TG Agent

Telegram automation for coding agents — Claude Code, OpenCode, Codex. Ships as a Claude Code plugin (skill + CLI): install the plugin and your agent can act on Telegram for you — read and send messages, manage channels and chats, download media, run LLM content pipelines. The content factory (collect → search → generate → publish) is one module among many.

Русская версия

How it works

• Skill + Claude Code plugin — this repo ships a skill (installed from the plugin marketplace, or copied manually) that teaches your agent the command catalog, the runtime model and safety rules. The skill drives the CLI — no MCP tool surface eating your context window.
• tg-agent CLI — the product surface: ~250 commands covering the full Telegram spectrum (channel, dialogs, messages, search, pipeline, photo-loader, analytics, scheduler, account, …). Every command is a self-contained one-shot: it opens a Telegram connection, does the work, exits. No daemon required.
• tg-agent worker — optional long-lived process for scheduled collection and queued sends.

Quick Start (coding agents)

Prerequisites

• Python 3.11+
• Telegram API credentials from my.telegram.org/apps

1. Install the plugin (Claude Code) — or the skill alone

/plugin marketplace add https://github.com/axisrow/tg_content_factory
/plugin install tg-agent@tg-agent-marketplace

The short form (axisrow/tg_content_factory) clones over SSH — without GitHub SSH keys it fails; use the full HTTPS URL as above. If the marketplace is not visible right after marketplace add, run /reload-plugins before plugin install.

OpenCode / Codex / other agents — skip the plugin: point the agent at skills/tg-agent/SKILL.md as an instruction file (AGENTS.md include, system-prompt attachment, etc.), or copy the skill folder:

cp -r skills/tg-agent ~/.claude/skills/tg-agent

2. Install the CLI

The skill drives the tg-agent CLI, which must be on the machine where your agent runs:

pip install tg-agent

Create .env in your working directory (auto-loaded by every command):

TG_API_ID=your_api_id
TG_API_HASH=your_api_hash
SESSION_ENCRYPTION_KEY=your_random_secret    # required before adding/importing accounts

For several projects on one machine you can skip the per-project key copies: store it once in the OS secret store with tg-agent key set (macOS Keychain / Windows Credential Manager / secret-tool). An explicit SESSION_ENCRYPTION_KEY still wins over the stored secret.

3. Authorize an account (interactive, once)

tg-agent account add

You will receive a Telegram login code — enter it yourself. Verify with tg-agent account list.

4. Just ask your agent

• "Show yesterday's posts from @durov"
• "Send this draft to my Saved Messages"
• "Collect new messages from my channels and find mentions of <keyword>"

Features

• Built for coding agents — the CLI is the contract: new capabilities land as CLI commands first (tested there), other surfaces follow only if at all
• All chat types — channels, supergroups, gigagroups, forums, public and private
• Multi-account with automatic flood-wait rotation
• 3 search modes — local DB (FTS5), direct Telegram API, AI/LLM-powered
• Scheduled collection — incremental fetching; runs in the background tg-agent worker
• Keyword monitoring — plain text and regex, with Telegram bot notifications
• Content factory — LLM pipelines: generate → moderate → publish, image generation included
• Built-in anti-spam filters — deduplication, low-uniqueness detection, cross-channel spam, subscriber-ratio and non-Cyrillic filters
• Analytics — top posts, trends, activity heatmaps, trending topics and emojis
• Security — session encryption (Fernet + PBKDF2), HMAC-signed web session cookies
• Docker-ready

Legacy surfaces

These still work but are frozen: development is paused indefinitely, and new capabilities must not be built on them. The CLI + skill is the only actively developed interface.

• Web dashboard (FastAPI + Bootstrap 5) — python -m tg_agent.main serve, then http://localhost:8080 (password from WEB_PASS). It is the optional tg-webui package (github.com/axisrow/tg_webui, PyPI: pip install tg-webui; requires tg-shared, currently not on PyPI); pip install tg-agent alone has no web panel
• TUI and the embedded agent chat (tg-agent agent chat; claude-agent-sdk / deepagents backends). The interactive TUI is the optional tg-tui package (github.com/axisrow/tg_tui, PyPI: pip install tg-tui); one-shot agent chat -p "..." works without it
• MCP server (python -m tg_agent.main mcp-server)

Legacy: split deployment (Docker / k8s)

serve spawns an embedded Telegram worker inside the same process by default. For split deployments pass --no-worker and run a dedicated worker service:

# container 1 — web UI + API only
python -m tg_agent.main serve --no-worker

# container 2 — Telegram worker (shared SQLite volume)
python -m tg_agent.main worker

Docker

cp .env.example .env
# fill in your credentials
docker-compose up -d

The image is worker-only (automation-first): it runs the Telegram worker runtime — collection, scheduler, queues — with the managed PID file and no web surface. The legacy web panel is not part of the image; it is the separate tg-webui package (pip install tg-webui, needs tg-shared on PyPI).

Semantic Search Roadmap Note

The current semantic and hybrid search implementation was originally built around runtime sqlite-vec loading. That turned out to be too fragile as a mandatory foundation: installing the sqlite-vec package alone is not enough, because the active Python/SQLite build must also support sqlite3.enable_load_extension(...). In practice, the same pip install can therefore produce different operator outcomes across machines, including "package installed but semantic search unavailable."

The roadmap is being corrected toward a portable SQLite-first semantic backend that works on standard Python builds without enable_load_extension. Until that backend lands, treat sqlite-vec as a transitional dependency rather than a guaranteed feature toggle. The public UX stays the same: semantic indexing, semantic search, and hybrid search remain the target interface.

