Enloq

A Telegram client that helps people read busy conversations and write clearer replies.

Enloq combines incoming messages into readable blocks, turns voice notes and images into clear text, softens hostile messages without losing their content, suggests complete replies that can build on your draft, and turns group conversations into daily briefings with links back to their sources. The original messages remain available, and sending is always a separate, deliberate action.

The product has a React web client and an Expo / React Native app, backed by a Python service. Enloq comes from en (縁, human connection) and loqui (Latin, “to speak”).

Open any screenshot at full size. These are high-resolution captures of the real phone-width web interface, using authored synthetic conversations and prepared outputs. They demonstrate presentation, not model quality, live Telegram delivery or native-device acceptance.

What it does

• Read with less friction. Show incoming text together as it was sent, and in chats with Voice and images On, rewrite voice notes, video circles and images into clear text. With softening On, a block with hostile wording is softened, keeping every fact, request, deadline and threat, and marked Softened; frustration or anger alone does not trigger it. Switch between AI text and originals; retain original media, observed edits, replies and source navigation.
• Write with control. Press Suggestions for up to three complete messages. A started draft is the idea they develop, and text in [square brackets] instructs the writer without appearing in them. Choosing one puts it in the draft for editing, Undo restores the previous draft, and nothing is sent until you press Send, which asks first while bracketed text remains. When the latest block shows a clear emotion or calls for a reply, suggestions also appear on their own and can offer Accept, Decline and Later stances. Each chat's AI Menu, opened by the robot in its header, switches automatic suggestions, emotion, softening, voice and images and memory, all Off until you turn them On for that chat, and sets the suggestions' style: Friendly, Professional or Idioms.
• Remember what matters. In chats with memory on, note what each person and the owner care about for the suggestions to read, and collect the owner's traits and habits for them alone.
• Catch up on groups. Opt individual groups into daily briefings, follow citations to exact messages, and return to the briefing. Ask a group a question and get an answer from its own messages with source links, or a plain statement that they hold none. Saved Messages has its own tab and does not go through AI processing.
• Keep ordinary messaging useful. Switch accounts, find people, start a private conversation, reply, and work with photos, albums, GIFs, stickers, voice notes and videos. Recent background processing covers seven days.
• Use web and mobile together. Pair trusted browsers and phones with the service. Approved members have their own accounts, preferences and budgets; the owner manages access and server-wide AI settings.

Architecture

Layer	Implementation	Responsibility
Service	Python 3.13, FastAPI, Pydantic, Telethon, SQLite	Telegram session ownership, sync, durable work, sending, access checks and spend accounting
Web	TypeScript, React, Vite	Reading, composing, streamed updates, pairing and settings
Mobile	TypeScript, React Native, Expo	iPhone and Android clients, device trust, media and notifications
Shared client rules	Framework-independent TypeScript in shared/	Routes, times, reply presentation and reading rules
AI and speech	A model proxy, bounded inference roles and Whisper transcription	Rewriting, suggestions, classification, briefings and speech, with explicit routes and budgets

The deployed architecture places the service on an owner-managed VPS behind HTTPS. Telegram authorizations, message caches and durable work stay with that service; browsers and phones are clients, not independent Telegram session owners. The former Mac service is a separate fallback, not the home of all current accounts or trust. Operations describes hosting, pairing and recovery; the mobile runbook describes builds and platform-specific delivery limits.

Member access is scoped in the application. The operator of the shared host can access its stored data, as disclosed in the member agreement. This is not an end-to-end encrypted AI processing boundary. Provider handling and current model routes are documented in Inference and ADR 39.

Engineering decisions

Generating and sending have different lifecycles. Incoming generation waits for a quiet window. Streamed text is provisional until final validation, and suggestions appear only as complete sets. A chosen suggestion only changes the editable draft; it cannot send a message. Account processing controls are independent of the reading view.

A decision model decides before a writer is called. Perplexity's Decider answers named questions with probabilities and writes no text. Its answers decide whether a text block is softened, the only rewrite of incoming text, whether automatic suggestions are prepared, and which group messages briefings leave out as ads or bots. Measurements gives its per-call latency and cost from small checks, with their limits.

An uncertain send is not safe to repeat. Sends are ordered per chat and retain their operation identity, intended account, peer and media. Failed and uncertain outcomes remain visible with explicit recovery actions. Checking delivery observes the original operation rather than sending another copy. See media delivery.

This is the beginning of the README. Read the rest at the original. Read the full README ↗