What's new in Houston

Houston 0.4.22

A release about choice and recovery. Switch the AI behind a chat without starting over, give each scheduled task its own brain, and get clearer prompts when a provider needs you to reconnect. Drop-in on top of 0.4.21, no breaking changes.

Added

• Switch the AI mid-conversation. You can move a live chat to a different AI provider at any time without losing the thread. Houston carries your conversation over to the new provider: the full history when it fits, or a short summary (with your go-ahead) when it does not. Handy when one provider runs out of credit, or you just want a second opinion partway through.
• Each scheduled task can pick its own AI. Routines can now choose their own provider, model, and effort level instead of always inheriting the agent's default. Pin a faster, cheaper model for a routine check-in, or a higher-effort one for the heavy weekly job. Leave it untouched and it keeps following the agent.

Fixed

• Houston picks the AI you actually use. Agents with no provider set used to fall back to Claude, which failed if you only use OpenAI. Houston now falls back to your last-used or only connected provider, so agents from the Store, scheduled tasks, and quick sends all start on the right AI.
• A clear reconnect prompt for OpenAI. When OpenAI ends your login session on their side, the chat showed a red border but no way to fix it, and any hint vanished on reload. Now you get the same inline reconnect button Claude already had, and it sticks around until you reconnect.
• Clearer messages when Claude is rate-limited. A Claude "too many requests" response used to show up as an unknown error with a "Report bug" button, sometimes doubled. It now reads correctly as a temporary rate limit, shown just once.
• Your own message shows in the mission card. Starting a task from a skill used to show a raw behind-the-scenes marker instead of what you wrote. The mission card now shows your message.
• Tool steps no longer take over the chat. A long list of an agent's tool steps could expand and crowd out the conversation. That panel is now capped, so the chat stays readable.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows x64: auto-update from 0.4.7 or later pulls you forward automatically. The MSI is still not OS-code-signed (SignPath integration still pending), so Windows SmartScreen may warn on a fresh install.
• Windows ARM64: there is no auto-update path from an emulated x64 install. Download the ARM64 MSI manually, uninstall the x64 build first, then install the ARM64 one.

Houston 0.4.21

A reliability release. A more reliable Stop button, installed skills show up, and a few provider papercuts are gone. Drop-in on top of 0.4.20, no breaking changes.

Changed

• Latest Claude Code under the hood. Houston now bundles the newest version of Claude Code that powers your Claude agents, so they run on Anthropic's most recent fixes and improvements.

Fixed

• More reliable Stop button. The Stop button already worked, but in some cases an agent could keep going in the background, especially when it had started subagents or other helper tasks. Stop now reliably halts the agent and everything it started, even mid-thought or right after you send a message. While an agent is working, the send button also shows a clear stop icon, so you always know how to halt it.
• Agents with a lot of context start again on Windows. A busy agent that had built up a large amount of memory and notes could fail to start on Windows. That is fixed, so even your most-used agents launch normally.
• Installed skills show up. A skill you install now appears for your agent right away and stays in sync if the agent later updates it. This was most noticeable on Windows, where installed skills could stay hidden.
• The Connect prompt only shows where it belongs. A "Connect Claude" button used to appear in chats that do not use Claude, like an OpenAI chat, and stick around. Now a reconnect prompt only ever shows for the provider that chat actually uses.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows x64: auto-update from 0.4.7 or later pulls you forward automatically. The MSI is still not OS-code-signed (SignPath integration still pending), so Windows SmartScreen may warn on a fresh install.
• Windows ARM64: there is no auto-update path from an emulated x64 install. Download the ARM64 MSI manually, uninstall the x64 build first, then install the ARM64 one.

Houston 0.4.20

Claude Fable 5, Anthropic's most capable model, is now a pick for your agents. Plus a simpler way to set custom routine schedules and a couple of chat fixes. Drop-in on top of 0.4.19, no breaking changes.

Added

• Claude Fable 5. Fable 5 is Anthropic's most capable model, and you can now choose it for any agent from the model picker. It is the strongest option for hard, multi-step work. One thing to know about cost: Fable 5 uses about twice the credits of Opus 4.8 per run, so reach for it when a task really calls for it and keep Opus 4.8 or Sonnet 4.6 for everyday work.
• Simpler custom routine schedules. Setting a custom schedule is now a plain "Repeat every N minutes, hours, days, weeks, or months" picker instead of anything cron-shaped. Choose weeks and you get a day-of-week selector with Every day, Weekdays, and Weekends shortcuts. Choose months and you pick the day of the month. The plain-language summary updates as you go, so you always see exactly when the routine will run.
• Routines now speak Spanish and Portuguese. The Routines tab, the schedule builder, run history, and the plain-language schedule summaries are all translated, so routines read in your language from end to end.

Changed

• Clearer model descriptions. Every Claude model in the picker now has a refreshed one-line description, so the trade-off between speed, capability, and cost reads at a glance.

Fixed

• Links in chat are clickable. A plain web address an agent writes in chat now opens in your browser with a single click, no copy and paste.
• Codex no longer interrupts during a web search. Fixed an error that could break a Codex chat while it was searching the web.
• No more flashing Claude sign-in dialog on desktop. A sign-in dialog meant for the web and mobile apps used to flash on screen and dismiss itself when you connected Claude on desktop. It no longer appears there.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows x64: auto-update from 0.4.7 or later pulls you forward automatically. The MSI is still not OS-code-signed (SignPath integration still pending), so Windows SmartScreen may warn on a fresh install.
• Windows ARM64: there is no auto-update path from an emulated x64 install. Download the ARM64 MSI manually, uninstall the x64 build first, then install the ARM64 one.

