Channels

A channel connects an external platform to the same local Yachiyo instance. Messages arrive as threads with the same persona and the same memory — this is the same assistant, not a separate bot that happens to share a name.

Credentials live in ~/.yachiyo/channels.toml, deliberately separate from config.toml so bot tokens never mix with the rest of your settings.

Set everything up in Settings → Channels.

Supported platforms

Platform	Transport	Needs
Telegram	Bot API	Bot token from @BotFather
QQ	NapCatQQ over OneBot v11	WebSocket URL, optional token
QQ Official Bot	QQ Bot API (DM only)	App ID + client secret from the QQ Developer Portal
Discord	Bot gateway	Bot token from the Discord Developer Portal

Each platform can carry its own model override, so a chatty Telegram bot can run on something cheaper than your desktop default.

Telegram

Create a bot with @BotFather, paste the token into Settings → Channels → Telegram, and enable it. Restart the service from the same pane if you change the token.

QQ (NapCatQQ)

Yachiyo does not talk to QQ directly — it connects to a running NapCatQQ instance over OneBot v11. Point wsUrl at NapCat’s forward WebSocket (typically ws://localhost:3001) and set the token if you configured one.

QQ Official Bot

The official Bot API path, using an App ID and client secret from the QQ Developer Portal. Direct messages only — no group support on this transport. Use the NapCatQQ path if you need groups.

Discord

A standard bot token from the Developer Portal. Discord supports both DMs and server text channels; server channels go through group discussion mode.

Access control

Nobody gets to use your assistant by messaging your bot. Users and groups are registered on first contact and start in a pending state.

Users — pending → allowed or blocked, with a role of owner or guest. Each carries a token quota (usageLimitKTokens, null for unlimited) and accumulated usage.

Groups — pending → approved or blocked. Approving a group starts its monitor immediately if the app is running.

Note the two vocabularies differ: users are allowed, groups are approved.

Approve from Settings → Channels, or just ask — it can list who is waiting and act on your answer:

Who’s pending on my channels right now?

For scripts and the exact commands, the CLI:

yachiyo channel users --json
yachiyo channel groups --json
yachiyo channel groups set-status <group-id> approval
yachiyo channel groups set-status <group-id> block

See the channel CLI reference.

Approval is a real decision

Everything in the guest model below is what stands between a stranger on Telegram and your assistant. Approve people, not requests — and be aware that asking Yachiyo to “approve everyone pending” hands it a decision you should be making yourself.

The guest model

A guest is not you, and Yachiyo treats them accordingly:

• Sandboxed workspace. Each guest and group gets its own workspace directory.
• Filtered memory. Memory results containing any of your memoryFilterKeywords are hidden from guests.
• Separate profile. Guests get their own USER.md mode, so what the assistant learns about a visitor never lands in your profile.
• Token limits. Per-user quotas, enforced.
• Guest instruction. A custom block injected into the system prompt for guest conversations — what guests should know about you, and any rules for how the assistant should handle them.

Group discussion mode

Having a bot reply to every message in a group chat is intolerable. So group mode runs a probe: a cheap model watches the conversation and decides whether the bot has anything worth saying. Only then does the full model get involved.

The monitor is a three-phase state machine:

Phase	Behavior
Dormant	Buffering messages, not checking. Wakes on activity.
Active	Checking periodically (default every 60s) to see if it should speak.
Engaged	Recently spoke; checks more often (default every 30s) to hold a conversation.

It falls back down the ladder after a few checks with nothing worth saying — three misses by default at each step. Messages that arrive while dormant sit in a buffer with a wake delay (default 60s) so a burst of chatter produces one considered response rather than three reflexive ones.

Group tuning

Setting	Default	What it does
Group model	tool model	Model that runs the probe
Check interval	60s active / 30s engaged	How often the probe evaluates
Verbosity	0	0 = the normal throttle curve, 1 = never throttled
Group context window	64k tokens	Sliding window budget for probe messages
Rewrite model	unset	Rewrites outgoing replies into the persona’s voice; unset sends them as generated
Vision	off	Pass images from group messages to the probe
Image to text	off	Pre-describe images as alt text before the probe sees them

The group probe thread also does a rolling handoff: once the raw transcript outgrows roughly twice the context window, older material is summarized rather than dropped, with hysteresis so it does not re-summarize every turn.

DM threads have their own compaction threshold (default 64k tokens), set separately in the same pane.

Sending messages yourself

yachiyo send channel pushes a message straight to a platform as the bot — no thread, no inference:

yachiyo channel users --json                       # find the target id
yachiyo send channel <id> "The build finished."

Useful in scripts and CI. The app must be running and the relevant channel service must be active. See the send CLI reference.

channels.toml holds secrets

Bot tokens and client secrets are stored in plain text in ~/.yachiyo/channels.toml. Unlike provider API keys, these are not redacted in output — treat the file as a credential store, and keep it out of any synced or shared folder.