Summarix

💡 The Problem

You follow thirty channels because each of them is sometimes worth it. In practice that means unread counters in the hundreds, the one post that mattered buried between ads and cross-promotion, and an hour a day of scrolling to find out that most of it was not news.

Summary bots promise to fix this and run into the same walls: they cannot read private channels, they skip images and charts, they lose posts when something fails, and their summaries come with no way to check what they are based on.

Summarix reads everything you follow and hands you the part worth your time — with the receipts.

🔥 Why it wins

🎬 What you get

An illustrative digest (channels and news are made up):

Digest · 13.09.2026Channels with news: 3 · posts: 27
Highlights• OpenAI shipped a cheaper vision model (AI Weekly)• Central bank kept the key rate at 16% (Markets)• ...
AI Weekly• A new small vision model is priced 4x below the previous one; benchmarks in the post show it on par for OCR (1, 4)• Two posts about an agent framework release - mostly a changelog (2, 3)
Markets• ...

Numbers in brackets are links to the posts. The model only cites post numbers; the bot turns them into links. You can check every line in one tap.

⚙️ How It Works

flowchart LR    A[📡 Your channels<br/>Telethon · your account] --> B[🆕 New posts<br/>since the last delivered digest]    B --> C[🖼 Image captions<br/>vision model · cost cap]    C --> D[✍️ Per-channel summary<br/>2-6 bullets citing post numbers]    D --> E[⭐ Highlights<br/>top 3 across channels]    E --> F[📬 One message<br/>then posts are marked read]

every day at DIGEST_TIME                      /digest, /digest 48          │                                          │          ▼                                          ▼   for each channel: new posts since the last delivered digest          │   Telethon, your account (private channels, albums)          ▼   images ──▶ vision model: one-line caption   (capped per run)          ▼   posts + captions ──▶ model: 2-6 bullets citing [post numbers]          ▼   2+ channels ──▶ model: top 3 highlights          ▼   HTML message(s) to you ──▶ only then posts are marked read

• Nothing is lost on failure. Read marks move after the digest is delivered. A crash, an API error or a failed send means the same posts come next time.
• One broken channel does not break the digest. It is listed at the bottom as "could not read".
• Ads and cross-promotion are skipped by the summary prompt.

Built to be trusted on a bad day, not just on a demo.

✨ Features

🧭 Three ways to use it

One engine underneath — Summarix reads your channels through your own account with Telethon. The three ways differ in who writes the retelling and where you read it.

Under the skill and the MCP server sits scripts/tgdigest.py — a CLI you can also run yourself. The bot has its own session and channel list; the skill and the MCP server share one, one process at a time, guarded by an OS file lock.

Run all three: the bot sends the morning digest, and Claude answers "what did my channels say about smart glasses this month".

🚀 Quick Start

You need four things:

# 1) Clonegit clone https://github.com/AndreySukhanov/summarix-telegram.gitcd summarix-telegram
# 2) Keyscp .env.example .env            # fill in the keys
# 3) Once: log in to your Telegram account (phone, code, 2FA password)docker compose run --rm bot python -m scripts.login
# 4) Rundocker compose up -d

Docker pulls the ready image ghcr.io/andreysukhanov/summarix-telegram (amd64 and arm64: a VPS, a Raspberry Pi, a Mac) — nothing to build. docker compose build builds it from source instead.

Without Docker, as one command: pipx install summarix-telegram, then in a folder with your .env run summarix-login once and summarix-bot (or point SUMMARIX_HOME at that folder). From a clone: pip install -r requirements.txt, then python -m scripts.login and python run_bot.py. On Windows, start_bot.bat starts it and refuses to start a second copy.

Then open your bot, press Start, and add channels.

💬 Commands

The bot menu has buttons for all of it — a digest per group, a style switch, the channel list.

🤖 Claude Code plugin

Read your channels without leaving the terminal. Summarix ships as a Claude Code plugin with a summarix-telegram skill: ask in plain words — "digest of my channels for 3 days", "roast style, last week", "what did my channels say about smart glasses", "find channels about AI agents" — and Claude collects the posts, retells them with a link to every post and answers follow-ups.

/plugin marketplace add AndreySukhanov/summarix-telegram/plugin install summarix-telegram@summarix-telegram

Prefer a plain skill without the plugin? Copy skills/summarix-telegram into ~/.claude/skills/ - it is the same file.

