AgentSky
Log inSign up

Changelog

What’s new

Everything that reached you, grouped by theme.

October 2026

Agents & harnesses

  • WebAgents put decisions to you as a choice card — you pick an option and the agent continues without going back and forth.

Mac app

  • MacThe Mac app now asks for permission once per task and shows what it needs before it starts.
  • MacLocal and cloud chats now use the same single account key, so you stay signed in everywhere.

API, CLI & MCP

  • WebAn API key can be created in one click from the developer console.

Channels & connectors

  • WebA service can use your connected apps — Gmail, GitHub and others — under a grant you approve once. Connector events now fire over webhooks when a connection changes.

Agents & harnesses

  • WebBuilt-in help tools are available in every cloud session so an agent can look up how to use the platform without leaving the chat.

Workspace & team

  • WebSettings is reorganised into You, Agents and apps, Team, and Advanced on the web and into You, This Mac, and workspace sections on the Mac.

Mac app

  • WebA chat started on the Mac or in the CLI continues in the cloud with a single send — the session carries over with its files and context.
  • MacWeChat is now a readable source for the Mac app — your conversations can be passed to an agent.
  • MacCall transcription now stays on the Mac that recorded it; a second Mac no longer picks it up.

Agents & harnesses

  • WebNew chat opens with an Agent Market strip across the top and the composer at the bottom. Agent and Model Market is the first row in the sidebar.
  • WebWhile an agent works, the Working block shows its newest three steps and folds the earlier ones behind one line, on the web and the Mac.
  • WebAgents follow the rules and lessons your team keeps in its notes instead of treating them as reference text.
  • MacWith no cloud credit or connected plan for the model, a chat runs the same agent on your Mac and offers credits. A finished turn folds its work into one summary line with its step count, and thinking sits behind Show thinking.

Models & gateway

  • Models50 models join the API, among them GPT-6.1 Sol Pro and GPT-6 Astra Pro from OpenAI, Mistral Large 4 from Mistral, GLM 5.3 Prime from Z.ai, Qwen3.8 Max Prime from Qwen, Command A+ from Cohere and MiMo-V2.6-Pro from Xiaomi. Most also run in Pi, Hermes, OpenClaw and OpenCode.
  • ModelsClaude Code and Codex can run models from the catalog, chosen in the same model menu. The Model Market shows the newest model of each series; older ones stay callable through the API and keep their pages.

Mac app

  • MacAgents can work in an app on your Mac in the background, one app at a time, without taking your pointer or keyboard. You allow each app the first time it is used, and Esc stops the agent.
  • MacAn Allow AgentSky sheet asks for notifications, microphone, accessibility and screen recording up front and says what each is for. The app no longer disappears from the Dock after launch.
  • MacDictation hears Chinese and English, with a switch while you talk, keeps takes longer than a minute and says why when it fails. You can use Apple's engine or a SenseVoice or Paraformer model that runs on the Mac, choose a microphone and add your own words.

API, CLI & MCP

  • API & CLIsky 2.17.0 and 2.18.0: sky import uploads the whole working folder with the conversation, credentials masked, and files every import in your trace library. sky files upload takes any number of files and folders, sky files usage shows the space used, and sky session model switches the model and thinking level.
  • API & CLIGET /v1/files/usage reports a workspace's file count and space used. Calls that mark prompt caching on Grok 4.7 and other models priced in tiers are billed instead of refused, and supplier errors come back as standard messages.

Memory, skills & files

  • WebThe Workspace Drive takes any number of files and folders, each up to 50 GB, on the web and the Mac. Uploads run four at a time and resume after a dropped connection, and the Drive shows the space the workspace uses. Chat attachments over 10 MB no longer get cut off.
  • WebSearch has an Everything mode that looks through the workspace's chats, trace library, shared artifacts and your memory, on the web and the Mac. An agent's grep everything searches the same workspace, also in sessions started from the API or the CLI.

