-ffb86c){kind=icon}
Your own AI. In Telegram. Coding agent on your laptop.
One install. ~$2/month on DeepSeek. No limits. No lock-in.
bash <(curl -sL https://raw.githubusercontent.com/MozgAI/mavka/main/install.sh)macOS · Linux · Windows · 5 minutes · ~$2/month on DeepSeek · 5 AI providers · hands-free voice · AI installs itself · MIT
vs ChatGPT · Vibecoding · The math · Install · Cost · Privacy · Security
I was about to drop $5,000–$10,000 on a workstation to run local AI models at home. The plan: a beefy GPU rig, MLX, llama.cpp, the whole stack.
Then DeepSeek shipped V4 Flash — a 248-billion-parameter model at $0.14 per million tokens.
Read that price again.
A $10,000 home server cannot physically run a 248B model. I'd be stuck with something half the size, half the quality. Meanwhile, the same money on the API covers years of heavy daily use — and the hardware would be obsolete long before that.
Hardware depreciates monthly. Cloud frontier models get smarter weekly. New providers launch, prices keep dropping, and you can switch brains anytime — no lock-in.
So I stopped buying servers and built MavKa — a one-line installer that drops a full AI assistant into your Telegram, powered by these absurdly cheap cloud models. Five minutes. ~$2 a month on DeepSeek for an active chat companion. Done.
The lifehack of 2026: don't pay $20/month for ChatGPT Plus. Don't buy a $10,000 server. Top up $2 on DeepSeek and use it for a month.
| ChatGPT Plus | MavKa | |
|---|---|---|
| Price | $20/month | ~$2/month (active chat use, on DeepSeek) |
| Brain | One model, take it or leave it | 5 providers, swap anytime |
| Where you use it | Their app, their UI, their rules | Telegram + your terminal |
| Voice in | Yes | Yes (Groq Whisper, free) |
| Voice out (TTS) | App only | Yes — replies as voice notes (free Edge TTS) |
| Vision | Yes | Yes (Gemini) |
| Web Search | Yes | Yes (Tavily) |
| Memory | Black box, theirs | Persistent wiki, yours |
| Personality | Vendor-controlled | You shape it. Chef? Coach? Tutor? Yes. |
| Computer use (read/write files, run commands) | No | Yes — Pi Agent under the hood |
| Vendor lock-in | Stuck with OpenAI | Switch to Claude / DeepSeek / Kimi / Groq in 30 seconds |
| Where data goes | OpenAI servers, period | API providers you choose, keys you own |
| Open source | No | Yes (MIT) |
| Self-hostable | No | Yes — runs on your machine |
| Setup difficulty | Sign up → done | One bash command. AI walks you through the rest. |
OpenAI can't switch you to Anthropic. Anthropic can't put themselves on your phone. Neither can put a real coding agent on your laptop with the same brain that just answered your voice note.
MavKa is one install. Two superpowers. Five providers. Zero lock-in.
| Cursor | Claude Code / OpenClaw | MavKa | |
|---|---|---|---|
| Price | $20/mo | Pay-per-token | ~$5–10/mo heavy coding (DeepSeek) |
| Where you talk to it | IDE only | Terminal only | Telegram + terminal |
| Voice control (in + out) | No | No | Yes — fully hands-free |
| Photo / screenshot understanding | No | No | Yes |
| Switch model providers | Limited | Limited | 5 providers, anytime |
| Memory across sessions | Limited | Limited | Persistent wiki |
| Open source | No | No | Yes (MIT) |
| Install help | None — read docs | None — read docs | AI installs itself, in your language |
Most AI tools force a choice: a chat app on your phone, or a coding agent on your PC. MavKa is both — sharing the same brain, the same memory, the same conversation.
Voice in, voice out, photo analysis, web search, persistent memory. Talk to MavKa hands-free — send a voice note, get a voice note back. Driving, cooking, working out — keep going, MavKa speaks. No app to install. Same chat works on your phone, tablet, watch, anywhere Telegram runs.
Voice in via Groq Whisper (free 8h/day). Voice out via Microsoft Edge TTS (free, no key). Hands-free conversation costs zero on top of the LLM tokens.
Built on Pi Agent (the open-source engine behind Claude Code, OpenClaw and similar tools). Full file access, shell, computer use. Ask MavKa in Telegram to refactor a file on your laptop — it just does it.
