Changelog
Everything we ship to the YoloSesh app and this site, newest first. Preview software, so expect rough edges — grab the latest build anytime.
- Agent
Private by default, and a timeline for every turn
- AgentAddedAgentTurn on Waterfall in the Activity panel to see each turn on one timeline: every model request, tool call and helper as a bar, failures marked, and a live timer while it runs.
- ModelsImprovedModelsOpenRouter requests now skip providers that store or train on your prompts, and never go to China-hosted providers. Every model stays available through other hosts, and you can change both in routing settings.
- AgentAddedAgent/compress summarizes a long chat into its key points so the model keeps working with a smaller context. Add a focus, like /compress keep the auth decisions.
- AgentFixedAgentTool calls the model sends in the wrong shape are repaired instead of failing, even later in a long turn.
- WorkersImprovedWorkersMore helpers run at once by default: up to 4 builders and 8 explorers, with more tool calls in parallel.
- WorkspaceFixedWorkspaceA chat whose first message fails keeps its name, Claude Code and Codex only show as connected once they are signed in, and an error shows once instead of four times.
- PerformanceImprovedPerformanceThe git panels no longer wait on GitHub to refresh, and a brief network drop no longer fills the logs with errors.
- Agent
- Agent
A clearer view of your work
- AccountFixedAccountSign in again from the app after an expired or rejected session, with a clear explanation when account access needs attention.
- AgentImprovedAgentA cleaner composer keeps your mode, model, attachments and active work in reach. Prompts stay in their own bubble, and tool activity opens for detail when you need it.
- WorkersImprovedWorkersFollow each workstream with its goal, model, progress and definition of done. Inspect a helper's chat, steer its next step or stop work from the same view.
- WorkspaceImprovedWorkspaceGlobal updates and team conversations have clearer unread states, reactions and draft recovery, with better feedback when a connection or action needs attention.
- AgentFixedAgentQueued prompts, image attachments and live permission changes now follow the current chat reliably. Private chats no longer send memory-recall queries outside the device.
- DataAddedDataOpen chat insights from the sidebar to review recorded usage, cache behavior and tool activity, with an optional analysis and clear labels when information is unavailable.
- Account
- Models
Costs that match your bill
- ModelsFixedModelsCosts now match what your provider charges, checked against OpenRouter's own records. Cached tokens are counted once, retries add up every attempt, and estimated costs are labeled as estimates.
- DataImprovedDataEverything that spends tokens shows up in Usage, including code search embeddings, memory and background decisions, and a helper agent's cost counts toward its parent chat's limit.
- SiteFixedSiteLeaderboards count each request once, even when it went through Compete.
- AgentImprovedAgentTimes read the same everywhere (59s, 5m 07s, 1h 05m), code search shows every match with the right context, and Stop ends a turn within a second without losing what was already written.
- Models
- Agent
Search code by meaning, issues that become pull requests
- AgentAddedAgentThe agent can now search your code by meaning as well as by keyword, so a question like "where do we refresh sign-in?" finds the right files even when they never use those words. The index stays on your Mac and updates as you work.
- AgentImprovedAgentOnly real project code goes into that search: secrets, .env and key files, lockfiles, build output, vendored and generated code, and data dumps are left out, and any token in your source is blanked first. Add paths to .cursorignore or .yolosesh/indexignore to leave out more.
- AgentAddedAgentReview changes one piece at a time: every block in the Changes diff has its own Accept and Reject buttons.
- IntegrationsAddedIntegrationsGitHub issues can become pull requests (preview, off by default): label an issue "yolosesh" and the app works on it in its own branch and opens a PR that fixes it.
- AgentAddedAgentA command-line runner for scripts and CI: yolosesh run "<prompt>" runs one task and exits with a status code your pipeline can check.
- Agent
- Agent
Memory that sticks, finer model control
- AgentAddedAgentThe agent can now keep memory notes across chats: things worth remembering about your project and preferences are saved as short notes, and the ones a project needs live in the project folder so they travel with it. Every new chat starts with the notes already in mind, and there's a Memory panel in the dock to browse, edit or delete them.