See docs/semantic-search.md for the architecture note, migration story, and rationale for de-emphasizing mandatory sqlite-vec.

Configuration

Environment Variables (.env)

Variable	Required	Description
TG_API_ID	Yes	Telegram API ID
TG_API_HASH	Yes	Telegram API Hash
SESSION_ENCRYPTION_KEY	Yes*	Key for encrypting Telegram session strings in DB; may instead live in the OS secret store (tg-agent key set)
WEB_PASS	—†	Web panel password (legacy web dashboard only)
LLM_API_KEY	No	API key for AI-powered search
ANTHROPIC_API_KEY	—†	claude-agent-sdk only (legacy embedded agent chat)
CLAUDE_CODE_OAUTH_TOKEN	—†	Claude Code auth token for claude-agent-sdk (legacy)
AGENT_MODEL	—†	Claude SDK model override (legacy embedded agent chat)
AGENT_FALLBACK_MODEL	—†	provider:model for deepagents fallback (legacy)
AGENT_FALLBACK_API_KEY	—†	Explicit API key for the legacy fallback provider

* Required to add or import accounts and decrypt stored sessions. New sessions are always encrypted (enc:v2:); plaintext writes are refused. Generate a secret with python -c "import secrets; print(secrets.token_urlsafe(32))", keep it private, and retain it to read existing sessions. No config.yaml is needed: the key is read directly from .env.

\† Legacy-only: needed solely by the legacy web panel and embedded agent chat (see Legacy surfaces).

config.yaml

Supports ${ENV_VAR} substitution. Empty env vars are dropped (defaults apply).

Section	Description
telegram	API credentials (api_id, api_hash)
web	Host, port, password (default: 127.0.0.1:8080; non-loopback host requires a strong WEB_PASS) — legacy web panel
scheduler	Collection interval, delays, limits, max flood wait
notifications	admin_chat_id for keyword match alerts
database	SQLite path (default: data/tg_search.db)
llm	LLM provider, model, API key for AI search and content pipelines
agent	Legacy embedded agent chat settings
security	Session encryption settings

Legacy: embedded agent backend rules

• /agent uses claude-agent-sdk when ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN is configured.
• If Claude SDK is not configured, /agent falls back to deepagents when AGENT_FALLBACK_MODEL is set.
• ANTHROPIC_API_KEY and CLAUDE_CODE_OAUTH_TOKEN are never reused by deepagents.
• Developer override for forcing claude-agent-sdk or deepagents lives on the Settings page and applies only when developer mode is enabled.

CLI reference (selected)

tg-agent restart                                 # managed daemon: worker runtime, no web panel
tg-agent worker                                  # same runtime without the stop-first step
tg-agent channel collect --channel-id ID         # one-off incremental collection (no daemon)
tg-agent search "query" --limit 20               # search collected history
tg-agent messages read @channel --format json    # read message history
tg-agent dialogs send                            # real actions in real chats
tg-agent pipeline generate                       # LLM content factory
tg-agent serve                                   # legacy web panel (deprecated)

Full catalog for agents — skills/tg-agent/reference.md; every group also has --help.

telethon-cli

telethon-cli is installed with the project and reuses the same TG_API_ID and TG_API_HASH values from .env.

Optional CLI-only overrides:

• TG_SESSION sets a custom Telethon session path or name.
• TG_PASSWORD supplies the Telegram 2FA password for non-interactive runs.

Legacy TELETHON_* environment variable names are still accepted by telethon-cli for compatibility, but this project standardizes on TG_*.

telethon-cli login
telethon-cli users get-me --output json

Web Interface (legacy)

Page	Path	Description
Web login	/login	Sign in to the web panel with WEB_PASS
Dashboard	/	Stats, scheduler status, connected accounts
Telegram auth	/auth/login	Add Telegram accounts (phone + code + 2FA)
Accounts	/accounts	Manage connected accounts
Channels	/channels	Add/remove channels, keywords, import
Search	/search	Search messages (local / Telegram / AI)
Analytics	/analytics	Top posts leaderboard, engagement by content type, hourly patterns
Filters	/filter	Anti-spam filter report and controls
Scheduler	/scheduler	Start/stop/trigger collection and keyword search
Agent	/agent	Legacy embedded AI chat

Roadmap

• Portable semantic search on stock Python installs without mandatory runtime SQLite extension loading
• Agent-facing capability growth: every new feature lands in the CLI (and the skill) first
• LLM-powered content factory
• LLM-powered intelligent search
• LLM-based chat spam moderation
• Direct message handling
• Telegram action automation (broadcasts, etc.)

Development

# Install dev dependencies
pip install -e ".[dev]"

# Run parallel-safe tests (up to four workers locally)
pytest tests/ -v -m "not aiosqlite_serial" -n auto

# Keep aiosqlite-backed tests sequential within each file; run independent files in parallel
pytest tests/ -q -m aiosqlite_serial -n auto --dist=loadfile

# Run a single test
pytest tests/test_local_install_consistency.py -v

# Benchmark serial vs safe mixed-mode suite execution
python -m tg_agent.main test benchmark

 # Lint
 ruff check tg_agent/ tests/ packages/ conftest.py

CI Note

• push workflow checks the branch head only.
• pull_request workflow checks the merge result against main.
• A branch can therefore be green on push and red on pull_request if main introduced a lint/test failure that is pulled into the PR merge ref.
• Before rerunning PR checks, fetch and sync with origin/main so local verification matches CI.