One install. Two superpowers. Pennies a month.
Here's the dirty secret of every "self-hosted AI agent" out there: you need another AI to install it. People literally open ChatGPT or Claude Code in another tab and ask "how do I configure this auth.json" — because doing it by hand is fragile, error-prone, and the docs are written for engineers.
MavKa is the first AI agent where the AI installs itself. After you give it the API key, the installer hands you over to MavKa's own brain — running live, in your terminal, in your language — and walks you through the rest of setup conversationally:
- "Я не могу найти этот ключ" → it explains where to click, in Russian.
- "What does Tavily do? Do I need it?" → straight answer, then asks if you want to skip.
- "можем сделать это потом" → done, on to the next step.
No second tool. No copy-pasting questions to ChatGPT. No engineer in the family.
Slogan: "Even your mom can install it." And we mean it.
MavKa isn't just for developers. It's a personal AI agent with a brain, internet access, and your memory — useful for almost everyone who has a phone.
Voice-driven coding, repo refactors, debugging, commits, deployments — all from Telegram, all running on your laptop. See Vibecoding the Easy Way below.
Explain a calculus problem with a photo of the page. Summarize a 60-page PDF. Practice a foreign language out loud (voice in, voice out). Get tutored on any topic — MavKa remembers what you already know.
Hands wet, dough on fingers, baby on hip — talk to MavKa, get voice answers back. "What can I cook with these?" — photo of your fridge. "Best way to remove red wine from white shirt?" — voice while doing laundry. "Make me a 7-day meal plan, $80 budget, two people" — done. "Read this label, is it gluten-free?" — photo of the package.
Send a photo of your plate — MavKa estimates calories and macros. Discuss meal plans, training programs, swap exercises. Track progress in persistent memory across weeks.
Hold up your phone, say "I'm at the gym, dumbbells busy. Build me a chest workout with just barbells." — MavKa replies as a voice note. Keep training. No screen, no typing. "How do I do a Romanian deadlift correctly?" — voice in, voice out. Step-by-step in your earbuds.
"Read these ingredients — anything I should avoid for sensitive skin?" "Outfit advice for a job interview, budget $200" — describe what you have, MavKa picks combinations.
Photograph receipts → MavKa categorizes them. Build a personal budget. Walk through tax basics. Ask about a contract clause before you sign.
"What does this rash look like?" — photo + symptoms → general info + advice to see a doctor. "My BP reads 140/90 — should I be worried?" — context, what to track. MavKa is not a doctor. For anything serious, see a real one. It's a smart sidekick for tracking and curiosity, not a replacement for medical care.
Translate a menu via photo. Get hidden-gem restaurant picks for a city you're visiting. Draft a polite cancellation email to your landlord. Plan a trip itinerary. Find that one website you visited last week.
Real-time web search via Tavily. Photo analysis via Gemini. Voice in via Whisper. Persistent memory that learns who you are — your kids' names, your allergies, your projects, your goals.
MavKa is the agent you can ask anything. It has the internet, your camera, your voice, and a long memory. It's yours.
You've heard the term — vibecoding: you describe the vibe, the AI writes the code, you review and ship. The problem is the on-ramp:
- Cursor / Windsurf — slick, but $20/month and another IDE to learn
- Claude Code / OpenClaw — powerful, but it's a CLI tool with a learning curve
- GitHub Copilot — autocomplete-tier, not a real agent
- Local models — slow on consumer hardware, model quality lags 6 months
MavKa is different: the same AI that just answered your voice note in Telegram is sitting in your terminal, ready to read your repo, run tests, write commits, push branches.
You (in bed): "fix the auth bug we talked about yesterday"
MavKa (laptop): reads ~/projects/api/auth.ts, finds the issue,
runs the failing test, patches it, runs again,
reports back in your Telegram with a diff
Same brain. Same memory. Same conversation. One install, no IDE switch, no $20/month subscription. The cheapest, lowest-friction entry into agentic coding that exists right now.
If you want to start vibecoding without paying $20/month for Cursor or fighting Claude Code's CLI — MavKa is the on-ramp.
