tg keeps the runtime deliberately small: configuration, named sessions, authentication, locking, and Python process semantics. Telethon remains the Telegram API.

Telethon is the API; tg only provides the authenticated execution boundary. When a workflow is missing, write the missing logic as ordinary Python and run it through tg.

agent wants something in Telegram
        │
        ▼
      tg
        │
        ├── client.*        friendly Telethon methods
        └── functions.*     raw Telegram API when needed

Three commands plus direct Python execution.

tg login
tg doctor
tg skill
tg script.py

The Python distribution is tg-harness. The installed command is tg.

Give it to your agent

Install from PyPI:

uv tool install tg-harness

Or install the current GitHub version:

uv tool install git+https://github.com/speech115/tg.git

Then give the agent this instruction:

Use tg for Telegram. Run tg doctor first. If the account is not authorized, ask
me to complete tg login. For Telegram work, use one tg program per decision
boundary, prefer Telethon client methods, and fall back to functions.* / types.*
for raw Telegram requests.

Requires Python 3.12+ and a POSIX system (macOS or Linux).

Configure once

Create Telegram API credentials at https://my.telegram.org/apps, then create ~/.config/tg/config.toml:

[telegram]
api_id = 123456
api_hash = "your-api-hash"

Alternatively, set TG_API_ID and TG_API_HASH in the environment.

Set TG_CONFIG when the config lives elsewhere.

Authorize the default account:

tg login
tg doctor

The default account is main. Named accounts map directly to Telethon session files:

tg --account work login
tg --account work doctor
tg --account work script.py

~/.local/state/tg/
├── main.session
├── work.session
└── another.session

Account names must match [A-Za-z0-9_-]+.

Run ordinary Python

For a one-off task:

tg <<'PY'
dialogs = await client.get_dialogs(limit=10)
for dialog in dialogs:
    print(dialog.name)
PY

For reusable logic:

tg script.py arg1 --flag
tg --account work script.py arg1 --flag

Every run gets:

client  # authenticated Telethon client
functions  # raw Telegram request constructors
types  # raw Telegram types
account  # selected named account

It also gets normal __file__, sys.argv, and local-import behavior.

Prefer the friendly API when it fits:

messages = await client.get_messages("me", limit=20)

Drop to the raw API when it does not:

result = await client(functions.users.GetFullUserRequest(id=types.InputUserSelf()))

How it works

                            one tg process
                                   │
                     authenticated Telethon client
                                   │
               ┌───────────────────┴───────────────────┐
               │                                       │
          client.* helpers                      raw TL requests
               │                                functions.* / types.*
               └───────────────────┬───────────────────┘
                                   │
                              Telegram API

config      ~/.config/tg/config.toml (or TG_CONFIG)
sessions    ~/.local/state/tg/<account>.session
locking     one process per named session

Workflow logic stays in ordinary Python scripts.

Agent skill

The repository ships skills/tg/SKILL.md.

Use tg skill to print the bundled instructions. Its main rule is simple: bundle deterministic operations into one tg process and stop only at a real decision boundary. That avoids reconnecting for every API call and keeps agent behavior both faster and simpler.

Trust boundary

tg is intentionally not a sandbox.

Code passed to it has the permissions of the selected Telegram account and can read, send, edit, delete, download, join, leave, and perform raw Telegram API operations.

Treat these as secrets:

• api_hash
• Telethon .session files
• any exported authorization material

The runtime keeps sessions outside the repository and serializes access to each named session with a lock.

See CONTRIBUTING.md for development and integration instructions.

License

MIT. See LICENSE.