Set up AI and extensions

This guide covers Grid 0.66.8. Check that your app and runtime are updated; model availability still depends on published packages in your chosen registry.

Open File → AI & extensions, or choose AI setup in Grid Agent. Your saved AI choice is shared by Chat, new Agent tasks, and new workbook AI_PROMPT and AI_ASK requests. Formulas that name an exact model with PREDICT keep that choice.

Choose how to run AI

Connect an AI account

1. Choose Add API key for your provider, paste the key, and choose Save and use.
2. Choose Test connection & find models to check access without generating an answer.
3. Select a model under Model preferences, then save the preference.

The key is encrypted on the Grid instance. Saving a key alone does not check its validity, model access, or account balance. Use provider switches back to a saved account. Your provider and model choice survive a restart.

Connect your model server

Choose Connect server, enter its OpenAI-compatible API address, and supply a key only if the server requires one. Choose Connect & find models, select a model, then Save and use server. If discovery is unavailable, enter the model name your server exposes.

A local address refers to the computer running Grid. Use the server's HTTPS address when it runs elsewhere. Grid connects to this server; it does not install or start it. Edit connection lets you change the server later.

Install a local model

1. Open Local models. Under Help me choose for, choose Chat and writing or Images and documents, then search for a model or choose a family shortcut.
2. Choose Find a good fit for this computer for a recommendation, or View beside a result. Review its download size, memory needs, selected CPU or GPU, accepted inputs, and publisher terms.
3. Choose Install. Progress shows downloaded bytes; Pause download and Resume download retain completed transfers where the publisher supports it.
4. In Installed, choose Try it, then Use this model to make it the shared AI choice. Open Grid Agent continues into chat.

For pictures, the image option recommends a compatible model whose package lists image support. Look for Text and images in its details. Image support not listed means the available metadata does not declare it. If no image model fits, refine the search or connect an image-capable provider or model server.

Already writing an image question in Chat? Choose Choose an image model beside your attachments to open this setup. Your question and images remain in the draft when you return.

Only published, compatible releases can be installed. Families come from your selected registry, so catalog additions do not normally require another app update. New model architectures may require an engine update. Gated downloads still require the publisher's access process; Grid does not accept terms for you.

Apple GPU execution uses Metal, and compatible NVIDIA releases use CUDA. Apple Neural Engine/NPU execution is not available. A CPU-only release remains CPU-only. Grid does not silently send a local-model request to a cloud provider.

Manage model offers removal. For an update, install a newer exact release, try it, and select it; older versions remain until you remove them.

Find and activate an extension

Choose Extensions, search, open a release, choose a workspace folder, and select Install. Review permissions & activate then opens Package Center for that folder. Review the permissions before activating the extension. Package Center also retains its npm, Python, and other upstream-package tools.

Choose your registry

Registries starts with Grid Registry. To add another, enter its name, HTTPS address, and the verification code supplied independently by its administrator. Have a registry setup link? can fill those fields from a setup link or JSON. Review them, then choose Verify and use registry. Saved registries can be selected again; a failed check leaves your current selection intact.

The selected registry supplies both signed extensions and local models. Upstream npm/Python registries keep their separate Package Center settings.

Continue with images or live input

Choose an image-capable model, then follow Ask about images, recordings, and live sources. The saved connection is the same for these requests.

Use AI in a workbook

The 0.70.0 development workbench adds Dataset, Predict, Train, Evaluate and Process surfaces. Check the opened workbook's capabilities before starting; the saved AI connection supplies a text model, while trained predictions select an exact model version and its feature mapping.

Capture the worksheet range, Table, Frame or CSV you want to use into a retained Dataset. Train produces a candidate version; selecting that version is a separate action. Predict starts with one example and lets you select a Dataset for a batch. Evaluate compares subjects on the same retained cases; scoring saved results makes no new subject calls.

An opened desktop document keeps these datasets, jobs and model versions with that physical document. A workbook with the same display name has separate data. Native Process operations require the host's explicit installation for the document and participant; opening the workbook does not start a run. Review the selected setup, then choose Start. Publishing a result to a declared input and preparing a trained formula are separate choices.

A participant can also select a named saved author, explain or research profile. Explicit Start uses that profile's saved provider, instructions and limits; a conflicting connection refuses. Use author for a Grid-source proposal: native checks validate its exact source and every model/tool call charges the Process budget. Author/research tools inspect supplied source only; explain has no tools. The language's default stays the AI-setup model binding. Compiled Process Evaluate cases use the same selected profile.

In Process Start setup, choose Installed methods to read the host's catalogue. Select a matching method and review its limits before pressing Start; selecting alone starts no work. Use saved plan without this method returns to the ordinary plan. In a saved run's definition, expand Retained turns beside an ASK step to inspect its original replies. Sequence and Diagram use the same frozen step identities, and the full Conversation remains available below. Choose Inspect saved values and receipts for a settled run's selected frozen page. Review the original participant settings if requested, then expand a step's Saved values · receipts disclosure. Values and completed receipts keep their own frame identities; repeated values also show their occurrence. This check and local entry paging make no model or operation call.

See Workbook surfaces for the configuration and capability contracts.