagine, /tts, /music, /video Image, speech, music and video generation
/schedule, /poll Scheduled tasks; polls

Alongside these it speaks MCP, so any Model Context Protocol server you already run — your database, your issue tracker, your internal API — becomes something you can reach from your phone. This is the piece that turns "Claude, but on Telegram" into an actual remote control for your own systems.

Two features round it out. Memory is per-user with full-text search, recalled into every reply, so context carries across sessions rather than dying with the chat window; /remember, /recall and /forget let you manage it directly. And smart model routing picks Haiku, Sonnet or Opus by query complexity, so a one-line question does not get billed at deep-reasoning rates.

Voice messages work too: send a voice note and it answers the transcript, free via Groq's Whisper tier, no paid key required.

The bridge, which is the part nobody else has

If you use Claude Code, this is the feature worth the install on its own.

When a Claude Code session on your machine finishes a turn or stalls on a question, you get a triage card on Telegram — status first, the pending decision pulled to the top. Reply to the card and your answer is injected back into that session via claude --resume. Pick from a session list if several are running.

Turn on --approval and every Bash, Write and Edit call waits for your tap, with the call preview and Approve / Deny / Always-allow buttons on the card. Standing rules resolve the repetitive ones without asking.

Which means a long-running agent stops being something you have to babysit at a desk. It runs, it pages you when it genuinely needs a human, you answer from wherever you are, and it continues.

Who this is actually for

Telechat is explicit about its boundary: it is a single-operator tool, and in CLI mode the bot runs with your local permissions. That is the right trade for a personal assistant on your own hardware, and the wrong one for a shared corporate deployment. Knowing which of those you are building matters more than any feature on the list.

If you already use Claude and you have caught yourself wishing it were on your phone without handing your history to a third-party wrapper, that is the whole pitch.

npm install -g telechat && telechat init
Enter fullscreen mode Exit fullscreen mode

telechat.fyi · github.com/telechatai/telechat

Top comments (0)