LangFord

Language (Model) For Organizational Reasoning & Delegation
A Telegram-first AI assistant that can actually do things: read and summarize news, draft replies, create calendar entries, pull information from your email, and more — by choosing the right tools on its own.

---

What it is

LangFord is a lightweight agent runtime that lives in Telegram. You chat with it like a normal assistant, but under the hood it can call real integrations (Calendar, email, web) to complete requests end-to-end.

Think: “Hey, remind me tomorrow 9:00 to call Alex and add it to my calendar” → it schedules it.
Or: “Summarize today’s headlines about AI regulation and send me the 5 key points” → it fetches, filters, and summarizes.

---

Highlights

• Telegram-native UX: no extra UI, no dashboards — just messages.
• Tool selection done by the model: you don’t need to remember commands like /calendar_add.
• Practical workflows:
  • create calendar events
  • email reading / extracting action items
  • web/news digest
  • quick research with sources (when enabled)
  • “do this for me” style delegation with guardrails
• Composable tools: adding a new capability is “just another tool”, not a rewrite.
• Provider-flexible: designed around an LLM client + an agent loop (works with common OpenAI-compatible endpoints).

---

Example chats

Calendar

• “Schedule ‘Project sync’ next Tuesday 14:30 for 45 minutes.”
• “Move my dentist appointment to Friday afternoon.”
• “What do I have tomorrow morning?”

Email

• “Check my inbox for anything urgent from HR and summarize it.”
• “Find the last email about ‘invoice 2026-02’ and tell me what’s needed.”
• “Draft a reply: confirm receipt and ask for the missing attachment.”

News

• “Give me a morning briefing: AI, geopolitics, finance — short.”
• “Summarize the latest on [topic] in 5 bullets and link sources.”

General assistant

• “Turn this message into a polite reply.”
• “Extract tasks from this text and create reminders.”

---

How it works (in one minute)

LangFord runs an agent loop:

1. You send a message in Telegram.
2. The agent decides whether it can answer directly or needs a tool.
3. If needed, it calls tools (calendar/email/web) with structured inputs.
4. Results are stitched into a final response, with the action completed.

The goal is to keep the interface simple (chat), while making the backend capable (tools).

---

Project layout (high level)

Folder names may evolve, but the structure is intentionally simple.

• interface/ – Telegram bot entrypoint + message handling
• agents/ – agent orchestration / routing logic
• tools/ – integrations (calendar, email, web scraping, etc.)
• prompts/ – system prompts + tool instructions
• core/ – shared utilities (config, clients, helpers)

---

Quickstart

1) Install

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

2) Configure environment

Create a .env file (or export env vars). Typical values:

• TELEGRAM_BOT_TOKEN – your bot token from BotFather
• OPENAI_API_KEY (or your OpenAI-compatible key)
• Optional: credentials for Google / Microsoft integrations (depending on which tools you enable)

If you don’t configure calendar/email, LangFord still works as a chat assistant — it just won’t execute those actions.

3) Run

python -m interface.telegram_bot

---

Tools & integrations

LangFord is built to support multiple tool backends. Out of the box (or by default templates), you’ll typically see:

• Web/news via HTML extraction + article parsing
• Google integrations (Calendar / Gmail) via OAuth + Google APIs
• Microsoft integrations (Graph) via MSAL (optional, if enabled)

If you don’t want an integration, disable the tool — the agent will stop using it.

---

Adding a new tool

1. Create a new tool module in tools/ (keep it small and single-purpose).

2. Define:

   • what the tool does
   • the inputs it expects
   • the output shape (keep it predictable)

3. Register it with the agent/router so the model can discover it.

4. Add a short usage note in prompts/ so the model knows when to call it.

A good tool is boring: deterministic inputs, deterministic outputs, clear failure modes.

---

Safety & guardrails

This project is meant to be useful, not reckless. Recommended defaults:

• confirmation before destructive actions (delete/move/cancel)
• explicit scopes for email/calendar access
• logging of tool calls (so you can audit what happened)
• rate limits to avoid runaway loops

---

Roadmap ideas

• Daily scheduled briefing (news + calendar overview)
• “Inbox to tasks” mode (extract tasks → calendar/reminders)
• Personal knowledge base (lightweight memory with opt-in)
• More tool adapters (Notion, Slack, Jira, GitHub, etc.)

---