Channels & Gateway

Run a gateway

hara gateway is an opt-in daemon that lets you drive your local hara from a chat app. Each inbound message spawns a headless run on that chat's own resumable session; the reply comes back in chat. The daemon connects out (long-poll or WebSocket), so it needs no public webhook — it runs fine on your laptop behind NAT.

hara gateway --platform <name>

Platforms at a glance

Platform	Transport	Images	Notes
Telegram	long-poll	✅ two-way	token from @BotFather
WeChat	iLink long-poll	✅ two-way	QR login
Discord	WebSocket	✅ two-way	needs Message Content Intent
Feishu / Lark	WS long-connection	✅ two-way	p2p DMs (v1)
Slack	Socket Mode	✅ two-way	App + Bot tokens
Mattermost	WebSocket + REST	✅ two-way	self-host or cloud
Matrix	`/sync` long-poll	✅ two-way	unencrypted rooms (v1)
DingTalk	Stream Mode	text only	reply window is per-message
WeCom	AI-Bot WS	✅ two-way	connects out; AES media
Signal	signal-cli (JSON-RPC)	✅ two-way	needs a local signal-cli daemon

Common setup

HARA_GATEWAY_ALLOWED — comma-separated user ids allowed to drive the bot. Empty = nobody (safe default; the gateway is never wide-open).
--cwd <dir> — point it at a real project. The default is a safe scratch workspace (~/.hara/workspace), so a full-auto bot never lands on a real repo by accident.
Each (chat × directory) is a stable, resumable session.
Session hygiene: a chat idle past 8 hours starts its next message on a fresh thread (one-time notice with /resume <id>; the old thread persists). Tune with HARA_GATEWAY_IDLE_HOURS; 0 disables.

Credentials per platform

Platform	Env
Telegram	`HARA_TELEGRAM_TOKEN`
WeChat	`hara gateway --platform weixin --login` (QR; the scanner is auto-allowed)
Discord	`HARA_DISCORD_TOKEN`
Feishu / Lark	`HARA_FEISHU_APP_ID` · `HARA_FEISHU_APP_SECRET` (+ `HARA_FEISHU_DOMAIN=lark`)
Slack	`HARA_SLACK_APP_TOKEN` (`xapp-…`) · `HARA_SLACK_BOT_TOKEN` (`xoxb-…`)
Mattermost	`HARA_MATTERMOST_URL` · `HARA_MATTERMOST_TOKEN`
Matrix	`HARA_MATRIX_HOMESERVER` · `HARA_MATRIX_TOKEN` · `HARA_MATRIX_USER_ID`
DingTalk	`HARA_DINGTALK_CLIENT_ID` · `HARA_DINGTALK_CLIENT_SECRET`
WeCom	`HARA_WECOM_BOT_ID` · `HARA_WECOM_SECRET` (create an AI Bot in the WeCom admin console)
Signal	`HARA_SIGNAL_RPC_URL` · `HARA_SIGNAL_NUMBER` (run a local signal-cli daemon)

# Telegram
HARA_TELEGRAM_TOKEN=… HARA_GATEWAY_ALLOWED=<your id> hara gateway --platform telegram
 
# Slack (Socket Mode — connects out, no Request URL)
HARA_SLACK_APP_TOKEN=xapp-… HARA_SLACK_BOT_TOKEN=xoxb-… HARA_GATEWAY_ALLOWED=U0123 \
  hara gateway --platform slack

In-chat commands

Command	Does
`/pwd` · `/cd <dir>`	show / switch the project the chat operates in
`/sessions` · `/new` · `/resume <id>`	list / fork / jump between threads in the current dir
`/voice` · `/say <text>`	toggle spoken replies · speak one message
`/send <path>`	send a file (image inline, anything else as an attachment)

Anything else is a task — hara runs it on that chat's session and replies.

Voice & media

Voice in — a voice note is transcribed and handed to hara as text, tagged so it knows it came from voice.
Voice out — replies can be spoken via pluggable TTS (below).
Images & files — send a photo and hara sees it (inline for a vision model, else described via your visionModel); ask it to send a file back and it uses the send_file tool. Both work in plain conversation on every platform marked two-way above.

Text-to-speech

Env var	For
`HARA_TTS_PROVIDER`	`say` (macOS, default) · `openai` · `cmd`
`HARA_TTS_VOICE`	voice name (e.g. `Tingting` for Mandarin on macOS)
`HARA_TTS_MODEL` · `HARA_TTS_BASE_URL` · `HARA_TTS_API_KEY`	an OpenAI-compatible `/audio/speech` endpoint
`HARA_TTS_CMD`	local command — text on stdin, `{out}` → output path

Keep it running

hara gateway runs in the foreground; supervise it with launchd / systemd / pm2, or:

nohup hara gateway --platform <name> > ~/.hara/<name>-gw.log 2>&1 &