What happened
webAI’s changelog dates version 3.1.0 to November 9, 2026 and labels the release “Persistent canvas, Array, and Meridian Air.” The Home view now keeps a canvas in the center with Chat beside it. While a reply is being written, webAI says the canvas displays the searches it ran and the pages it read. Either column can expand, and Command-backslash collapses Chat or restores it. In a smaller window, the canvas yields space to the conversation.
The release changes Array, webAI’s assistant for web work. webAI says Array checks each web search on the device and can leave out personal details, including booking codes and passport numbers. The canvas displays the information it held back. Array can also turn an answer into a document. When a user asks for revisions, it edits only the sections the user names rather than broadly rewriting the document.
Model handling changes as well. TwIL 8B is now available. Meridian 1 and Air appear together in the model catalog as Meridian Air, while Meridian can inspect images attached by a user. webAI says Air is the default on Macs with 64 GB of memory or more. The model list now shows exact download sizes and blocks a download when the file will not fit in available disk space. The release notes also say the application reinstalls its local model engine if that engine goes missing, rather than failing to load a selected model.
Other parts of the product receive operational changes. In Groups, a user can stop a Persona while it is writing and retain the question for another attempt; administrators can add members through Members. Knowledge bases open from Library, where users can add documents, search indexed passages, manage suggested prompts, rebuild a knowledge graph, and change settings. Users can create a knowledge base from a Persona and subsequently manage it in Library. Connector controls are now separate for allowing an assistant to send or change something and for requiring it to ask first. A Persona can be granted access to Apple apps one at a time. webAI also lists interface fixes, faster handling around completed tool calls, and general stability improvements.
What the archive adds
New Runtime’s earlier coverage includes an OpenCode release with session-persistent terminals and agent-published artifacts from October 31, 2026 and a SuperARIS preview of a persistent multi-agent research workbench from November 8, 2026. Those notes concern adjacent products that preserve work across interactions rather than webAI itself. This is a new event.
Trend signal
This weakly supports agent-ready software exposes capabilities. webAI’s connector settings separate permission to send or change data from the requirement to request confirmation, while Array exposes a visible account of information held back from a search. Both move from an undifferentiated assistant permission toward individual, inspectable controls. The release does not establish how those controls behave under every connector or failure condition.
Why it matters
Local AI products are useful only when people can understand the boundaries around them: what information goes into a search, what an agent can alter, and whether a chosen model will fit the machine that runs it. v3.1.0 puts several such boundaries inside the ordinary interaction flow rather than treating them as background configuration.
Array’s visible record of withheld details gives a user something to inspect before relying on web search with sensitive context. Section-limited editing offers a similarly bounded contract for document work: a user can request a change to named sections rather than invite an unrestricted rewrite. Separate connector choices matter for teams because “can act” and “can act without asking” are materially different operational permissions.
These controls are not a substitute for a team’s access design, retention policy, or review process. Their value is that they create explicit points at which those policies can be tested against the application’s observed behavior.
Limits and uncertainty
The published release notes are product claims from webAI and are not independent measurements. Detection accuracy for Array’s treatment of personal details, the range of information it recognizes, and behavior on unusual or ambiguous queries remain unverified here. The practical security properties of connector permissions, including logging, enforcement across integrations, and behavior after an interrupted request, also need testing in a real deployment.
Performance for Meridian Air, including image handling and the implications of its default selection on 64 GB Macs, is not independently benchmarked. Teams with regulated, confidential, or high-consequence workloads should not infer compliance or complete data protection from a visible redaction step or a confirmation control alone.
Skill to build now
Create a small acceptance test for every agent workflow that can search, edit, or send data. Use a harmless synthetic booking code and a unique marker in a test document. Confirm that Array displays the withheld value before search, that a request to edit one named section leaves the other sections unchanged, and that a connector configured to ask first actually pauses before an external action. Record the result with the application version and machine configuration.
That test turns feature descriptions into an operational check. It also gives a team a repeatable regression test when a model, connector, or desktop application version changes.
Reader takeaway
webAI 3.1.0 is most notable for making local-AI boundaries more inspectable: it can show what Array withholds from a search, constrain edits to named document sections, expose model download sizes, and separate connector action permissions from confirmation requirements. Treat those as controls to verify in your own workflow, not as guarantees to assume.