Chat-first Aria with agents, teams and projects — plus flow control that survives a closed tab

What changed

v0.8.0 5 October 2026 Chrome, Web app Breaking changes & migration

New 20

Aria & agents Chat with Aria from the right-click menu, the address bar, a shortcut or the notch. Ways to open Aria
Aria & agents Agents with their own skills, rules, memory and activity log. Agents
Aria & agents Teams of agents that hand work to each other in a shared thread. Teams
Aria & agents Projects group chats, files and agents. Projects
Aria & agents Live browser control and routines, so an agent can act on the page you are on, with safety checks. Browser control & safety
Marketplace Agent and team templates to start from a ready-made agent. Agent & team templates
Aria & agents A change plan before anything is applied, and highlighted nodes after a build. Building workflows in chat
Local AI Local AI task nodes: generate, summarise, translate, classify, caption, detect objects and transcribe on your device. Local AI
Local AI A Local Models page to install, browse and delete in-browser models. Local models
Nodes Wait of a minute or more survives closing the tab, for up to 30 days. Wait
Nodes Poll Until checks a URL until your conditions hold or a deadline passes. Poll Until
Nodes Split in Batches processes items in chunks, with durable checkpoints. Split in Batches
Workflows A badge, a sound and a desktop notification when a run waits for you, with the card on the page you are on. Waiting
Workflows Continue On Fail and per-node retry with backoff. Error handling
Nodes File nodes: Save as File, Convert to File, Extract From File, Compression, Edit Image and XML. Convert to File
Nodes Digest collects items across runs and releases them together. Digest
Nodes Hotkey Trigger starts a workflow from a keyboard shortcut. Hotkey Trigger
Integrations Readwise / Reader, Attio and Pipedrive integrations. Readwise
Integrations Mistral chat and Cohere embeddings. Mistral
Security Anonymous usage telemetry with a Privacy settings section to turn it off. Privacy and data

Improved 11

Workflows A richer {{ }} expression language; $('label') works for labels with spaces. Expressions
Nodes The Code node reads upstream data as input, and Python runs in an isolated worker. Code node
Workflows A redesigned Executions view with an attention band and a docked detail panel. Executions
App Every screen works in the side panel at 320–450 px. Side panel
App An outcome-first first run with a welcome tab and a templates-first home. Quick intro
Local AI Ollama models are detected automatically and appear in the model picker. Ollama
Integrations Google nodes use the right OAuth scopes and refresh tokens on their own; an inline guide creates your OAuth client. Create your OAuth client
Integrations Extension-only integrations work from the web app through a secure fetch bridge. Extension-only integrations
Marketplace Browse without an account, featured and trending workflows, and editing a published listing. Publishing a workflow
Nodes Loop allows up to 1,000 iterations; Screenshot Page uses native capture. Loop
App Heavy libraries load only when needed and only the last 100 executions per workflow are kept locally. Performance

Fixed 6

Workflows No more double runs or re-runs on resume, and resumed runs no longer disappear.
Workflows Schedules handle day-of-month and DST correctly, and keep durable-wait alarms when they sync. Schedule
Workflows Sub-workflow failures and cancels reach the parent, and nested loops rewind correctly.
Security A “Not authenticated” error after sign-in, caused by the storage key being created twice.
Aria & agents The assistant no longer freezes the UI when it opens.
Workflows Opening a past run no longer locks the editor canvas.

In detail

Chat-first Aria, agents and teams

Aria moved from a workflow-drafting assistant to a chat you can open from anywhere:

Chat core with page context and knowledge, reachable from the right-click menu, the address bar (omnibox), a keyboard shortcut, and the notch quick chat.
Agents with their own skills, rules, memory and activity log. Every built-in agent skill was checked and fixed so it actually works.
Live browser control and routines, so an agent can act on the page you’re on, with safety checks.
Teams of agents that hand work to each other in a shared thread.
Projects that group chats, files and agents, with a reorganised side panel and a redesigned project context panel.
An agent and team templates marketplace to start from a ready-made agent.
Characters with living eyes give each agent a recognisable face.
Agent and team workspaces now fill the full height of the window, with pinned side rails and consistent tab headers.

Relevant docs:

Building workflows with the assistant

The workflow builder inside the assistant is easier to follow and to trust:

A thinking timeline shows each build stage as it happens.
A change plan lists what will change before anything is applied, and the canvas highlights the changed nodes afterwards, with a streamed summary of the build.
A real read-only canvas preview in the dock shows the draft workflow.
Deterministic structure fixups and schema-guided decoding make drafts from local models more reliable.
A grouped model selector (Auto / Cloud / On-device), a Local/Cloud indicator, and a nudge to go offline when an on-device model would do.
Ollama models are detected automatically and appear in the model picker.
The assistant no longer freezes the UI when it opens.