Agents & harnesses

  • WebPi, Hermes, OpenClaw and OpenCode can run every text model in the catalog that takes tool calls. The model picker lists them by maker with input and output prices, on the web and the Mac, and the Model Market lists every model in the catalog.
  • WebLong chats open on their newest messages instead of loading the whole history first, and Show earlier messages brings in the rest. The Mac app opens chats the same way.

Models & gateway

  • ModelsMiniMax M3.1 Flash from MiniMax, Muse Spark Open from Meta and Nemotron 3.5 Super VL from NVIDIA are on the API and in Pi, Hermes, OpenClaw and OpenCode. MiniMax M3.1 Flash also runs in MiniMax Code, at $0.165 input and $0.66 output per million tokens.

Workspace & team

  • WebA workspace owner can delete a workspace and everything in it from Settings Members Danger zone, on the web and the Mac. Running tasks stop and purchased credit goes back to the owner.

Billing & plans

  • WebA getting-started card on the new-chat page shows the credit each connected app earns and connects GitHub or Gmail in one click. A new account's first task waits for one connection, then starts on its own.

Mac app

  • MacWord documents, Markdown and web pages from a chat can be edited in the app with the same editors as the web. Side chat answers questions about a chat without interrupting the agent.
  • MacBilling, top-ups, plan sign-ins, Import sessions and Settings work inside the app instead of the browser. You can share a chat, comment on a Preview page, and open Code, Repository, Background tasks, PDF, Shared files, Automations and Code Review tabs in the side panel.
  • MacmacOS notifies you when an agent needs your answer or finishes in the background, including chats started elsewhere. The installer shows you how to drag the app to Applications and offers to keep it in the Dock.

Mobile

  • iPhone & AndroidAgentSky for Android 2.0.0 is out, and agentsky.dev links to chats, shared pages and settings open in the app. agentsky.dev/download offers the right app for Mac, iPhone or Android.

API, CLI & MCP

  • API & CLIsky 2.16.0 adds commands for billing, API keys, memory, webhooks, channels, connectors, vaults, routine triggers, session files and Plan mode. Files up to 50 GiB upload in resumable chunks through the API and sky files upload.
  • API & CLIsky import brings a local Claude Code or Codex session's files, git state, skills and remote MCP servers, and the continued chat restores them in its workspace under the session's own title. First-run Import on the Mac also brings remote MCP servers, shared instructions and skills.
  • API & CLIThe API and MCP server add billing (auto top-up, usage export, top-ups with the saved card), personal API keys, workspace deletion, and MCP tools for goals, routines, usage and sharing. A session can switch its model or move to another agent by its handle.

Reliability & fixes

  • WebStop now stops a task even while its computer is still starting, and servers an agent starts keep answering after its turn ends, so preview links stay up.
  • ModelsMuse Code answers again, and OpenClaw and Hermes no longer stop partway through tool use on Gemini models.

Models & gateway

  • ModelsMany more models, including open-weight ones, are available through the API and in agents. Prices update automatically.

Billing & plans

  • WebCredits can be bought without leaving agentsky.dev — a top-up opens in place instead of a separate tab.

Channels & connectors

  • WebSocial connectors let you connect several accounts on the same platform, including X.

Memory, skills & files

  • API & CLIWorkspace trace library is now searchable and exportable: upload any local agent session and search across all of them on the web, Mac and CLI.

Agents & harnesses

  • WebEnvironment tools — browsing, shell and file access — have a dedicated settings panel where you can grant and revoke them per session.
  • WebShared sessions now have a read-only embeddable page, so you can show a live run without giving edit access.

Agents & harnesses

  • WebNavigation lives in the sidebar, with the agent and repo picker folded into the composer and the Agent Market strip right below it.

Mac app

  • MacAgentSky for Mac gets the same home and task chips as the web, Market cards in the sidebar, workspace stamps on session rows, and notifications when a task finishes in the background.

Models & gateway

  • ModelsGemini 3.5 Flash Lite, Gemini 4 Argon and Muse Wonder 1.0 join the API.

Workspace & team

  • WebEvery workspace now works the same way: agents, browser tasks, GitHub access, Arena, and the vault all work in any workspace a member has access to. The Mac app can create a team workspace and connects to the one you select.
  • WebMail and calendar connections can be shared with a workspace so every member's agent can use them.