MavKa is provider-agnostic. Choose at install time, swap anytime by re-running the installer:
| Provider | Why pick it | Active chat-bot use (real) |
|---|---|---|
| DeepSeek V4 Flash | 248B parameters at $0.14/M tokens — the lifehack. (recommended) | ~$2/mo |
| ChatGPT GPT-4o-mini | Mainstream OpenAI quality, familiar API | ~$5/mo |
| Kimi 2.6 Moonshot | Long-context (262K), strong on code | ~$25–35/mo |
| Opus Claude 4.7 | Anthropic's flagship — smartest model on the market | ~$200–400/mo (no caching) |
| Groq Llama 3.3 70B | Free tier with daily limits, fastest inference | $0 (within free quota) |
Why DeepSeek is the recommendation: no other provider gets you a 248B model anywhere near $0.14 per million tokens. It's a generational price-to-quality leap. Every other provider listed here is at least 2–200× more expensive for the same daily chat volume. The reason they're still in the menu: nobody likes lock-in, and some users have credits or strong preferences.
bash <(curl -sL https://raw.githubusercontent.com/MozgAI/mavka/main/install.sh)powershell -ExecutionPolicy Bypass -c "irm https://raw.githubusercontent.com/MozgAI/mavka/main/install.ps1 | iex"
⚠️ Windows support is currently in BETA. Mac/Linux are battle-tested, Windows hasn't been deeply tested yet. Same flow, same providers, same AI-guided setup. Uses Git Bash (auto-installed via winget) for Pi Agent's shell — WSL not needed. Please file an issue if anything breaks.
# macOS / Linux
git clone https://github.com/MozgAI/mavka.git
cd mavka
less install.sh
bash install.sh# Windows
git clone https://github.com/MozgAI/mavka.git
cd mavka
notepad install.ps1
powershell -ExecutionPolicy Bypass -File install.ps1Required:
- AI Provider — DeepSeek / ChatGPT / Opus / Kimi / Groq (one click)
- API key for the provider you picked
- Telegram Bot Token → @BotFather (free)
- Your Telegram User ID → @userinfobot (free)
Optional (skip to enable later):
- Groq Key → voice transcription via Whisper, free 8h/day
- Gemini Key → photo analysis, free tier
- Tavily Key → web search, free 1000/month
After step 3, MavKa's own AI takes over and walks you through the rest conversationally — in your language, on the same screen. No googling, no docs.
After install: open Telegram, say hi to MavKa. Done.
# macOS / Linux
bash <(curl -sL https://raw.githubusercontent.com/MozgAI/mavka/main/uninstall.sh)# Windows
powershell -ExecutionPolicy Bypass -c "irm https://raw.githubusercontent.com/MozgAI/mavka/main/uninstall.ps1 | iex"Removes MavKa, autostart, local config. Doesn't revoke API keys — do that yourself on each provider.
Let's be honest. MavKa is "self-hosted" but the AI brain itself runs in the cloud. Here's exactly what leaves your machine:
| Data | Goes to |
|---|---|
| Text messages | The AI provider you picked (DeepSeek / OpenAI / Anthropic / Moonshot / Groq) |
| Voice notes | Groq API (Whisper transcribes → text → your AI) |
| Photos | Google AI Studio (Gemini) |
| Web search queries | Tavily API |
| Telegram messages | Telegram Bot API |
What stays local: your API keys, bot config, persistent memory, personality preset, conversation cache.
What does NOT stay local: the actual content of every message you send — by design. That's the trade-off for not running a 248B model on your laptop.
If you need air-gapped privacy, MavKa is the wrong tool. Run llama.cpp with a local model instead.
The honest pitch isn't "private," it's: you own the keys, you pick the providers, and you can swap them anytime. OpenAI can't do that. Anthropic can't do that. You can.
See SECURITY.md for full details on data flow, key storage, and how to nuke everything.
MavKa runs Pi Coding Agent under the hood, and Pi is intentionally a "YOLO" agent — its author says so openly. It has no built-in permission popups, no per-action confirmations, no allow-lists. When the LLM says "run this shell command", Pi runs it.
Treat MavKa like SSH access to your user account. That means:
- It can read any file your user can read.
- It can write/delete files in your home directory.
- It can run shell commands as you (no
sudo, but everything else). - It can make outbound network requests.
