mrstack 1.1.1

Proactive AI Butler for Claude Code — Telegram Bot with Memory, Jarvis Mode, and Daily Coaching

Mr.Stack

Claude that works while your MacBook sleeps. Morning briefings, nightly coaching.

Claude Code + Telegram = Your 24/7 AI development partner

Observes your Mac | Remembers conversations | Learns your patterns | Speaks first — an AI partner that doesn't wait to be asked.

Quick Start

# Install (pick one)
uv tool install mrstack    # recommended
pip install mrstack         # alternative
# git clone https://github.com/whynowlab/mrstack.git && pip install -e .

# Setup & run
mrstack init               # interactive setup wizard
mrstack start              # start the bot
# or: mrstack daemon       # register as background service

All CLI commands

mrstack init          # Interactive setup wizard
mrstack start         # Start bot (foreground)
mrstack start --bg    # Start bot (background)
mrstack stop          # Stop bot
mrstack daemon        # Install system daemon (launchd/systemd)
mrstack daemon -u     # Uninstall daemon
mrstack status        # Show status
mrstack logs [-f]     # View/follow logs
mrstack config        # Edit .env configuration
mrstack jarvis on/off # Toggle Jarvis mode
mrstack patch         # Re-install Mr.Stack modules
mrstack update        # Update to latest version
mrstack version       # Show version info

Legacy install (git clone && ./install.sh) is still supported. See Installation for details.

---

What is Mr.Stack?

An always-on Telegram AI partner built on Claude Code.

It extends Claude Code beyond the terminal — through Telegram, it works 24/7, even when your MacBook lid is closed, you're away from your desk, or you're asleep.

A typical AI assistant sits in an office waiting for your call. Mr.Stack is more like a senior colleague sitting next to you, looking at the same screen.

Three pillars:

• Proactive alerts — Scans your Mac every 5 minutes and reaches out via Telegram when something needs attention
• Persistent memory — Remembers conversations and carries context across sessions
• Pattern learning — Analyzes your work habits and sends data-driven coaching every night

---

Key Features

1. Telegram AI Assistant

Give Claude instructions with a single Telegram message.

• Write / edit / debug / refactor code
• Process files, photos (image recognition), and voice messages
• Voice in, voice out (Whisper + TTS)
• Auto-analyze clipboard content (error → root cause, URL → summary, code → review)

2. Memory System — An AI That Doesn't Forget

Most AIs have goldfish memory. Close the window, everything resets. Mr.Stack separates working memory (volatile) from long-term memory (persistent), much like the human brain.

Working Memory (Volatile)

• Context snapshots: Last 1 hour of system state (in-memory only, 12-slot ring buffer)
• Current session: Lives within Claude's context window
• Recent interactions: Last 7 days from interactions.jsonl → used for pattern analysis

Long-Term Memory (Persistent)

Every 3 hours, recent conversations are analyzed and only meaningful information is extracted for permanent storage:

~/claude-telegram/memory/
  people/          → People mentioned in conversations (overwrite)
  projects/        → Project progress and technical decisions (overwrite)
  decisions/       → Decisions and their reasoning (append)
  preferences/     → Your preferences — coding style, tools, language (overwrite)
  daily/           → Daily summary archive (1 file per day)
  knowledge/       → Newly acquired knowledge (overwrite)
  patterns/        → Work pattern data (interactions.jsonl)

The key insight: it updates, not accumulates. Like Git's HEAD, people/, projects/, and preferences/ always maintain only the latest snapshot. Files don't grow forever.

The only append-mode data is daily/ (1 per day, ~1KB) and interactions.jsonl (~300B per conversation). After a full year of use, total storage is 10–15MB.

Say "continue what I was working on yesterday" and it picks up exactly where you left off.

3. Always-On System (Jarvis Mode)

Like a linter watching your code in real-time, Mr.Stack scans your entire Mac every 5 minutes:

• Active application (osascript)
• Battery level and charging state (pmset)
• CPU load (sysctl)
• Git branch and working tree status
• Chrome active tabs
• Recent terminal commands

From this data, your current context is classified into 7 states:

CODING · BROWSING · MEETING · COMMUNICATION · BREAK · DEEP_WORK · AWAY

Based on state classification, 7 triggers fire proactively:

Situation	Mr.Stack's Response
Battery below 20%	"Battery at 12% — save your work"
Returned from away	"Welcome back. You were on the feature/auth branch"
3+ hours of continuous coding	"You've been coding for 3 hours — take a break"
5+ app switches in 10 min	"Too many context switches. Try to focus"
Same file for 30 min, no commits	"Are you stuck? Need help?"
Terminal error detected	"Want me to analyze that error?"
Deep work mode (2+ hours focused)	Only critical alerts pass through

Each trigger has a debounce cooldown. No alert spam. During deep work: interrupt masking — everything blocked except battery warnings.

