Four doors into the same agent: two chat platforms, one personal account, one microphone. The answer arrives where the question already was.
Four connectorsStarted by the bridgePython · NodeOne IM helper setv1.0.0
What a connector is
The answer arrives where the question already was.
A connector is a long-lived client that Quanta's bridge starts for you. It receives a message, turns it into an agent turn, and sends the answer back into the same thread — with de-duplication, chunking and attachment download shared across all of them.
None of them is a daemon you launch by hand. Quanta holds the credentials, starts the client, and routes each message as if it had been typed into the window — same memory, same pages, same tools.
Long connection or gateway
Feishu and QQ hold a socket open, so nothing needs a public callback URL, a tunnel or a server of your own.
Reply inside the window
QQ only accepts a passive reply for a while after the message. The connector tracks that clock; the agent does not get to be late.
Local stays local
WeChat and voice run on your machine against your own account or microphone. Nothing is relayed through a Quanta server.
SkillA written procedure plus a script. Loaded when its description matches the request.SKILL.md
PluginCode inside the agent runtime. Changes how a run behaves rather than what it can call.dsh ACP
ConnectorA door into the agent from somewhere else — a chat app, a microphone.bridge
MCP ServerA separate process exposing tools over stdio JSON-RPC.stdio
The four
Three inboxes and a microphone.
Each demo below is the same shape: a message comes in, a turn runs, an answer goes back. What changes is the chrome and the rules.
CONNECTOR-01feishu
A long connection, no callback URL.
A Feishu self-built app opens a WebSocket long connection and receives every message on it; replies go back through the OpenAPI. Quanta supplies the app_id and app_secret, and each message becomes an agent turn.
Self-built app.Not a marketplace app — your tenant, your bot, your permissions.
No public address.The connection runs outbound, so a laptop behind a router works.
Needs lark-oapi.One Python dependency. Attachments are downloaded and filed like any other input.
Press send: the message arrives on the long connection, runs as a turn, and comes back through the OpenAPI. Long replies are chunked; a file would be downloaded first.
CONNECTOR-02qq
Official bot, passive replies.
QQ's official bot gateway is a WebSocket too, but it answers with a rule: a reply is only valid for a short window after the message, and it must carry that message's id. The connector keeps the clock and the id, so a slow turn still lands.
Passive, not push.The bot may not message first. A reply is bound to one incoming message.
msg_id bound.The id travels with the turn so the answer has somewhere to go.
Needs websocket-client.No framework, no SDK — one small dependency and a socket.
The reply is sent as a passive answer inside the window, with the incoming message's id — the reason the footer shows a clock instead of a chat box.
CONNECTOR-03wechat
Your own account, your own machine.
The WeChat connector bridges a personal account to the agent on the same computer. It is not a platform API, and it is not pretending to be one: the client runs locally, keeps the session on your disk, and needs an npm install before it starts.
Local session.The login lives on your machine, not in a service. Sign out and it is gone.
npm install first.It is the only connector here with a Node dependency tree.
Use it on your own account.A personal account is not an enterprise tenant. Keep it to yourself, on your own machine.
A personal chat, answered by the same agent. Files sent to it are downloaded, filed, and readable in the next turn.
CONNECTOR-04voice · rhema
Voice in, voice out, one sidecar.
Rhema is a streaming VAD / ASR / TTS sidecar. It decides when you have started and stopped speaking, transcribes while you talk, and speaks the answer back — the device is not asked to type.
Streaming, not a file.VAD cuts the utterance; the transcript grows while you speak.
MLX on Apple silicon.The ASR path wants mlx-audio and huggingface_hub; on other machines it falls back.
Interruptible.Speaking over the answer stops the voice and keeps the transcript — the text is the record.