Every time I wanted an AI agent to communicate with humans through chat, I had to build separate integrations for each platform. Slack, Discord, Telegram: they all speak different protocols, handle auth differently, and require you to manage WebSocket sessions, reconnects, and rate limits yourself. It's a lot of boilerplate.
Pantalk is a lightweight daemon (pantalkd) that handles all of that upstream complexity. Your agent talks to it through a simple CLI or Unix domain socket with a JSON protocol, and the daemon takes care of the rest. Currently supports Slack, Discord, Mattermost, Telegram, WhatsApp, IRC, Matrix, Twilio, Zulip and iMessage. It's also local-first, sqlite persistence, no external dependencies.
I also added a notification system specifically designed for agents: it tracks when a bot is mentioned or DM'd, so an agent can poll for relevant events without having to process all message traffic.
It's early and I'm actively working on it. Would love feedback on the design, especially around the IPC protocol and whether the agent-first primitives make sense.
Hey HN, I'm the author of Pantalk.
Every time I wanted an AI agent to communicate with humans through chat, I had to build separate integrations for each platform. Slack, Discord, Telegram: they all speak different protocols, handle auth differently, and require you to manage WebSocket sessions, reconnects, and rate limits yourself. It's a lot of boilerplate.
Pantalk is a lightweight daemon (pantalkd) that handles all of that upstream complexity. Your agent talks to it through a simple CLI or Unix domain socket with a JSON protocol, and the daemon takes care of the rest. Currently supports Slack, Discord, Mattermost, Telegram, WhatsApp, IRC, Matrix, Twilio, Zulip and iMessage. It's also local-first, sqlite persistence, no external dependencies.
I also added a notification system specifically designed for agents: it tracks when a bot is mentioned or DM'd, so an agent can poll for relevant events without having to process all message traffic.
It's early and I'm actively working on it. Would love feedback on the design, especially around the IPC protocol and whether the agent-first primitives make sense.
Happy to answer questions.