# safepilot

> Self-hosted AI assistant for Telegram that plans tasks as a DAG, classifies risk, and requires explicit approval before dangerous actions. Rust + SQLite.

- **Category:** AI Assistant
- **Licence:** Apache-2.0
- **Latest version:** v0.2.0
- **Open source:** oss
- **Languages:** Rust
- **Platforms:** Docker Compose, Cargo
- **Tags:** SQLite, Automation, Assistant, Ai Assistant, Secure, LLM, Personal, Open Source, AI, Rust, Telegram bot, Self Hosted, Approval workflow, Own your data, Safe, Security First

## Description

SafePilot is a self-hosted AI assistant that operates through Telegram, turning messages into durable, checkpointed automation runs. It uses a risk-tiered execution model where every task is classified by risk level, with explicit checkpoints before high-risk actions run. The system persists context in SQLite with optional compression, supports both direct LLM API calls and an agent mode with tool-calling loops, and integrates with many third-party services. It runs as a single Rust binary with no external database server, deployed via Docker Compose or cargo build. Security features include SSRF-protected fetch, allowlisted shell commands with bash/sh refusal, cleared subprocess environments, optional bubblewrap sandboxing, and at-rest encryption of sensitive database fields. The project is in early development (v0.2.0) with a security-first design philosophy.

## Links

- [Homepage](https://github.com/3DCF-Labs/safepilot)
- [Website](https://safepilot.dev/)
- [Docker](ghcr.io/3dcf-labs/safepilot)
- [GitHub Repository](https://github.com/3DCF-Labs/safepilot)
- [Docker Deployment Guide](https://github.com/3DCF-Labs/safepilot/blob/main/docs/docker.md)
- [Security Model](https://github.com/3DCF-Labs/safepilot/blob/main/docs/security-model.md)
- [Hardening Guide](https://github.com/3DCF-Labs/safepilot/blob/main/docs/hardening.md)
- [Architecture](https://github.com/3DCF-Labs/safepilot/blob/main/docs/architecture.md)

---
Source: https://selfhost.directory/project/safepilot