AstrBot

AstrBot's pitch is narrower and more credible than most "AI assistant" projects: it is an agent runtime that happens to live in chat apps, not a chat app that happens to call a model. The pieces that back that claim are named and concrete — Sub-Agent orchestration, multi-step workflows, tool calls, context management, native MCP and Skills as the extension standard, and an Agent Sandbox that isolates execution so a misbehaving tool call doesn't take the host down. Cron Jobs let the agent act on a schedule rather than only in response to a message, which is the difference between a bot and a worker. The knowledge base is unusually specific for an open-source project. Parsing goes through Microsoft Markitdown for PDF, DOCX, and Markdown, and retrieval combines dense search with sparse BM25 — a hybrid setup that matters most for Chinese and multilingual corpora, where pure embedding search tends to underperform. A single conversation can reference multiple knowledge bases, which is the practical pattern for teams that keep product docs, internal wikis, and support macros in separate stores. Model support is provider-agnostic: OpenAI, Google, Anthropic, and any OpenAI-compatible endpoint, with model switching and multimodal coverage across text generation, vision understanding, STT, TTS, and embeddings. That breadth matters for two reasons. It lets you route cheap models to cheap tasks and strong models to hard ones, and it lets you run against a local or self-hosted endpoint when data cannot leave your network. What you are actually buying is control. The project reports 41.9k GitHub stars, 200k+ weekly active users, 1000+ community plugins, and 100+ contributors under AGPL-v3 — enough of a community that the plugin ecosystem is a real asset rather than a bullet point. Deployment paths are varied (one-click launcher, desktop app, Docker, uv, pip) and the WebUI centralizes configuration, debugging, logs, and plugin management, which removes some but not all of the operational burden. The honest trade: there is no managed tier in the sense of someone else running it for you, and the deployment story assumes Docker, uv, or pip plus a Python environment and real command-line comfort. Throughput, latency, and uptime are yours to manage, and the project makes no official SLA or dedicated support commitment in the material available. AGPL-v3 is a genuine constraint if you intend to modify and redistribute inside a closed-source product. Roadmap direction is community-driven across 100+ contributors rather than set by a paying customer base, and there is no native mobile app — you reach it through the web interface. If any of those are dealbreakers, a hosted assistant is the better fit.