4. Pattern Learning — Logs Become Insights

Every conversation is logged as a structured event in interactions.jsonl:

{
  "ts": "2026-02-28T10:00:00",
  "hour": 10,
  "state": "CODING",
  "request_type": "debug",
  "duration_ms": 5200,
  "tools_used": ["Read", "Edit", "Bash"]
}

Like Prometheus collecting metrics, these logs are analyzed as time-series data:

• Peak hours: "Most productive between 10–12" (max throughput window)
• Request type distribution: "40% debugging, 30% features, 20% questions" (work profiling)
• Pattern detection: "Requests weekly planning every Monday at 10am" (recurring event detection)
• Routine learning: Confidence > 0.7 triggers preemptive preparation (predictive prefetching)

5. Daily Coaching Report

Every night, a report like this arrives via Telegram:

[Daily Coach] 2026-02-28

Productivity: 8/10 (+1 from yesterday)

What went well:
• 2-hour focus block in the morning → 4 commits
• Clear specs → 0 rework

Areas to improve:
• 7 context switches between 2–4pm — least efficient window
  → Tomorrow: try a single-task block during this time
• 40% debugging ratio — writing tests first could cut this in half

Weekly trend:
• Instruction clarity improving (2.1 → 2.6)
• Most productive hours: 10–12 (consistent)

What a scrum master should say during sprint retro — delivered nightly, backed by data. No flattery, just numbers. Manual access via /coach.

6. Context-Aware Tone (Persona Layer)

The same information should be delivered differently depending on context. Terse while coding, summary-first when returning.

State	Tone	Example
Coding	Technical, concise	"auth.py:42 — missing null check"
Deep work	Silent (urgent only)	—
Returning from break	Warm, summary	"Welcome back. PR #23 awaiting review"
After 10pm	Concerned	"You've done enough today. Pick it up tomorrow"

7. Scheduled Jobs (9 cron-based automations)

Like a CI/CD pipeline — configure once, runs automatically on schedule.

Job	Schedule	Model
Morning briefing (calendar + tasks + news)	Weekdays 08:00	Sonnet
Evening summary + coaching report	Daily 21:00	Opus
Weekly review	Fridays 18:00	Opus
Calendar reminder	Weekdays 09/12/15/18	Haiku
Conversation → memory auto-update	Every 3 hours	Sonnet
GitHub notification check	Every 2 hours	Haiku
Notion work log sync	Daily 22:00	Sonnet
AI news summary + Notion save	Daily 08:30, 20:30	Sonnet
OAuth token expiry monitor	Daily 10:00	Haiku

Model routing: Automatically selects the right model per job complexity. Simple lookups (Haiku) → analysis/summary (Sonnet) → deep reflection (Opus). Reduces overnight usage by 60–70%.

8. External Integrations (MCP)

Mr.Stack connects to external services via MCP (Model Context Protocol). Each service is optional — core features work fine without them.

Service	Capability	Setup
Google Calendar	View/create events, reminders	Google Cloud OAuth 2.0
Notion	Automatic work log recording	Notion Integration API key
GitHub	PR, issue, notification monitoring	gh auth login (CLI auth)
Playwright	Web automation	npx playwright install
AppleScript	macOS control	Accessibility permission

Jobs for unconnected services are automatically skipped (no errors).

9. Claude HUD — Real-Time Usage Monitor

A real-time status bar (HUD) is displayed at the top of your Claude Code terminal. Usage data is fetched directly from the Anthropic API, so it's always accurate.

Displays:

• Model name — Currently active Claude model
• Context usage — Context window utilization
• Plan usage — Usage percentage on 5-hour / 7-day windows
• Tool activity — Currently running tool status
• Agent status — Active agent information
• Todo progress — Task list completion rate

install.sh automatically installs and configures the HUD. No extra setup needed.

---

How Is This Different?