- ModelsAddedModelsPer-assistant model controls: in Multitask models you can now pick a specific model, speed and reasoning effort not just per category but for each background assistant (planner, researcher, editor and the rest), with quick reset chips to drop back to the category default.
- DataImprovedDataLocal history stops piling up: chat search, run logs, usage records and schedules now keep only a window of recent data (weeks, not forever), and the app quietly prunes the rest in the background while it runs. Everything you actively use stays untouched.
- PerformanceImprovedPerformanceA round of background fixes keeps long-running sessions tidy: worker bookkeeping gets cleaned up when a session ends, repeated config reads are cached, and a slow or stalled remote connection now pushes back instead of buffering without limit.
- Agent
- Models
Steadier model streams, sharper tool search
- ModelsFixedModelsReasoning models that go quiet for minutes while they think no longer get cut off mid-thought. The idle cutoff for model streams is now five minutes, so long silent reasoning survives and your answer arrives instead of an error.
- ModelsFixedModelsIf you explicitly pick a model, the app keeps it — even when it drops off a provider's fallback list. Your choice only changes when you change it.
- ModelsAddedModelsLong chats that get condensed to stay under the context limit are now summarized by a small, fast model instead of your main one. Same result, a fraction of the cost and wait — and switching your main model no longer changes what does the condensing.
- IntegrationsImprovedIntegrationsThe plugin, server and skill catalog got its refresh: 500+ listings, each with its health checked, so what you browse is what installs.
- AgentImprovedAgentSearching connected tool servers now ranks the hits, so the tool you're after is near the top instead of buried under dozens of loose matches. Tool search in research mode is also less locked down.
- WorkspaceImprovedWorkspaceThe goal strip stays honest when your todo list is long: it shows how many items it cut off and how many are still open, instead of silently hiding the rest.
- IntegrationsAddedIntegrationsAutomations get a Run-on picker, clearer error surfaces when a run fails, and timer fixes, so a scheduled run does what you expect when it fires.
- SiteImprovedSiteTightened the app's web-content rules and raised its window size floor, part of keeping the app current with its platform updates.
- Models
- Agent
Challenge: critics that test the work
- AgentAddedAgentChallenge, a new option in the goal step: read-only critic agents try to poke holes in your plan, the finished result, or both before the goal is marked done.
- AgentAddedAgentChoose how many challenge rounds to run (1 to 5) and when: after each wave of workers, or once after all of them finish.
- AgentAddedAgentCritics back every finding with evidence from your code, library docs, the web or public code search, and suggest a check that would settle it. Findings without evidence are dropped.
- AgentAddedAgentThe agent has to fix or explain every serious finding before it can finish. Challenge is off unless you turn it on.
- WorkersImprovedWorkersWorkers started from a folder that holds several projects now work in the one project their task names, and say clearly when a task spans more than one.
- ModelsImprovedModelsEach worker's model is now picked by how hard its task is, so simple lookups run fast and harder work gets a stronger model.
- Agent
- Workers
Longer answers, steadier workers
- AgentFixedAgentLong answers no longer stop halfway. When a reply hits the length limit, the agent picks up where it left off.
- WorkersFixedWorkersWorkers that write out a tool call as text instead of running it get one nudge to fix it. If they do it again, they're reported as failed, not done.
- WorkersImprovedWorkersWhen a sub-agent fails, its error is carried into the final report so you can see what went wrong.
- PerformanceImprovedPerformanceLong sessions respond faster because more of your earlier context is reused between turns.
- DataAddedDataDoctor can compact and tune your local databases to free up space and keep the app quick.
- PerformanceImprovedPerformanceOn a device you've already signed in on, the app starts working right away while your sign-in is checked in the background.
- Agent
- Browser
One screen to get started
- AccountImprovedAccountSigning in and accepting the Terms now happen on one screen.
- BrowserFixedBrowserBrowser tasks no longer give up after a minute. They get the full time a task can take.
- AgentImprovedAgentWhen a tool call times out, the message says where to raise the limit in Settings.
- BrowserFixedBrowserThe agent stops suggesting browser commands that aren't installed on your Mac.
- WorkersFixedWorkersThe worktree guard understands variables set earlier in a command, so it catches writes that would land outside a worker's own copy.
