# Agent Harness Installer
An installer that deploys an agent harness onto Intel hardware and wires it to a local, OpenAI-compatible inference endpoint. The installer deploys one of three open-source agent projects, each published independently: | Agent | Verified package | | --- | --- | | `openclaw` | npm `openclaw` — github.com/openclaw/openclaw | | `deepagents-code` | `dcode` (Deep Agents Code) — github.com/langchain-ai/deepagents | | `hermes` | Nous Research's official installer — github.com/NousResearch/hermes-agent (pinned installer script from the repo, not npm) | The inference backend is **configurable**, but every request goes through the **Inference Router** (IR): agents are wired to the Router's endpoint, never directly to a model server. By default the Router sits in front of a local **OpenVINO Model Server** (OVMS) serving an OpenAI-compatible `/v3/chat/completions` endpoint on Intel Core Ultra iGPU and Arc (CRI) GPUs; the Router itself is the published `intel/inference-router` image. Setting `HARNESS_LLM_ROUTER_ENDPOINT` instead points the Router at any existing OpenAI-compatible endpoint you already operate (e.g., Ollama, vLLM, an internal gateway), with `HARNESS_LLM_ROUTER_API_KEY` as an optional bearer token. OVMS is not started in that case, and this installer never starts, stops, or health-checks that endpoint (`status` reports it as "routed externally"). Host mode has no fresh-install default model: name it with `--model` or the model environment variables. The installer can reuse a previously exported model; otherwise it prompts on a terminal and fails non-interactively without a selection. ## Agentic AI Security Explore [Agentic AI Security](./agentic-ai-security.md) for security controls for agent deployments, starting with [Agent Sandboxing using OpenShell](./agentic-ai-security/agent-sandboxing.md). The `--runtime` option is **optional**. Interactive installation offers a choice of `host` or `openshell`; pressing Enter selects `host`. Non-interactive installation also defaults to host mode unless a runtime is explicitly selected with `--runtime` or `HARNESS_RUNTIME`. Select `--runtime openshell` to deploy OpenClaw, Hermes or Deep Agents Code in a sandbox. The installer supplies default sandbox policies and agent/provider configuration as a functional starting point. Fresh managed OpenShell deployments default to `OpenVINO/Qwen3.5-9B-int8-ov` when no explicit or reusable model selection exists, with model-tuning defaults for OpenClaw and Hermes. These sandbox defaults do not change host-mode model selection. See the [sandbox setup and policy boundaries](./agentic-ai-security/agent-sandboxing.md#getting-started) for prerequisites and limitations. ## Documentation - [Get Started](./get-started.md): install flow, usage and [model sources](./get-started.md#model-sources) (Hugging Face, Ollama, remote URL and the other Model Download hubs). - [Run Without a Checkout](./get-started/run-without-a-checkout.md): the standalone `curl | bash` bootstrap path. - [Uninstall](./get-started/uninstall.md). - [Known Limitations](./known-limitations.md). :::{toctree} :hidden: ./get-started.md ./how-it-works.md ./agentic-ai-security.md ./known-limitations.md :::