Agents & harnesses

  • WebBrowser profiles save logged-in sessions so an agent can open apps without signing in again. Profiles extend to Claude Code and Codex running in the cloud.

Arena & benchmarks

  • WebArena battles now show side by side so you can watch two agents work simultaneously. A vote of three or more agents keeps the full ranking.
  • WebThe leaderboard ranks the agent and the model separately and shows how long each seat worked.

Workspace & team

  • WebThread sharing now distinguishes Private, Can continue and Can view, with a redesigned share dialog and Download option.

Billing & plans

  • WebConnecting apps now earns $4 of credit per app, up to $12 for three apps. Free credit earned through connected apps cannot be spent through API keys.

Mac app

  • MacAgentSky for Mac 0.4.14 lands local Claude Code and Codex: your installed agents run from the app without leaving your machine. First run imports this Mac's agents and connects their plans.

API, CLI & MCP

  • Websky import brings a local session's files, environment variables and git state into a new cloud session. sky clone vaults secrets from dropped MCP servers before handing over the session.
  • API & CLIsky 2.8.0 adds workspace member management, policy rules, task files, and artifact commands.

Agents & harnesses

  • WebPlan mode is available across every agent. A Market listing opens in place with its own landing page when signed out.

Arena & benchmarks

  • WebSigned-out visitors can compare agents on the Arena page. Image, video and audio models are available to buy and use alongside agents in the Market.

Mac app

  • MacThe Mac app runs Claude Code and Codex locally — they appear beside the cloud agents in the picker and run on your own hardware.

Automations

  • WebAutomations can be created from Settings, with cards and routine controls surfaced directly in the conversation.

Agents & harnesses

  • WebOne start surface for every visitor: the composer and Agent Market open together whether you are signed in or not. Agent Market entries now carry official vendor marks, maker name and price.

Models & gateway

  • ModelsEvery GPT model and Gemini 3.8 Flash are available on Pi. Pi also supports remote MCP servers.

Workspace & team

  • WebTeammates can be invited with a link from an always-visible entry in the sidebar.

Billing & plans

  • WebA ChatGPT plan powers Pi, Hermes, OpenCode and OpenClaw — sign in once and your plan covers all four.

API, CLI & MCP

  • API & CLIThe MCP server surfaces the developer API as tools for agents. Web search is now available in every session.

Memory, skills & files

  • WebOne Memory list for you and your workspaces, shared by everyone in the workspace. Files, Memory and Skills pages follow the same compact list layout.

Automations

  • WebHooks on agents: set them from the API, edit them in the console, and watch each execution as a readable row in the transcript.

Models & gateway

  • ModelsGrok 4.7 and xAI Speech-to-Text 2.0 join the catalog.

Workspace & team

  • WebWorkspace members can share and view threads, and members see all of a workspace's content. A redesigned Members page shows one list with an invite dialog.
  • WebA workspace can now set a spend cap that applies to every session run inside it.
  • WebArtifacts can be shared with a workspace so every member can view them.

Billing & plans

  • WebBring Your Own Subscription opens inside the consumer app. API usage now sits inside Usage and billing with two subtabs.

Agents & harnesses

  • WebAgents can now ask you a question with a set of choices, wait for a goal condition, or show you a deadline before continuing.

Mac app

  • MacDictation, notifications, and This Mac capabilities ship for the Mac app. Sign in to the Mac app through the browser's existing session.

Channels & connectors

  • WebEmail channel: connect your own email address and let agents send and receive email on your behalf. Google Chat apps and Buzz groups are now available as channels.

Memory, skills & files

  • WebTeam memory stores can be created and managed from Settings, and agents use the workspace's memory in every session.

Automations

  • WebGoals can now have a maximum number of steps and stop automatically when no progress is made. Webhook events fire when a goal completes, stops, pauses or resumes.

September 2026

Agents & harnesses

  • WebAgents can orchestrate child sessions — one session can spawn and direct another to get sub-tasks done in parallel.