Relevant docs:

Local AI and local models

On-device AI now covers far more than chat:

Local AI runs TensorFlow.js and transformers.js models in a worker, with task nodes for text generation, summarising, translation, classification, question answering, image captioning, object detection and audio transcription.
The Local Models page lets you see, install and delete in-browser models, browse the Hugging Face Hub with real model sizes, and filter by task or tool support. A tool-capable badge marks models that can call tools.
Prompt-based tool calling lets WebLLM and Transformers Chat models run the Tools Agent, so agents can work fully on-device.
The app checks your GPU before installing a WebLLM model it can’t run.

Relevant docs:

Durable flow control

Long-running workflows can now pause and resume safely:

Wait of one minute or more survives closing the tab or the extension being unloaded (up to 30 days).
Poll Until checks a URL until your conditions are met or a deadline passes.
Split in Batches processes items in chunks, including durable checkpoints between batches.
Nested and batch suspend: approvals and forms can pause inside loops and sub-workflows.
When a run needs your response, you get a top-bar badge, a sound, and a desktop notification, and the approval or form card can appear on the web page you’re on.
Loop now allows up to 1,000 iterations.

Relevant docs:

Building and running workflows

A richer {{ }} expression language, and $('label') references now work for node labels that contain spaces.
The Code node reads upstream data as input (including fan-in), and Python runs in an isolated worker per run.
Continue On Fail per node, and per-node retry with a backoff strategy and retry-on-condition.
A failed node now stops only its own branch, and a node with no output no longer stops the run.
Output schema contracts validate what a node produces and show expected output previews.
Workflows can declare a minimum extension version, and installing is blocked on older versions.
The canvas never loses an edit: a save status, a leave guard, and Cmd+S / Cmd+Z.
Node notes are hidden by default and revealed from a pin badge; the flow story opens only the first time.
Element Change can watch a list of selectors picked from the page.

Relevant docs:

New nodes

Save as File
Convert to File
Extract From File
Compression (gzip / gunzip)
Edit Image
XML
Text Formatter gained encode/decode operations
Debug Helper
Digest — collect items across runs and release them together
Execution Data
RSS Read (re-enabled)
Hotkey Trigger
Document loaders for the Indexer: URL, File and Current page
Screenshot Page now uses native capture for the viewport, the full page, or one element

Integrations and AI providers

New integrations: Readwise / Reader, Attio and Pipedrive.
New AI providers: Mistral chat, plus Google and Cohere embeddings.
Google nodes use the correct OAuth scopes, refresh their tokens on their own, and explain credential problems clearly.
An inline “Create your OAuth client” guide for Google and Microsoft.
Extension-only integrations (such as Notion) now work from the web app through a secure fetch bridge to the extension, including file uploads and custom hosts for Custom API and Supabase.

Relevant docs:

Files and binary data

Workflows can now carry files. A $binary model with sources, sinks and an output panel preview moves files between nodes; large files are stored locally or in the cloud and expire automatically. Gmail can send attachments and Notion can upload file and image blocks.

Executions

The Executions view was redesigned with an attention band for runs that need you, a sortable table, and a resizable docked detail panel. Opening a past run no longer locks the editor canvas.

Relevant docs:

Executions

Marketplace and publishing

Browse without an account — you sign in only when you act.
Install opens the canvas with a Ready-to-run checklist.
Featured and trending workflows, and a more discoverable request board.
Edit a published listing through the publish wizard, with a pending revision status while it’s under review.
Generate a preview image from your canvas.
A new Explain the steps wizard step; narration is now required to publish.
My published workflows and the workflows list both support multi-select bulk actions (delete, export, move storage, activate/deactivate, tags).
Imported workflows are treated as untrusted until you review them.

Relevant docs:

Side panel and first run

Every screen works in the side panel at 320–450 px.
An outcome-first first run: a welcome tab, a Start here page, and a templates-first home.
A branded loading screen, clearer error states, and no more dead ends.
When the extension updates, the release notes open in a new tab.

Relevant docs:

Firefox groundwork

This release lays the groundwork for a Firefox version of the extension: a Firefox build, a sidebar in place of Chrome’s side panel, sign-in and integration OAuth that work in Firefox, browser-specific copy, and a built-in data-collection consent. The Chrome AI node explains that it only works in Chrome.

Relevant docs:

Browser compatibility

Privacy and security

Anonymous usage telemetry (funnel events and error reports) with a new Privacy settings section where you can turn it off.
The extension validates message senders, the fetch bridge re-checks redirects, and the injected notch and chat are hardened against the host page.
Secrets in the local vault are always encrypted, and the storage encryption key is created only once (fixes a “Not authenticated” error after sign-in).
If sign-in requires a password reset, the app now tells you so.

Relevant docs:

Permissions and security

Reliability and performance

Engine fixes: no more double runs or re-runs on resume, resumed runs no longer disappear, sub-workflow failures and cancels reach the parent, nested loops rewind correctly, and failures are attributed to the node that actually failed.
Scheduler fixes: correct day-of-month and DST handling, and durable-wait alarms are kept when schedules sync.
Faster app: heavy libraries (WebLLM, node SDKs) load only when needed, the chat and notch are injected on demand, the dashboard loads in parallel, the canvas uses a shared graph index, undo history is capped, and only the last 100 executions per workflow are kept locally.

Relevant docs:

Ask Aria