OpenClaw released version 2026.9.1 on September 3, 2026, with a faster path for new installations. The system detects existing Claude Code or Codex access and API keys, verifies them, and opens the web dashboard from the Gateway.
From installation to the first chat
The new path reduces decisions during setup, but it does not remove configuration. People who need to select providers, credentials, or advanced options can still use the project’s custom setup process.
OpenClaw also identifies how a user is authenticated and retains configured providers. This prevents repeated work when a compatible account or valid key is already available on the computer.
Hosted services or local models
The agent can use hosted services or connect to local models through Ollama and llama.cpp. The update improves discovery of those servers, but it does not automatically include a model suitable for every computer.
With an API, the main inference workload runs outside the PC. This approach works on conventional hardware, but it may create usage costs and send information to the provider, depending on the service, integrations, and instructions.
Hardware depends on the model
There is no universal specification for local inference. Memory needs vary with model size, quantization, context length, and the number of active tools or sessions.
A computer with 16 GB of RAM can experiment with smaller models, while 32 GB provides more room. A GPU with 24 GB of VRAM supports heavier workloads, but demanding local agent loops may still face latency and capacity limits.
Improvements that keep the system running
The release can roll back an update when its follow-up check fails, preserve configuration references, and wait until plugins are ready. These measures reduce interruptions, but they do not guarantee trouble-free migrations.
Personal skill libraries are also available on shared Gateways. Each person can keep individual automations separate from the common collection and import packages, although external skills should be reviewed before receiving access to data or tools.
Permissions need careful attention
OpenClaw can read files, start processes, and control a browser when those functions are enabled. The project offers isolation to restrict tool access, but sandboxing is disabled by default.
The web dashboard controls conversations, settings, and execution approvals. Official guidance recommends keeping it on the local machine or using a protected connection instead of exposing it directly to the Internet.
Easier setup without unrestricted autonomy
Persistent approvals reduce repeated prompts for selected tools, but they also increase the effect of an incorrect instruction. New users should prefer temporary permissions, restrict accessible folders, and review every sensitive action.
The faster setup makes OpenClaw more approachable, not infallible. A person can begin with an API on ordinary hardware or test local models that suit available resources, while retaining supervision over credentials, tools, data, and results.
