Text Your Laptop Like You Text Your Mate
"The agent should not live behind a login screen — it should be wherever you are, ready the moment you need it." — Manus team, on why AI agents belong in messaging apps
The Story
You have a friend who's always available. Not the kind that lives in a browser tab you forget about, or a terminal window buried under twelve other windows, or an IDE sidebar you only open when you're already coding. This one's a contact in your phone. You text it the same way you text anyone else.
The difference is this friend is also a developer, a researcher, a sysadmin, and a personal assistant. It remembers what you told it three weeks ago. It can actually go and do things — write code, search the web, check your systems, file a report. And it talks to you in the same place where you talk to everyone else: your messaging app.
This is the fundamental shift. Not a better chatbot. Not a smarter autocomplete. A persistent, capable agent that lives where you already are.
The Idea
Your AI agent has a Telegram account. You text it throughout the day, the same way you'd message a colleague. Except this colleague has perfect memory, can write code, search the internet, and run commands on your server — all from a text conversation on your phone.
Walking to the station, 8:14am
remind me to email Sarah about the Q1 numbers tomorrow at 10
Done. I'll ping you tomorrow at 10am. Want me to pull the latest Q1 figures before then so you have them ready?
On the couch, 7pm
what was that error we debugged on Friday?
The auth middleware was failing silently on token refresh. The issue was in middleware/auth.go line 142 — the error from refreshToken() was being swallowed instead of propagated. We patched it Friday afternoon and added a retry with exponential backoff. The fix has been running in staging since then with zero failures.
In bed, 11:30pm, just had an idea
I think the onboarding flow should skip the database step if they're connecting to a remote server. Can you draft that change?
Good call. If they've already entered a remote connection string, the local database setup is redundant. I'd skip the database step when the connection mode is remote and update the progress indicator to show 4 steps instead of 5. Two files to change — want me to make the full change and run the tests?
At the pub with friends, Saturday afternoon
how much did we spend on API calls this month?
March so far (17 days in):
Anthropic: $34.20 (Claude Sonnet, mostly coding agent)
OpenAI: $12.80 (embeddings for memory)
Brave Search: $0.00 (free tier, 1,847 of 2,000 queries used)
Total: $47.00
You're on track for about $83 this month — up 15% from February. The increase is mostly from the session branching work last week, which was heavy on the coding agent. Normal months run closer to $70.
What Makes It Different
It can do things. This isn't a question-answering bot. It writes code, runs commands, searches the web, checks your systems, creates files, deploys changes. When you ask it to draft a change, it actually drafts the change.
It remembers. Not just this conversation — every conversation, across every channel. The error you debugged on Friday. The architectural decision from January. The preference you mentioned once about preferring explicit error handling. It's all in PostgreSQL on your machine, retrieved with vector search when relevant.
Same person everywhere. The agent you text on Telegram is the same one that responds in Slack, the same one in the terminal, the same one on the desktop app. One identity, one memory, many interfaces.
It has voice. Not literally — but it develops a consistent personality shaped by your interactions. It learns your communication style, your preferences, your shorthand. After a few weeks it doesn't feel like a tool. It feels like texting someone who knows you.
It's proactive. It doesn't just wait for you to ask. It notices things. That deadline you mentioned is tomorrow and you haven't started. That dependency you pinned three months ago has a critical CVE. That pattern in your API costs that suggests a runaway loop.
How It Works
- Channels — Telegram channel receives your messages and streams responses back. Same protocol handles Slack, HTTP, stdio, and the desktop app.
- Memory — Three-tier memory system: session context (current conversation), working memory (active projects and recent facts), and long-term memory (everything, stored in PostgreSQL with pgvector embeddings). Relevant memories are retrieved automatically based on what you're talking about.
- Coding agent — A full agentic coding loop with 9+1 tools: read, write, edit, exec, grep, find, ls, truncate, workspace info, and create sub-agents. When you ask it to draft a change, it actually reads the code, understands the context, and writes the diff.
- Search — Multi-provider search (Perplexity, Brave, Google) with intelligent routing, result aggregation, and two-tier caching. Deep research mode for complex questions.
- Scheduler — Cron-based and one-shot scheduling. "Remind me tomorrow at 10" creates a scheduled job that fires a message through your preferred channel.
- Remote access — The agent runs on your machine (or a server you control). Telegram is just the interface. The agent has access to your codebase, your tools, your databases — everything it would have if you were sitting at your desk.
- Judge — A safety layer that evaluates risk before the agent acts. Reading a file? Low risk, go ahead. Running a database migration? High risk, ask for approval on your phone first.
What Breaks Without This
ChatGPT lives in a browser tab. You have to go to it, open a new conversation, re-explain context every time. It can't do anything — no code execution, no file access, no system commands. It forgets everything between sessions.
Claude Code is powerful but lives in the terminal. You have to be at your desk, in a terminal, in the right directory. You can't text it from the pub. It doesn't remember last week's conversation.
Cursor / Copilot live in the IDE. Great for the specific task of writing code in an editor. Useless for everything else — reminders, research, system monitoring, deployment, casual questions while you're walking somewhere.
Siri / Alexa have no memory, no real agency, and no useful integration with your actual work. They can set a timer. They cannot debug your auth middleware.