- Account
- Editor
Edits that don't get stuck
- EditorFixedEditorEdits stopped looping. If an edit fails, the agent hears why instead of trying the same thing again.
- EditorImprovedEditorThe editor shrugs off small formatting slips and points to the closest matching line when it can't find your text.
- WorkersImprovedWorkersOrchestration won't start two workers on the same task. Failed sub-agents show as failed, in the order they started.
- ModelsFixedModelsA rejected model key is no longer retried over and over, and the error says which connection needs attention.
- Editor
- Agent
Session goals
- SiteAddedSiteThis changelog. Every app and site change, newest first.
- AgentAddedAgentSession goals. Describe what done looks like, and the agent keeps a live check on it, with the goal pinned until it's met.
- WorkersAddedWorkersThe side chat shows a live, step-by-step timeline for the sub-agent that's running.
- AgentImprovedAgentLong commands move to the background after two minutes. You get your turn back, and the agent can check in on them later.
- WorkspaceFixedWorkspaceYou can select and copy chat text in the app again.
- AgentFixedAgentStarting a brand-new chat no longer crashes when it checks for a session goal.
- Site
- Agent
Editor mode
- AgentAddedAgentEditor mode joins the mode list. Build, Plan, Multitask and Orchestrate got reordered to match how people use them.
- AgentImprovedAgentExternal coding agents you connect show the same tool names as the built-in agent.
- WorkspaceFixedWorkspaceA copy that fails now tells you with a toast instead of quietly doing nothing.
- Agent
- Models
Quicker model responses
- ModelsAddedModelsAn Optimize button that tunes each model for the quickest, most reliable responses.
- AgentImprovedAgentCompaction keeps more of the summary, so long chats don't get cut off mid-thought.
- WorkspaceFixedWorkspaceThe glass sidebar looked gray in light mode. It's translucent again, and text stays easy to read.
- PerformanceFixedPerformanceThe app recovers on its own when the background service refuses a launch.
- DataFixedDataBackground service logs rotate while it runs, so they no longer grow without limit.
- Models
- Models
Edit files in the side panel
- EditorAddedEditorThe file panel on the right is a real editor now. Tweak a file without waiting for a turn.
- ModelsAddedModelsVideo generation models are in the catalog, along with more ways to bring your own key.
- WorkersAddedWorkersThe work board tracks each step in more detail and lets you cancel a task.
- ModelsFixedModelsIf a model can't take images, the request is retried without them instead of failing the turn.
- AgentFixedAgentA queued prompt no longer jams the composer after a turn errors out.
- Editor
- Site
A weekly digest
- SiteAddedSiteA leaderboard digest email with new releases and the week's top builders.
- SiteImprovedSiteThe download page picks the right build for your Mac's chip.
- WorkersFixedWorkersThe work board stays current instead of showing stale steps.
- Site
- Agent
Model picker and workers
- ModelsImprovedModelsThe model picker shows recommended picks and full details for every model, with a loading state while it fills in.
- WorkersAddedWorkersYou choose whether sub-agents show up in the sidebar, and two workers can't step on the same checkout.
- AgentAddedAgentPlans belong to the chat that made them, with a plan command to bring them up.
- PerformanceImprovedPerformanceChats load faster and keep their sidebar order across restarts. Automatic titles stopped repeating or coming out empty.
- AgentFixedAgentPlans saved under another project open correctly.
- AgentAddedAgentA Stop button in the composer, plus progress notices while the agent writes out a long tool call.
- WorkspaceAddedWorkspacePermission chips in the composer and info tips throughout Settings.
- IntegrationsAddedIntegrationsA summary of every MCP tool you have, and usage insights for your sessions.
- BrowserAddedBrowserThe in-app browser can load extensions.
- WorkspaceFixedWorkspaceVery short runs no longer make the chat list flash, and renaming a chat by hand works again.
- AgentFixedAgentIf a chat was unloaded during a restart, it reloads and retries your turn instead of failing.
- PerformanceImprovedPerformanceCheckpoints skip files that haven't changed, so saving a snapshot is quicker.
