I mostly vibe-built MicroClaw — a Telegram bot that turns a chat into an agent-style AI assistant.
The idea came from nanoclaw, a TypeScript/WhatsApp agentic assistant. I liked the concept but wanted something small, hackable, and easy to self-host, so I rewrote it in Rust and used Telegram as the interface.
At its core, it runs a full agent loop: Claude can call tools (bash, file ops, web search, etc.), inspect results, and keep iterating until the task is done (up to 25 tool calls per message).
Some design choices I’m particularly happy with:
- Sessions are fully resumable, with complete tool history persisted in SQLite
- Long-running sessions auto-compact context by summarizing old messages
- An agent skills system (compatible with Anthropic’s Skills spec) for specialized tasks like PDF/DOCX/XLSX generation
- Explicit plan-and-execute flow with todo tracking for multi-step tasks
- Natural-language scheduled tasks
It’s built in Rust with Tokio, teloxide, and direct Anthropic API calls (no SDK wrapper). It’s a single binary, configured with three env vars, and that’s it.
Fair warning: this is still a toy / experimental project. There’s no sandboxing, permission model, or security hardening — the agent has full access to bash and the filesystem. Please don’t run it on anything you care about without understanding the risks.
Contributions, feedback, and criticism are very welcome.
This was vibe-built in about a day as a quick learning experiment, so the scope is intentionally small and rough edges are expected.
It’s not positioned as a secure or production-ready system — just a playground to see what breaks first.
Hi HN,
I mostly vibe-built MicroClaw — a Telegram bot that turns a chat into an agent-style AI assistant.
The idea came from nanoclaw, a TypeScript/WhatsApp agentic assistant. I liked the concept but wanted something small, hackable, and easy to self-host, so I rewrote it in Rust and used Telegram as the interface.
At its core, it runs a full agent loop: Claude can call tools (bash, file ops, web search, etc.), inspect results, and keep iterating until the task is done (up to 25 tool calls per message).
Some design choices I’m particularly happy with: - Sessions are fully resumable, with complete tool history persisted in SQLite - Long-running sessions auto-compact context by summarizing old messages - An agent skills system (compatible with Anthropic’s Skills spec) for specialized tasks like PDF/DOCX/XLSX generation - Explicit plan-and-execute flow with todo tracking for multi-step tasks - Natural-language scheduled tasks
It’s built in Rust with Tokio, teloxide, and direct Anthropic API calls (no SDK wrapper). It’s a single binary, configured with three env vars, and that’s it.
Fair warning: this is still a toy / experimental project. There’s no sandboxing, permission model, or security hardening — the agent has full access to bash and the filesystem. Please don’t run it on anything you care about without understanding the risks.
Contributions, feedback, and criticism are very welcome.
OP here.
This was vibe-built in about a day as a quick learning experiment, so the scope is intentionally small and rough edges are expected. It’s not positioned as a secure or production-ready system — just a playground to see what breaks first.
Happy to answer questions.