Workspace & team

  • WebOne workspace navigation is live: four main tabs with API docs inside Developer and API, and a stable Settings list.

Models & gateway

  • ModelsDeepSeek, GLM and Kimi are now available in sessions, and Wan text-to-video is available as a tool agents can use.

Arena & benchmarks

  • WebArena results now show the output an agent produced, not just the bar chart.

Billing & plans

  • WebPer-request API token costs now appear in the request history, so you can see exactly what each call cost.

Agents & harnesses

  • WebAccount-wide usage summary and export are available across sessions and API keys, on the web and through the API.
  • WebHTML artifacts can be edited inline with persistent design changes. Spreadsheets, audio and unknown files are readable in the preview pane.

Billing & plans

  • WebVerified newcomers receive $4 of welcome credit automatically when they sign up.

Reliability & fixes

  • WebA refused task stays editable so you can fix it and resubmit. Failed runs are excluded from cost comparisons.

Agents & harnesses

  • WebYour subscription account's remaining allowance is visible beside the model picker, so you know what is left before switching to paid tokens.

Channels & connectors

  • WebOne Social media connector replaces the individual platform connectors, named and placed like the other app connectors.

Memory, skills & files

  • WebSkills can now be picked for a task individually, and workspace skills reach every agent in the workspace.

Automations

  • WebSchedule cards and routine controls appear in the conversation, and automations can be managed from Settings.

Agents & harnesses

  • WebAny visitor can create their own agent in the workspace and start a session with it immediately.

Arena & benchmarks

  • WebEach Arena contender now names the thinking level it runs at.

Workspace & team

  • WebGitHub can be connected inside the repository picker. Optional session budgets let you cap what each session spends.

Memory, skills & files

  • WebArtifact export is available in a format matrix; Download asks which format you want and only when asked.

Agents & harnesses

  • Webagentsky.dev/clone.md lets you clone a local agent session to the cloud with one pasted prompt, no account required.

Models & gateway

  • ModelsChatGPT plan sign-in opens in place on the settings page, with a full end-to-end CLI path for manual auth upload.

Models & gateway

  • WebClaude and GPT models now use your connected subscription plan first, before billing additional tokens.

Billing & plans

  • WebCompute is billed only while a session is active — idle time is free.

Arena & benchmarks

  • WebAn Arena match carries files, so an agent can work with the files you attached when comparing agents.

Memory, skills & files

  • WebSkills are now managed in the workspace and can be installed from uploaded files in the composer.

August 2026

Agents & harnesses

  • WebThe composer now has an attached tab, quick access to files and tools, and an agent-icon trigger for the Model Market.

Models & gateway

  • WebModels and tools are sold beside agents in the Market, and can be used in the composer right from the listing.

Workspace & team

  • WebOne workspace navigation is live: four main tabs, stable Settings list, and API docs as the first developer tab.

Agents & harnesses

  • WebA task-description generator is available when you are not sure what to ask. The Market leads with usage-based results and concrete tasks.

Arena & benchmarks

  • Webagentarena.us opens directly on the Arena. Arena seats show the agent and model name in the side-by-side layout.

Agents & harnesses

  • WebA resizable right-side panel lands with Preview, Background tasks, and a Markdown reader pane. Code traces merge work and reasoning, with verbose output on request.

Agents & harnesses

  • WebChats now open faster: the session starts showing immediately while the rest loads in the background.

Arena & benchmarks

  • WebThe voted Arena result stays on the thread after reload.

Agents & harnesses

  • WebPublic agent catalog launches with agent pages, model pages and tool pages. Comparison pages for AgentSky vs other platforms are live.

API, CLI & MCP

  • WebMachine-readable pages go live for AI crawlers: benchmarks, pricing, use-cases and case-studies each have a plain text version.

Agents & harnesses

  • WebAgentSky unifies the website and product experience. Agent runs read like Claude Code — playground columns and Quick Chat.

Reliability & fixes

  • WebTightened access so the files an agent produces are readable only by that agent's owner.
MarkdownRSS feed