RIP OpenClaw… this 100% private AI Agent is insane

Agent Zero can run as a fully local, privacy-first AI agent by combining a Docker-isolated agent environment with locally hosted language and utility models via Ollama—so sensitive prompts, files, and analysis stay on the user’s machine. The setup matters because many popular “autonomous” coding agents can be risky on a workstation: they may delete files, leak data, or require trusting third-party services. Agent Zero’s approach—running inside a Docker container—aims to keep the agent’s actions contained while still giving it practical autonomy.

The walkthrough starts with installing Agent Zero from the official site (agent-zero.ai) using a one-line install script. During setup, the user creates a new instance (named “YT”) and keeps the default port (5080). Agent Zero then pulls a Docker image that includes a full Linux environment and tools, which the creator frames as the reason it’s safer than running multiple other agents directly on the host system. After installation, the browser interface shows a warning until an LLM is connected.

Next comes the local model layer. The guide recommends Ollama over LM Studio for Agent Zero compatibility. Ollama is installed via its own one-line script, then models are managed through terminal commands like “ollama list” and “ollama run <model>.” Model choice is tied to hardware: model size ranges from smaller ~1.2B options up to very large 122B-class models, but the practical constraint is GPU VRAM (or shared GPU/CPU memory on Apple silicon). The walkthrough demonstrates running a 122B model locally and notes the speed trade-off (around tens of tokens per second) while emphasizing that the model runs without sending data off-device.

Agent Zero then needs three local components configured in its settings: a chat model, a utility model, and (optionally but importantly) an embedding model. For the chat model, the provider is set to Ollama, the exact Ollama model name is entered, and the context length is matched to the model’s configuration (the guide stresses it must not be smaller in Agent Zero than in Ollama). The “Chat model API base URL” is pointed to a local Docker-accessible host string (HTTP host docker internal 111434). Once connected, Agent Zero can generate responses with “reasoning” while remaining fully local.

The utility model is used for faster background tasks and long-term memory operations (including vector embeddings and markdown-based storage). The walkthrough shows swapping from a heavy 122B chat model to a smaller, faster local option—using GLM 4.7 flash (30B)—to reduce latency for memory-related work. Finally, embedding model issues are treated as a common “gotcha”: the default Hugging Face embedding setup may fail with Ollama, so the guide switches to “nomic-embed-text” by pulling it in Ollama and updating the embedding provider and base URL.

With the system running, the guide demonstrates a high-stakes use case: analyzing private photos locally. Users drag images into Agent Zero, then prompt it to read image metadata (GPS coordinates, dates, camera models), use vision to describe contents, sort images into categories, and generate a markdown travel report with timelines and patterns. The workflow relies on multiple tool calls and terminal execution inside the Agent Zero environment, and the guide highlights that the resulting report is produced without uploading photo data to external AI services.

The guide closes by recommending local agents for sensitive domains—medical records, financial documents, credentials, legal contracts, journaling/therapy notes, business secrets, and even offline survival planning—arguing that the privacy and control trade off against slower performance. It ends with a pitch to move beyond tinkering and build an AI product, but the core takeaway remains: Agent Zero plus Ollama can deliver autonomous, multi-step assistance while keeping data on the machine.