- IntegrationsFixedIntegrationsServices finds command-line tools installed with Go.
- Models
- Site
A more accurate leaderboard
- SiteFixedSiteYour personal totals match the public ranks now, across refreshes and live tiers.
- SiteFixedSiteThe leaderboard explains when you're not ranked yet, and spend shows in US dollars.
- AccountImprovedAccountWork you've already authorized keeps running on your Mac even when our servers are down.
- PerformanceFixedPerformanceChats open while the app checks for updates, and first launch is much less sluggish.
- DataFixedDataPrivate skills and MCP activity stay on your Mac.
- Site
- Agent
External coding agents, natively
- AgentAddedAgentConnected coding agents use their own local sign-in. No API key needed.
- ModelsAddedModelsNative connections for external coding agents, with a choice of reasoning level.
- EditorImprovedEditorText edits are more reliable, and an interrupted stream no longer messes up the next turn.
- AgentImprovedAgentLong tool output is kept and searchable, so the agent can look back at it.
- AgentFixedAgentQueued chats and resumed jobs keep their place through app rebuilds.
- Agent
- Workspace
SSH and privacy
- WorkspaceFixedWorkspaceImporting SSH hosts and opening remote folders works reliably. Previews and downloads keep the right workspace.
- DataImprovedDataConnections you mark private stay out of reporting, and their details never leave your Mac.
- AgentFixedAgentChats show the turn that's actually running instead of a stale running status.
- SiteFixedSiteModel filters apply to spend rankings on the leaderboard.
- Workspace
- Agent
An agent that learns
- AgentAddedAgentThe agent proposes lessons from your past sessions, kept on your Mac, and you review how they worked out.
- Agent
- Workers
Queues you can act on
- WorkersAddedWorkersQueued work is actionable, and worker approvals show up where you'll see them.
- WorkspaceImprovedWorkspaceBetter recovery for remote sessions, and attention alerts on your phone.
- AccountImprovedAccountThe app and web updates apply when you sign in.
- Workers
- Workspace
Motion, glass and steadier connections
- WorkspaceAddedWorkspaceA glass sidebar, on by default on macOS 26 and later, tinted so text stays readable over your wallpaper.
- WorkspaceImprovedWorkspaceSpring motion across the app. Transitions follow your click or keypress, and messages slide in without jumping while they stream.
- PerformanceImprovedPerformanceIf the app loses its connection, your project and draft stay put. A clear notice offers Retry and help.
- BrowserAddedBrowserResearch chats work with plain web links, no search setup required.
- BrowserFixedBrowserSandbox onboarding, bookmark import and resuming a connected coding agent all work from a fresh sign-in.
- Workspace
- Agent
Canvas cards and your Mac from your phone
- AgentAddedAgentCanvases show up as a glass card right in the chat.
- IntegrationsAddedIntegrationsServices became Providers, with a page for each connection.
- WorkspaceImprovedWorkspaceCalmer step rows, a sticky prompt for each turn, and a grouped skills picker.
- AgentImprovedAgentNew chats start in Build mode.
- IntegrationsImprovedIntegrationsThe MCP directory keeps only the best servers in each category, and page reading for research works without any key.
- AccountAddedAccountPair a phone with your Mac using a QR code and a two-way confirmation, then open your Mac from the web app.
- SiteAddedSiteSee and revoke paired devices on the site, and choose which folders remote access can reach.
- SiteAddedSiteThe web app installs to your home screen and fits a phone.
- AgentAddedAgentAutomation suggestions drawn from your recent chats, showing where each one runs.
- Agent
- Workspace
A big polish day
- ModelsAddedModelsA new model picker with favorites, recents and an effort dropdown.
- IntegrationsAddedIntegrationsWhen a tool needs you to sign in, an inline card asks, and the agent waits.
- AgentAddedAgentTurn skills on, off or remove them in bulk, plus a Stop everything button.
- SiteImprovedSiteA cleaner download page, and a marketplace with filters, sorting and icons.
- SiteFixedSiteEvery member shows up on the global leaderboard, and turning sharing off in the app is respected.
- WorkspaceAddedWorkspaceOne shortcut system for the whole app. Hold Command to see key badges in the sidebar.