What MavKa does to mitigate this:
- On macOS/Linux, the installer enables Pi's
@anthropic-ai/sandbox-runtimeextension by default. It deny-lists~/.ssh,~/.aws,~/.gnupg, and secret-looking paths (*.pem,*.key,.env*) for shell commands. It usessandbox-execon macOS,bubblewrapon Linux. It does NOT sandbox theread/write/edittools — those still have full FS access through Node. - API keys are stored in
~/.pi/agent/auth.jsonmode 0600 (only your user can read). - You can completely uninstall with
bash <(curl -sL .../uninstall.sh)— and rotate API keys on each provider's site.
What MavKa does NOT do:
- No per-command confirmation prompts (that fights Pi's design).
- No Docker/VM isolation.
- No code review of LLM-suggested actions before they execute.
If that's not your trust level, don't install MavKa on a machine with credentials you can't afford to lose. Run it in a VM, or wait for a future version with stronger sandboxing. Be honest about who's getting SSH-equivalent access here: you, the LLM you picked, and (in theory) anyone who can compromise that LLM provider.
These are real costs at typical chat-bot use (~150–200K tokens/day total — the kind of person who actually uses MavKa daily for chat, voice, photos, web search, light coding).
| Provider | Active chat-bot use | Heavy coding |
|---|---|---|
| DeepSeek V4 Flash (recommended) | ~$2/mo | ~$5–10/mo |
| Groq Llama 3.3 70B | $0 (within free tier; lighter chats only — limits hit fast on heavy days) | ~$10–20/mo over free TPD |
| ChatGPT GPT-4o-mini | ~$5/mo | ~$15–25/mo |
| Kimi 2.6 Moonshot | ~$25–35/mo | ~$50–80/mo |
| Opus Claude 4.7 | ~$200–400/mo (no caching) | ~$50–150/mo with aggressive prompt caching |
Tools layered on top (free):
| Service | Pricing |
|---|---|
| Groq Whisper (voice) | Free (8h/day) |
| Gemini Vision (photos) | Free tier |
| Tavily Search | Free tier (1000/mo) |
Why DeepSeek is the only honest answer: at $0.14/M input + $0.28/M output, an active daily user pays about $2 a month. Not $0.05. Not "$2 lasts a year." About $2 a month — and that's still cheaper than every alternative on this list by 2× to 200×. The $2 starter credit DeepSeek requires you to top up will last roughly one month of active use. When it runs out, top up another $2 and keep going.
Prices as of May 2026. AI providers change pricing frequently — verify on the official pricing page before committing. Numbers above are realistic estimates for an active chat-bot user, not lab benchmarks.
You (Telegram)
↓
Telegram Bot API
↓
Pi Agent (your machine) ← pi-telegram extension
↓
Your AI provider (cloud) ← Groq + Gemini + Tavily for tools
Your computer runs a lightweight Node.js process. The heavy thinking happens in whichever cloud you picked. Your data, memory, and config stay local. The brain is rented per token — switch providers anytime by re-running the installer.
- macOS — Apple Silicon & Intel
- Linux — Ubuntu, Debian, Fedora, Arch, ARM
- Windows — Windows 10 (1809+) and Windows 11, native PowerShell, no WSL needed (uses Git Bash). Beta — please report bugs.
- No computer at all? → Mavka Light — runs on Fly.io free cloud, $0/month. For people who want a Telegram AI bot without owning a Mac/Linux/Windows machine.
macOS / Linux:
tail -f ~/mavka-bot/mavka.log # logs
tmux attach -t mavka # attach to console (Ctrl+b d to detach)
bash ~/mavka-bot/launch.sh # restartWindows:
mavka logs # tail logs
mavka stop # stop
mavka start # start
mavka restart # restart
mavka status # scheduled task status
mavka uninstall # remove autostartI spent half a day manually setting up this same agent for my brother. SSH, configs, dependencies, Pi Agent patches, launchd plist. Hours.
Then it hit me: if a developer needs half a day, a normal person has zero chance.
MavKa is the fix. One command. Five minutes. Done. And while I was at it, I made sure no provider can ever lock you in.
Early days. Tested on macOS (Apple Silicon) and Arch Linux. Windows is in beta.
If something breaks — open an issue on GitHub. If you can fix it — send a Pull Request. Everyone welcome.
MIT — do whatever you want with it.
Stop paying $20/month for AI you don't own.
Stop fighting CLIs to ship code with AI.
One install. ~$2 a month. Forever yours.
Made with 🍃 — open source forever.