	Typical AI Bot	Mr.Stack
Interaction	Request-response (must be asked)	Event-driven (proactive alerts on conditions)
Security	API keys sent to external servers	Runs only on your Mac. Zero servers
Availability	Tied to terminal session	macOS daemon — 24h even with lid closed
Memory	Stateless (session end = reset)	Persistent memory (permanent storage, context carried)
Learning	None	Time-series pattern analysis + routine prediction
Coaching	None	Daily productivity metrics + data-driven suggestions
Data	Cloud storage	100% local (SQLite, JSONL, Markdown)

---

Prerequisites

Before you begin:

Requirement	Description	Where
macOS	Apple Silicon or Intel	—
Claude Code	Anthropic's AI coding tool (Max plan recommended)	claude.ai/download
Telegram account	Messenger to communicate with the bot	telegram.org
Telegram Bot Token	Create a bot via BotFather	Send /newbot to @BotFather
Telegram User ID	Your ID for receiving notifications	Send any message to @userinfobot
claude-code-telegram	Base bot framework	github.com/nicepkg/claude-code-telegram

Get Your Telegram Bot Token (2 min)

1. Search for @BotFather in Telegram
2. Send /newbot
3. Enter a bot name (e.g., My Stack Bot)
4. Enter a bot username (e.g., my_stack_bot)
5. Save the token (1234567890:ABCdef...)

Find Your Telegram User ID (30 sec)

1. Search for @userinfobot in Telegram
2. Send any message
3. Save the numeric ID

---

Installation

Choose one of two methods:

Method A: Manual Install

Step 1: Install the base bot

Install claude-code-telegram first.

uv tool install claude-code-telegram

Configure your .env file and verify the bot runs correctly.

Step 2: Install Mr.Stack

git clone https://github.com/whynowlab/mrstack.git
cd mrstack
./install.sh

The install script automatically:

• Locates your claude-code-telegram install path (uv / pipx)
• Copies the Jarvis module
• Patches 3 config files
• Adds ENABLE_JARVIS=true to .env
• Prompts for your Telegram User ID
• Creates the memory directory

Step 3: Restart the bot

# If running as a launchd daemon:
launchctl stop <your-service-name>
launchctl start <your-service-name>

# If running manually:
# Stop the process and restart

Step 4: Verify

curl http://localhost:8080/health   # If API server is enabled

Send any message to your bot on Telegram. /jarvis — check status | /coach — coaching report

---

Method B: Install with Claude Code

If you have Claude Code installed, open it in your terminal and say:

Look at github.com/whynowlab/mrstack and install Mr.Stack for me.

Claude Code will read the README and install.sh, then guide you through setup. If you don't have the base bot (claude-code-telegram) yet, just ask Claude to install that too.

---

(Optional) External Service Integrations

MCP integrations are optional. Mr.Stack's core features work without them. Set up only the services you need.

Google Calendar

1. Create a project in Google Cloud Console
2. Enable Calendar API
3. Create OAuth 2.0 Client ID (Desktop app)
4. Add client_id and client_secret to MCP config
5. Register the Google Calendar server in mcp-config.json:

{
  "mcpServers": {
    "google-calendar": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-google-calendar"],
      "env": {
        "GOOGLE_CLIENT_ID": "your-client-id",
        "GOOGLE_CLIENT_SECRET": "your-client-secret"
      }
    }
  }
}

6. Enable MCP in .env:

ENABLE_MCP=true
MCP_CONFIG_PATH=/path/to/mcp-config.json

Notion

1. Create a new integration at Notion Integrations
2. Copy the Internal Integration Token
3. In your target Notion database, click "Connect" → add the integration
4. Add to mcp-config.json:

{
  "mcpServers": {
    "notion": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-notion"],
      "env": {
        "NOTION_API_KEY": "ntn_your_api_key"
      }
    }
  }
}

GitHub

The easiest one. In your terminal:

gh auth login

That's it. Mr.Stack uses the gh CLI to check GitHub notifications, PRs, and issues.

Playwright (Web Automation)

npx playwright install chromium

---

Commands

Command	Description
/new	Start new conversation
/status	Session & cost overview
/repo	Switch project
/jobs	List scheduled jobs
/usage	Usage report
/help	Command reference
/verbose	Output verbosity (0/1/2)
/voice	Toggle voice responses
/clipboard	Clipboard auto-analysis
/jarvis	Pause/resume Jarvis
/coach	Daily coaching report
/propose	Spec-based change proposal
/apply	Execute spec implementation
/archive	Archive completed work
/explore	Exploration mode

---

Data & Privacy

• All data stored only on your Mac (~/claude-telegram/memory/)
• No external server transmissions. No cloud. No telemetry. Zero-trust architecture.
• System snapshots are refreshed in-memory every 5 min, never persisted to disk
• Claude API calls only on trigger activation (max 10/hour, rate-limited)

---

Tech Stack

Layer	Technology
Runtime	Python 3.11 + asyncio (async event loop)
AI Engine	Claude Code SDK (Opus / Sonnet / Haiku)
Interface	Telegram Bot API
Process Management	macOS LaunchAgent (daemon, auto-restart)
API Server	FastAPI (health check, webhooks)
Storage	SQLite (conversations) + JSONL (event logs) + Markdown (memory)
Integrations	MCP — Google Calendar, Notion, Playwright, AppleScript
Voice	Whisper (STT) + TTS (speech synthesis)

---

Compatibility

Item	Version
claude-code-telegram	v1.3.0+
Python	3.11+
macOS	Ventura / Sonoma / Sequoia

---

FAQ

Does it work with the free version of Claude Code?

