ts4k 0.1.2

Token-efficient messaging gateway for LLM agents

ts4k — Token Saver 4000

Unified messaging gateway that gives LLM agents token-efficient access to messages across Gmail, O365, and WhatsApp. Normalizes, deduplicates, filters, and formats messages for LLM context windows.

Raw HTML email: ~8,000 tokens. After ts4k: ~400 tokens. 20x reduction.

Quick Example

$ ts4k wn
SOURCE|FROM|SUBJECT|DATE|ID|SIZE
g|[e-mail]|Q1 report draft|2026-02-24T09:15:00Z|g:19abc123|2.1kb
g|[e-mail]|Re: lunch tomorrow|2026-02-24T08:30:00Z|g:19def456|340b
o|[e-mail]|Budget approval|2026-02-24T10:00:00Z|o:AAMkAD789|1.8kb
w|[phone]|Hey, running late|2026-02-24T09:45:00Z|w:3EB05C42|120b

One command, three platforms, pipe-delimited output an LLM can parse in ~60 tokens.

Install

pip install -e .

Requires Python 3.12+. Creates two entry points:

• ts4k — CLI
• ts4k-mcp — MCP server for Claude Code and other MCP-compatible agents

Provider Setup

Each messaging platform needs a one-time setup. Pick the ones you use:

Provider	Guide	Auth Method	Complexity
Gmail	docs/setup-gmail.md	Google OAuth (browser)	Medium
O365	docs/setup-o365.md	Azure device code flow	Medium
WhatsApp	docs/setup-whatsapp.md	Local SQLite (no auth)	Low

You can run any combination. One provider is enough to get started.

Usage

See docs/usage.md for the full command reference, MCP server mode, output formats, and examples.

Everyday Commands

ts4k wn                        # What's new across all sources
ts4k wn --source g             # Gmail only
ts4k g g:19abc123              # Read a specific message
ts4k t g:thread456             # Read a thread
ts4k l -q "from:alice" -n 10   # Search messages
ts4k o                         # Overview of cached messages
ts4k st                        # Health + token savings stats

MCP Server Mode

ts4k-mcp                              # stdio (default, for Claude Code)
ts4k-mcp --transport http --port 9000  # HTTP for other clients

Exposes the same commands as MCP tools: updates, get, thread, list, overview, status, and more.

Architecture

Agent (Claude, etc.)
  |
  v
ts4k (normalize --> filter --> format)
  |-- Gmail Adapter    --> Google Gmail API
  |-- O365 Adapter     --> Microsoft Graph API
  |-- WhatsApp Adapter --> whatsapp-mcp (local SQLite)
  '-- Future adapters  --> Slack, Teams, Telegram, ...

• Adapters wrap platform APIs. ts4k doesn't reimplement them.
• Normalize strips HTML, deduplicates reply chains, collapses whitespace.
• Filter applies skip lists (senders, domains, patterns). Off by default.
• Format outputs pipe-delimited (default), JSON, or XML.

Platform failures are isolated — if one adapter is down, the others still return results.

Key Principles

• Metadata first, content on demand — default to minimum useful response
• No LLM calls inside ts4k — this is the data layer, not the intelligence layer
• Using a command IS the side effect — watermarks update on whatsnew, no separate save step
• Format is a feature — pipe-delimited saves ~60% tokens vs JSON

Configuration

All state lives in ~/.config/ts4k/ (override with TS4K_CONFIG_DIR):

~/.config/ts4k/
  sources.json              # Configured providers
  watermarks.json           # Last-fetched timestamps
  contacts.json             # Cross-platform identity map
  filters.json              # Skip lists
  stats.json                # Usage counters
  cache/                    # Local message cache
  google/{email}/token.json # Gmail OAuth tokens
  microsoft/{id}/token_cache.json  # O365 MSAL tokens

Tech Stack

Python 3.12+, Google API client, MSAL, httpx, MCP SDK, html2text, beautifulsoup4.

License

MIT