- WorkspaceImprovedWorkspaceShort automatic chat titles, one-line sidebar rows and a readable hover card.
- WorkspaceAddedWorkspaceLinks to pull requests, issues and commits show as tidy reference chips in chat.
- Models
- Workspace
Orchestration and secrets
- WorkersAddedWorkersOrchestration mode, where a coordinator splits up the work and hands it to workers.
- IntegrationsAddedIntegrationsPull secrets from the password manager you already use.
- WorkspaceAddedWorkspaceUpdate notifications, a restart from the menu bar, and your sessions restored after relaunch.
- AgentAddedAgentPlugin and settings suggestions, each with a reason.
- WorkspaceImprovedWorkspacePasting a big block collapses it into a neat placeholder in the composer.
- Workers
- Workspace
Services and slash modes
- IntegrationsAddedIntegrationsA Services view for the code hosting and deploy command-line tools you already use.
- AgentAddedAgentSlash commands to switch modes: /plan, /multi, /build, /ask and /explore.
- SiteAddedSiteThe plugin marketplace is served from yolosesh.com.
- WorkspaceImprovedWorkspaceLayout and browser state are remembered per project.
- WorkspaceAddedWorkspaceThe app version shows in the lower corner.
- Integrations
- Account
Teams and Doctor
- AccountAddedAccountTeams. Create one, invite people by email, and share a workspace on the web and in the app.
- PerformanceAddedPerformanceDoctor checks your PATH, shell and runtimes, and explains what it finds.
- BrowserAddedBrowserAn agent browser, with sign-in and consent steps, in Settings.
- IntegrationsAddedIntegrationsA CI runs panel and a language server status menu in the side dock.
- Account
- Site
Sign in with a code
- AccountAddedAccountSign up on the site with a code sent by email.
- SiteAddedSiteSeven new featured automations, each listing the tools it needs.
- SiteFixedSiteThe in-app leaderboard updates live again.
- AgentImprovedAgentFeatured automations reply with just their report.
- Account
- Site
YoloSesh on the web
- SiteAddedSiteUse YoloSesh in your browser at yolosesh.com/app, with a live preview at /demo.
- AgentAddedAgentBuilt-in go-to-market skills with their own section.
- IntegrationsAddedIntegrationsAn MCP menu in the composer that shows connected and disconnected servers.
- SiteAddedSiteBadges on the site for the tiers you've earned.
- WorkspaceAddedWorkspaceThe app includes its license and third-party notices.
- Site
- Workers
Multitask and SSH
- WorkersAddedWorkersMultitask mode. Background sub-agents do the work and report back to the lead.
- WorkersAddedWorkersWriting workers get their own branch, and you review, apply or discard their changes.
- WorkspaceAddedWorkspaceAdd an SSH host in a few clicks, with config import and a connection check.
- WorkspaceAddedWorkspaceA fast command palette that searches all your chats.
- SiteAddedSiteAppearance settings, plus custom handles on yolosesh.com profiles.
- Workers
- Workspace
A real workspace
- WorkspaceAddedWorkspaceAn integrated terminal, in-app browser tabs and a file tree, side by side with the chat.
- AgentAddedAgentAutomations. Describe what you want and get a setup, an editor and an honest run history.
- WorkspaceImprovedWorkspaceA slim composer with a + menu, and a sidebar that lists recent chats first.
- IntegrationsFixedIntegrationsMCP servers connect reliably, and tool approvals show on every window.
- PerformanceFixedPerformanceThe app no longer tries to index your whole disk.
- Workspace
- Agent
Day one
- AgentAddedAgentYoloSesh arrives: a fast, local-first coding agent for the Mac that keeps your code and chats on your machine.
- AgentAddedAgentCanvases: live, sandboxed pages the agent builds for you, with hot reload.
- WorkersAddedWorkersGit worktrees, so sub-agents can work without touching your checkout.
- PerformanceImprovedPerformanceLong chats scroll smoothly, and code highlighting runs off the main thread.
- EditorAddedEditorCode-aware tools: language servers and structural search.
- Agent
Older preview notes live in the repo history. Requests?Vote on what ships next.