Changelog

2026-10-18 — Talk to an agent while it works

You no longer have to wait for an agent to finish before saying more. While the assistant or any agent is running, the composer stays open: what you type is handed to the run as it goes (queued for its next step, or, if you choose Interrupt, delivered at once, cutting short whatever it was writing or doing except a change it has already started making). The same input appears next to each run in the Agents app, in Pulse and in Agent Center, so you can steer a sub-agent directly while its parent keeps going. Handing work to a sub-agent no longer freezes the agent that asked: it carries on, or sits idle at no cost until the result comes back, and you can keep talking to it in the meantime. Agents also now say when they are done instead of trailing off, and an agent that still has helpers working cannot declare itself finished until they report. Runs on the server behave the same way through their own message channel. All of this is on under the beta flag "steerable-runs" (Settings › General › Early features); with it off, nothing changes.

2026-10-18 — A trained model behind predicted tools

Predicted tools (below) now have a second, trained guesser alongside the keyword matcher: a small retrieval model (about 50 MB) that was distilled from a much larger tool router and learned what requests sound like for every app tool - including apps it never saw in training. When predictions are on for you, the browser downloads it once in the background and keeps it cached; until then the keywords guess alone, and the two are combined once it is ready. It runs on your graphics card when the browser allows it and on the CPU otherwise. On requests for apps it was never trained on, the combined guess puts every needed tool in its top 20 for 92% of requests, against 73% for keywords alone (and in the eight tools a run actually starts with for 79%, against 59%). Nothing about your request leaves the browser; the model is served from the same place as other site assets.

2026-10-17 — Predicted tools for the assistant (behind flags)

The assistant can now start a request with the tools it is likely to need, instead of loading each one as it goes. A small router runs in your browser, matches what you asked against keywords for every app tool, and adds its best few guesses to the assistant's starting tools; a wrong guess just means the assistant loads the missing tool itself, exactly as before. It can also work the guess out while you are still typing, so the assistant starts sooner - nothing you type leaves the browser before you send it.

All of this is off by default and rolled out with Vercel Flags: a measurement mode that predicts and records hits and misses without changing anything, predicted tools, routing while typing, and letting speculative turns run alongside on-demand tool loading. Once a flag is on for you, Settings › General › Early features lets you switch it back off. Speculative turns now also discard a guessed reply if the tool list changed before it resolved.

2026-09-17 — Connect Claude Code or Codex with one line

Community tools: a public repository (kumin-consulting/nextos-community-tools) holds one folder per community-built tool with the fields a public page needs beyond an app manifest - summary, search title, keywords, share image, screenshots with descriptions, questions, install line, publish and update dates - and the site renders each at /community/ with structured data, plus a listing at /community. The sitemap gained a community sitemap with real per-page dates, and every canonical, share URL and sitemap location now names the www host the site actually serves from.

The gateway now serves the Anthropic Messages API (what Claude Code speaks) and the OpenAI Responses API (what Codex speaks) as metered passthroughs: your plan's included model on the deployment's AI Gateway credential, or your own Anthropic or OpenAI key from the vault for that provider's models. Settings > Developer > "Connect Claude Code or Codex" mints a key for the tool and shows the one line to paste; the hosted connect script writes the tool's own config and keeps a backup. Background jobs on the included model now run on the deployment credential too (they used to fail for lack of an account key), and the field bar's @-mention of a department says which gate stopped a run and opens the fix.

The same Connect card now has a third option, "Your machine": a one-line bridge (connect.sh ... bridge ...) that registers a computer under Settings > Agents > Your machines and runs your NextOS agents through whichever of Claude Code or Codex you're already signed into there, on that subscription rather than any API key. Settings > Agents shows a badge per machine for each CLI it detected, and Settings > Privacy > Background runs lists "Claude Code (your subscription, on your machine)" and "Codex (your subscription, on your machine)" as places background work can run, with a note when no machine is running the bridge yet. See Bring your own machine for what the agent can do there and what we never touch.

2026-09-13 — Teams and governance, the gateway, cloud storage, password sign-in, revenue features and agent speculation

Four rounds of parallel agents (PRs #70–#99). A governed team layer: orgs with owners, admins, members and custom roles; one policy object covering models, MCP, skills, tools and tool arguments, budgets, rate limits, guardrails, routing, residency, caching, approvals, a two-person rule, break-glass elevation and key rotation; presets, history, a diff preview and a 30-day simulation; team-shared runs with review, shared agents with versions and canaries, schedules, evals, a team folder and team memory, an inbox for paused runs, notifications, spend forecasts, an overview dashboard, audit streaming to a SIEM, retention and export, SSO enforcement, SCIM, guest memberships, a public gallery and status page, templates and onboarding. The gateway gained per-key limits, exact and semantic caches, routing with streaming-safe fallback, structured output, a per-request cost cap, a prompt registry and a run feed; runs stream over a signed transport with a public viewer, replay, diff, forks and embeds. Sign-in with a username and password. Cloud storage beyond Google Drive: OneDrive, SharePoint, Dropbox, Box, S3-compatible and WebDAV, with mirroring and push webhooks. Revenue: annual plans and an Enterprise tier, metered gateway pricing, a 14-day team trial, upgrade nudges, seat and budget upsells, managed MCP, a funnel, dunning and a cancellation flow, buyer pages, integration pages and a templates marketplace. Agent speculation (branch prediction): read prefetch and speculative turns with commit/squash, controlled per org, account and agent with token guardrails. docs/loose-ends.md lists the seams.

2026-09-10 — Square billing, product pages, Windows apps and the sandbox executor

Plans and entitlements keyed by the account's email, with Square subscriptions, payment-link checkout, a signed webhook and server-side gates on everything that costs money. The product landing page moved to /nextos so the root of the site is the portfolio again, and every product page took on the portfolio's design language. 32-bit Windows programs run in the browser through Boxedwine with an install flow, and background runs gained a Vercel Sandbox executor with real shell tools and mid-run approvals.

2026-09-09 — Marketing pages, app product pages, and the checkout groundwork

A product landing page, a full app catalog with a product page per app, a pricing page, and this changelog. The app READMEs, package manager support and the Square SDK also landed this week.

2026-09-08 — The built-in app library grows to 43 apps

Bundled, hosted and native third-party app support, plus the App Store's package management UI, background execution, an OPFS filesystem backend, a policy-aware network proxy, WebAssembly language runtimes, and multi-window, multi-device support.

2026-09-07 — Every AI provider, wired end to end

Anthropic, OpenAI, Ollama, Mistral and Perplexity, each integrated with both text and media support, behind one agent kernel and one provider plugin system. The Claude Code extension shipped inside VS Code, running its own Node substitution since a native CLI cannot run in a WebContainer.

2026-09-06 — The Field shell and the real workstation

The Field shell (the strip, the pulse rail, snapping cells) replaced the classic desktop as the default. Real CPython (via Pyodide) and Node.js (via WebContainers) started running behind a pty in the Terminal. A Unix-style virtual filesystem in IndexedDB backs every app. Google Drive sign-in, sync, and live push (via a changes.watch webhook, Redis and Server-Sent Events) landed the same week.

2026-09-02 — The agentic OS

Agents that can operate every app under your approval, the first version of the OS shell, and the initial set of built-in apps.

2025-06-06 — The first commit

The portfolio and the first per-app READMEs.