Skip to content

Run your own models with a local AI harness

Locale brings chat, coding tools and projects into one workspace. Run supported models on your computer through the free Agent for Windows and Linux, use on-device browser AI, or choose a connected provider.

Request invite
Closed beta
This overview is public. An invitation is required for the product and accounts.

What is a local AI harness?

A model generates responses. A harness organizes the work around it: project context, available tools, approvals and the next step after a tool runs. Locale provides that workspace while keeping the selected model and where its data goes explicit.

Chats, browser files and previews live in the browser workspace. The Agent adds authorized device files, commands and supported local models.

Choose where the model runs

Local models on this computer
The running Agent executes supported native GGUF models on the connected computer. Compatibility, memory and disk space determine what can run. Device files and commands follow the Agent's permissions.
On-device browser AI
Supported desktop Chrome can run its foundation model in the browser. Availability depends on Chrome and the device. Locale asks for consent before setting it up.
Connected providers
A remote API or Locale account model processes requests outside the device. Browser-direct connections need an endpoint that permits browser access. Provider terms and charges apply.

Provider connections

Claude API models can use Locale's tools after you connect an API key in the browser and pair the Agent. Anthropic bills API use separately from Claude app subscriptions. The Agent checks permission before device file writes and commands.

Start with a supported local model

  1. Request beta access and open the browser workspace after admission.
  2. Install and pair the free Agent on the computer that will run the model, and keep it open while you use local models or device tools.
  3. Choose a supported model and review its compatibility, license, storage and memory needs before downloading it.
  4. Open a project, choose the local model and start with a small task.

A local model needs no paid remote request for its own generation. Downloads, hardware, electricity and any network services you choose have their own costs.

Choose a model that fits

The Agent checks the computer's CPU, GPU and memory before recommending a catalog model. The estimate includes context and runtime allocations, not only the download. Start smaller when the fit is uncertain.

For a supported Hugging Face GGUF link, the import dialog shows the files and download size before installing. Some packings need a particular runtime, so Locale checks the package and runtime pairing rather than treating every GGUF as interchangeable.

Keep data destinations clear

Local chats and browser files stay in browser storage unless a feature you choose transfers them. A local model runs on the connected computer; a remote model receives the request sent to its provider. Cloud storage and sync are separate choices.

Optional usage analytics excludes prompts and replies. Read the privacy notice and analytics choices.

Common questions

Can every model run locally?

No. The model must work with the installed runtime and fit the hardware. Its license and capabilities also matter.

Can a connected model use device tools?

Yes, when the provider supports tool calls and the Agent is paired. Device writes and commands require Locale's permission checks.

Does Locale make a remote model private to the device?

No. A remote model receives the request. Choose a local model when the computation needs to stay on the connected computer.

Is the whole product public?

No. Locale is in closed beta. Request an invitation or contact Locale.