What it does

Yao is a full-stack runtime shipped as a single binary. It runs three kinds of executors — LLM calls, sandboxed CLI agents (Claude Code, Codex, etc.), and plain TypeScript hooks — through the same request pipeline. You also get data models, REST APIs, server-rendered UI components, and a built-in chat interface, all without installing Node.js or Python on the host.

The interesting bit

The “cage, not an animal” framing is apt. Yao doesn’t try to be the AI; it wraps whatever model or CLI runner you bring, enforcing boundaries through Create and Next hooks that intercept every request. The built-in V8 engine runs TypeScript directly, and the CLI agent mode spins up isolated containers with VNC desktop access — useful for “computer use” tasks that need a real browser or IDE.

Key highlights

• Single binary for ARM64/x64; self-contained, no host dependencies
• Three executor modes (LLM, CLI agent, pure hook) mixable in one pipeline
• Native MCP support for tool connections via process, SSE, or STDIO
• Memory API with four scopes: request, session, user, team
• Built-in search stack: vector embeddings, knowledge graph, GraphRAG
• Skills ecosystem: drop SKILL.md packs into CLI agents for reusable capabilities

Caveats

• The README is heavy on feature lists and light on concrete usage examples or performance numbers
• “No containers needed on the host” is true for the binary itself, but the CLI agent mode explicitly uses containerized sandboxes — the distinction matters

Verdict

Worth a look if you’re building agentic workflows and tired of stitching together half a dozen services. Skip it if you need a mature plugin ecosystem or extensive third-party benchmarks; the project is ambitious but still proving itself.