It works, but Mr.Stack calls Claude for scheduled jobs, Jarvis triggers, and regular conversations, so Max plan is strongly recommended. Free plans hit rate limits quickly.

Do I need an Anthropic API key?

No. It uses Claude Code's built-in authentication. No separate API key or billing required.

How much does it cost?

Mr.Stack itself is free (open source). You need:

• Claude Code subscription: Pro ($20/mo) or Max ($100/mo recommended)
• Telegram: Free
• Server costs: None (runs on your Mac)

Does it work on Windows or Linux?

Currently macOS only. System state scanning uses macOS-specific tools (osascript, pmset, etc.). Linux support is on the roadmap.

Does it really work with the MacBook lid closed?

Yes. When registered as a macOS LaunchAgent (daemon):

• Auto-starts on boot
• Auto-restarts on crash
• Runs in background with lid closed (Power Nap / connected to power)
• Accessible via Telegram while you're out

Is my data sent externally?

Absolutely not. All data is stored in local files on your Mac (SQLite, JSONL). No external servers, no cloud storage, no telemetry whatsoever. The only external communication is Claude API calls to Anthropic's servers — identical to how Claude Code itself operates.

Will installing this break my existing claude-code-telegram setup?

No. install.sh only adds Jarvis-related code to existing files. All your settings, conversation history, and scheduled jobs are preserved. If Jarvis is already installed, it will ask before overwriting.

install.sh throws an error

Most common causes:

1. claude-code-telegram not installed → Run uv tool install claude-code-telegram first
2. Python below 3.11 → Check with python3 --version
3. Can't find site-packages path → The script will prompt for manual input

If unresolved, file an issue at GitHub Issues.

Bot responses are slow

Claude is invoked via the Claude Code SDK, so complex requests can take 30 seconds to 2 minutes. This is Claude's processing time — Mr.Stack's overhead is negligible.

Jarvis alerts are too frequent

Use /jarvis to pause, or increase _TRIGGER_COOLDOWNS values in context_engine.py. During deep work mode, everything except battery warnings is automatically blocked.

Does it use a lot of storage? Will data pile up over time?

Mr.Stack separates short-term and long-term memory.

Short-term (auto-recycled, no disk usage):

• Context snapshots: 12-slot ring buffer in memory → 0 disk
• Session conversation: lives within Claude's context window only

Long-term (disk, but barely grows):

• people/, projects/, preferences/ → Overwrite mode. Files don't grow; always latest state only
• daily/ → 1 file/day, ~1KB. 1 year = ~365KB
• interactions.jsonl → ~300B per conversation. 50/day for 1 year = ~5MB
• routines.json → Periodically refreshed, stays at a few KB

Estimated storage after 1 year: 10–15MB (smaller than a single photo)

Does it consume a lot of tokens?

Mr.Stack adds about 15–20 Claude API calls per day. Negligible on a Max plan.

Component	Token Usage	Details
5-min polling (system scan)	Zero	Local subprocesses only (osascript, pmset, git)
Pattern logging	Zero	Local JSONL file append
State classification	Zero	Simple if-else branching (app name → state mapping)
Jarvis triggers	3–5/day	Only on condition match. Hard cap: 10/hour
Scheduled jobs	10–15/day	Model routing auto-selects Haiku/Sonnet

The most expensive operation (5-min polling) uses zero tokens. Scheduled jobs use model routing to assign Haiku for simple tasks and Sonnet for analysis, minimizing cost. Opus is reserved for evening summaries and weekly reviews only.

Pro plan users should consider disabling memory-sync (every 3h) and github-check (every 2h), or extending their intervals:

# Manage scheduled jobs
~/claude-telegram/schedulers/manage-jobs.sh list              # List all
~/claude-telegram/schedulers/manage-jobs.sh toggle memory-sync   # Toggle off
~/claude-telegram/schedulers/manage-jobs.sh toggle github-check  # Toggle off

Can I use it on multiple Macs?

Install separately on each Mac. To sync memory data, share ~/claude-telegram/memory/ via iCloud or git.

---

Roadmap

• Weekly trend visualization report
• pip install mrstack one-line install

---

Author

@thestack_ai

---

License

MIT