CLIver is a Personal AI Lab — a two-package project for AI agent research and experimentation.

• cliver-core — the embeddable engine: a Re-Act loop with tool calling, OpenAI and Anthropic protocol providers, per-purpose model bindings, nine builtin tools, Agent Skills, MCP, and opt-in permission and execution-policy axes. No CLI, database, or config files — Python objects in, messages out.
• cliver — the CLI application built on cliver-core: config-directory and skill-discovery conventions plus a one-shot cliver command with interactive tool approval. An interactive TUI is not built yet.

Highlights

Model freedom. You choose the model. cliver-core speaks both OpenAI and Anthropic protocols and connects to any compatible endpoint — cloud providers like DeepSeek, GLM, and MiniMax, or local ones like Ollama and vLLM. Live-verified presets for several brands ship in cliver_core.presets.

Open architecture. Add custom tools with the @tool decorator, load Agent Skills from explicit source directories, and connect any MCP server over stdio or streamable HTTP. The core stays embeddable — app-specific behavior lives outside it.

Explicit policy. Tool permissions are opt-in: a SAFE/ASK gate with an application-supplied callback, plus an execution policy that controls subprocess environment, working directory, and read/write path scopes. With no policy configured, nothing changes.

Embeddable by design. AgentCore has zero CLI dependencies. Construct it with ModelBindings and use chat(), stream(), or run_skill() from your own scripts and applications.