The skill works on your clone of this repo: set SUMMARIX_HOME to it (or tell Claude once where it is), put TG_API_ID / TG_API_HASH in .env, and run the one-time login in your own terminal — python scripts/tgdigest.py login. No model key needed: Claude writes the digest itself, in any of the 13 styles.

It has its own Telegram session and its own channel list, and it only reads a recent window — it never moves the bot's read marks. Run the bot and the plugin side by side; nothing collides.

🪟 The digest beside the chat (optional mod)

On Claude Code 2.1.287 or later, one more install gives the digest a pane of its own:

/plugin install summarix-panel@summarix-telegram

When the skill finishes a digest, Claude also puts it in the pane - every point still linked to its post. < Older / Newer > page through the last 10 digests, kept across sessions; 24 h / 3 days / Week ask for a new one in one press; /summarix-panel opens the pane at any time. The mod needs no keys and reads nothing from Telegram itself. Details: mods/summarix-panel.

python scripts/tgdigest.py login                    # once: phone, code, 2FApython scripts/tgdigest.py add t.me/addlist/slug    # a folder, @name, t.me/name, t.me/+invitepython scripts/tgdigest.py fetch --hours 24 --style tldrpython scripts/tgdigest.py search "smart glasses" --days 30python scripts/tgdigest.py find "smart glasses" --top 10python scripts/tgdigest.py styles

🔌 MCP server

Not on Claude Code? The same collector runs as an MCP server, so any MCP client — Claude Desktop, Cursor, Claude Code — gets your channels as tools. The client's model writes the digest; the server needs no model key.

# Claude Codeclaude mcp add summarix-telegram -- /path/to/summarix-telegram/.venv/bin/python /path/to/summarix-telegram/scripts/mcp_server.py

Claude Desktop, in claude_desktop_config.json:

{  "mcpServers": {    "summarix-telegram": {      "command": "/path/to/summarix-telegram/.venv/bin/python",      "args": ["/path/to/summarix-telegram/scripts/mcp_server.py"]    }  }}

On Windows the interpreter is .venv\Scripts\python.exe.

Installed with pipx install summarix-telegram? Then the server is one command, and SUMMARIX_HOME is the folder with your .env and data/:

claude mcp add summarix-telegram -e SUMMARIX_HOME=/path/to/folder -- summarix-mcp

From the MCP Registry (io.github.AndreySukhanov/summarix-telegram): a client that runs local stdio servers starts it with uvx — install uv once, the rest is fetched on first run. Log in once with uvx --from summarix-telegram tgdigest login and the same SUMMARIX_HOME:

{  "mcpServers": {    "summarix-telegram": {      "command": "uvx",      "args": ["summarix-telegram"],      "env": { "SUMMARIX_HOME": "/path/to/folder" }    }  }}

It shares the session and the channel list with the skill - one process at a time, guarded by an OS file lock. Log in once in your own terminal with python scripts/tgdigest.py login — the login never goes through MCP. Every result with posts is marked as untrusted content: a post is data, not instructions.

One server, any MCP client, your channels.

💸 Cost

Per digest: one call per image (sent at low detail), one call per channel with news, one call for highlights. Twenty channels with a few posts each and the image cap at 20 means about 40 calls a day; on a small model that is usually a few cents a day — /status shows the real token count.

The cheapest hour you will save today.

🔐 Privacy and security

• The session file (data/owner.session) is full access to your Telegram account. It stays on your machine, is git- and docker-ignored, and should never be shared or committed.
• The bot answers only OWNER_ID; every other update is dropped before any handler runs.
• The account is used read-only: it never posts, joins or leaves anything. Channels behind an invite link must be joined by you first.
• Post texts and images are sent to the model provider you configure. Pick one whose data policy you accept, or point LLM_BASE_URL at a local model.
• Storage is a local SQLite file with channel names, groups and read marks only; post contents are not stored.

