// the take
OpenAgent is a self-hostable AI assistant platform (Go backend, React frontend) combining chat, RAG over your own documents, and agent loops that can browse the web, run shell commands, and call MCP tools. It's aimed at people who want a single-binary, self-hosted alternative to cloud AI assistants with enterprise auth and multi-tenancy baked in.
Ships as a genuinely self-contained single binary with native Windows support and no Docker/WSL requirement, which is unusual and low-friction for a Go+React stack. It has a real provider abstraction layer (30+ LLM and embedding providers) plus an MCP client wired directly into the tool-call loop, not bolted on as an afterthought. It also inherits mature enterprise plumbing from its lineage — OIDC/LDAP/SAML SSO, multi-tenancy, audit logging, a BPMN workflow engine — features most agent projects this size don't have yet.
This is a years-old general chatbot/knowledge-base platform (Casibase, same casbin org and Docker Hub namespace) repackaged under new 'agent' branding, and the codebase still carries that legacy surface: blockchain chain clients (ChainMaker, Ethereum), a Chinese TTS/STT vendor integration, and a dozen messaging-platform pipes (WeChat, Telegram, WhatsApp, X DM, Discord), all shipped in the same binary regardless of whether you use them. Giving the agent loop shell execution, browser-use, and MCP tool access inside one multi-tenant process is a large trust boundary, and there's no sandboxing or isolation story documented for the shell/browser tools — a prompt-injected or misbehaving agent can run commands directly on the host. The GitHub topic list mixes in unrelated hype terms (agi, openclaw, hermes-agent) alongside accurate ones, consistent with a recent repositioning rather than a ground-up agent design, so it's worth treating the 'agent' framing with some skepticism.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →