Build Your Own Open-Source Dots with Pi and Telegram
This tutorial demonstrates how to build open-source, always-on AI agents, inspired by OpenAI's 'dots' concept. Using a VPS or local machine (like a DGX Spark), the system utilizes a message gateway architecture (inspired by Hermes and OpenClaw) to allow specialized agents (e.g., personal assistant, research agent) to be accessible from multiple platforms, specifically Telegram. The core components include agent workspaces, custom skills (like Google Workspace CLI access), and persistent session routing managed by a local SQLite database.
Key takeaways
-
Agent Architecture
Each agent is contained within its own directory, defining its specific tools, skills, and context via an `agents.mmd` file. This allows for specialized, isolated workspaces (e.g., a research agent vs. a personal assistant).
-
Skills and Tooling
5:49
Agents can be equipped with skills, such as `firecroll pars`, `firecroll scrape`, and access to the Google Workspace CLI, enabling tasks like reading emails, summarizing messages, and creating calendar events.
-
Message Gateway Implementation
11:52
The `pi-gateway` acts as a background process, constantly listening to multiple platforms (demonstrated with Telegram). It uses a local SQLite database to manage sessions, ensuring incoming messages are correctly routed to the specific agent's workspace and session.
-
Setup Process
19:02
The gateway is installed using `uv tool install pi-gateway`. Agents are initialized using `pi gateway` commands, requiring the Telegram BotFather token and the user's unique ID for secure operation.