Your feed stays yours: nothing leaves your machine except what you send to the model you chose.

summarix_telegram/  bot/    main.py           wiring: settings, SQLite, Telethon, model, bot, scheduler, "/" menu    handlers.py       commands, groups, digest per group, search and find    settings_menu.py  /settings: style, prompt, languages, time, timezone, layout, word filter    keyboards.py      reply and inline keyboards    middlewares.py    owner-only gate    texts.py          interface strings in five languages  services/    reader.py         Telethon: resolve channels and folders, fetch posts, albums, photos    llm.py            prompts: image captions, channel summary, highlights, grouping by topic    styles.py         13 styles, each with a prompt per digest language    digest.py         build a digest by channel or by topic, mark read after delivery    export.py         a Markdown copy of each digest with news    search.py         /search query expansion and /find channel discovery    scheduler.py      daily run in your timezone    refs.py           parse @name, t.me links, invite and folder links  database/           channels, groups, digest runs, settings  scripts/    login.py          one-time account login    tgdigest.py       collector behind the Claude Code skill    mcp_server.py     MCP server: the collector as tools  utils/text.py       citations → links, safe HTML, splitting at 4096 chars, the word filter  cli.py              summarix-bot and summarix-login of the pipx packagescripts/              the old paths keep working: scripts/tgdigest.py, scripts/mcp_server.py, python -m scripts.loginskills/summarix-telegram/  the Claude Code skillmods/summarix-panel/  the optional panel mod for Claude Code.claude-plugin/     plugin and marketplace manifeststests/              offline: fake Telegram reader and fake model

🧪 Tests

pip install -r requirements.txtpytest

140 tests, no network: link and folder parsing, album grouping, citation linking and HTML escaping, message splitting, schedule in local time, read marks only after commit, recap windows, image cost cap, broken channels, groups, search and discovery, settings, the tgdigest companion, the MCP server (schemas and bounds, parts that fit, no stale results, locks that survive cancellation, safe errors, an isolated protocol stream), owner-only access. CI runs them with a flake8 pass on every push to main.

❓ FAQ

How do I summarize Telegram channels with AI?

Run Summarix: add the channels you follow (/add, or a Telegram folder link) and the bot sends one digest a day — 2-6 bullets per channel, the day's highlights first, every bullet linked to its post. /digest gives one on demand.

Can it merge the same news from several channels?

Yes. In /settings choose Layout → By topic: a story told by three channels becomes one point with links to all three posts. The code checks the grouping — every cited fact must stay and no post may be added — and sends the digest by channel if the check fails twice.

Can it summarize private Telegram channels?

Yes. It reads through your own Telegram account (Telethon), so any channel you have joined works, private ones included. A regular bot cannot do that on its own.

Is there an MCP server for Telegram channels?

Yes — scripts/mcp_server.py gives any MCP client (Claude Desktop, Cursor, Claude Code) your channels as eight tools: fetch posts, read further parts, search, find new channels, manage the list. See MCP server.

Can I use it in Claude Code?

Yes — install the summarix-telegram plugin with two /plugin commands and ask in plain words: "digest of my channels for 3 days". See Claude Code plugin.

Do I need an OpenAI API key?

For the bot, yes — any OpenAI-compatible API works: OpenAI, OpenRouter, DeepSeek, Qwen, or a local model behind LLM_BASE_URL. The Claude Code skill and the MCP server need no model key: the client's model writes the digest.

Does it post, join or leave channels on my behalf?

No. The account is used read-only, and the bot answers only its owner (OWNER_ID).

What does it cost?

Running it for yourself costs only model tokens — about 40 calls a day for twenty channels, usually a few cents on a small model. The code is Source-Available: free to read, run and learn from.

Which languages does it support?

Summaries in Russian, English, Spanish, French and Kazakh, whatever language the channels write in; the bot interface speaks the same five.

What does it store?

A local SQLite file with channel names, groups and read marks. Post contents are not stored.

🤝 Contributing

The bot is built from small, replaceable parts. Great first contributions:

• A new digest style — one entry in summarix_telegram/services/styles.py
• A new language — interface strings in summarix_telegram/bot/texts.py, a prompt per style
• A reader feature, tested against the fake Telegram client in tests/fakes.py

See CONTRIBUTING.md. Open an issue first for bigger changes — happy to discuss the design.

⭐ Support the project

If this project saved you time or showed you something new — star it. It genuinely helps others discover it.

💜 Sponsor development

This is a self-funded project — every digest, caption and search runs on paid model APIs, and every new style is tested on real channels before it ships. A donation goes straight into model credits and new features.

TERj89Gjf4qsMW7xyKQWGWtNmBQaDUPnnH

Thank you — even a small contribution keeps the digests coming. 🙏

📄 License

Source-Available — free to read, run and learn from. Commercial or derivative use requires written permission. See LICENSE.

Licensing contact: [email protected]