Known limitations

• Fable 5 costs more credits. It uses roughly twice the credits of Opus 4.8 per run. The model picker flags this, but it is worth knowing before you switch every agent over.

Houston 0.4.19

More control over how your routines use chats, update notes in your own language, and a reliability pass so Houston opens cleanly and the right things take you to the right place.

Added

• Choose how each routine uses chats. When you set up a routine, pick whether every run keeps going in one ongoing chat or starts a fresh chat each time, so the history matches how you think about that task.
• Update notes in your language. The "update available" card now shows what is new in English, Spanish, or Portuguese, matching the language you use Houston in.

Changed

• Mission Control matches your boards exactly. Mission Control now uses the same cards and behavior as each agent's board, so what you see lines up everywhere.

Fixed

• Houston opens even when a workspace is slow to load. The app no longer gets stuck on a blank window at startup if one workspace takes a while to respond.
• Routine notifications take you to the right chat. Clicking a routine's "finished" notification now opens that routine's chat instead of a generic view.
• Smoother mission dragging. Dragging mission cards now shows the same cursor on every system and no longer underlines the card title mid-drag.
• Clearer "connect" prompt. The "Waiting for you to connect" note for an integration now appears at the end of the agent's message, where you are already looking.
• Problem reports catch more (behind the scenes). Some errors from Houston's screen were being dropped before they reached us. They now get reported so we can fix them faster. Nothing for you to do.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows x64: auto-update from 0.4.7 or later pulls you forward automatically. The MSI is still not OS-code-signed (SignPath integration still pending), so Windows SmartScreen may warn on a fresh install.
• Windows ARM64: there is no auto-update path from an emulated x64 install. Download the ARM64 MSI manually, uninstall the x64 build first, then install the ARM64 one.

Houston 0.4.18

New ways to organize your missions, clearer scheduling and integrations, and a reliability pass so problems actually get reported when something goes wrong.

Added

• Drag and drop missions between columns. Rearrange mission cards across your board columns by dragging them, so the board matches how you actually work.
• Friendlier schedule picker. Setting a custom schedule for a routine now uses a simple interval picker instead of raw cron-style entry.
• Search highlights in archived missions. When you search archived missions, the matching words are now highlighted so the result is easy to spot.

Changed

• Clearer integration cards. An integration that still needs you to sign in now says "Waiting for you to connect," so the next step is obvious.
• Formatted update notes. The "update available" dialog now renders these release notes with proper formatting instead of raw text.
• Clearer problem reports. When Houston hits an error, the follow-up message now says "report sent" with a short reference code. Tap Copy to grab it and send it our way so we can pin down exactly what happened.

Fixed

• Routines run on the day you picked. A weekly routine now fires on the correct day of the week instead of being off by a day.
• Mission Control matches your boards. The chats and missions shown in Mission Control now line up with what is on each agent's board.
• Routine editor stays correct when you switch agents. Opening the routine editor after switching agents no longer shows leftover settings from the previous agent.
• Archived mission search lines up. The archived-mission search now sits and behaves consistently with the rest of the board.
• Problem reports carry the real details (behind the scenes). We fixed our build pipeline so crash reports point at the actual place a problem happened instead of unreadable scrambled code, and so problems in Houston's background engine get reported too. Nothing for you to do.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows x64: auto-update from 0.4.7 or later pulls you forward automatically. The MSI is still not OS-code-signed (SignPath integration still pending), so Windows SmartScreen may warn on a fresh install.
• Windows ARM64: there is no auto-update path from an emulated x64 install. Download the ARM64 MSI manually, uninstall the x64 build first, then install the ARM64 one.

Houston 0.4.17

Long conversations now keep themselves tidy on their own, your language choice sticks per workspace, and you can search inside archived missions, plus a few routine and connection fixes.

Added

• Conversations that keep going. When a chat gets close to full, Houston now compacts it automatically so the conversation can continue without losing the thread or making you start over.
• Your language stays put. The interface language you pick is now remembered for each workspace, so it stays the way you set it next time you open Houston.
• Search your archived missions. The Archived tab now has a search box, so you can quickly find a finished mission instead of scrolling.

Fixed

• One ongoing chat per routine. A scheduled routine now keeps a single conversation across its runs, instead of starting a brand new chat every time it runs.
• Custom schedules are kept. Editing a routine no longer resets a custom schedule back to Daily.
• Connected apps show the right state. A connected app like Gmail or Slack no longer gets stuck saying "Connecting." The live status and the reconnect button now work independently.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows x64: auto-update from 0.4.7 or later pulls you forward automatically. The MSI is still not OS-code-signed (SignPath integration still pending).
• Windows ARM64: there is no auto-update path from an emulated x64 install. Download the ARM64 MSI manually, uninstall the x64 build first, then install the ARM64 one.

Houston 0.4.16

Archive missions you have finished, manage your connected apps without leaving Houston, and keep an eye on how full each conversation is getting, plus a round of chat and routine reliability fixes.

Added

• Archive missions. Select one or more missions on the board and move them to the new Archived tab, so the board only shows what you are actively working on. Archived missions stay available and can be brought back anytime.
• Manage connected apps in-app. Reconnect or disconnect a connected app, like Gmail or Slack, right from Houston, without going elsewhere to fix a broken connection.
• Conversation length indicator. The chat composer now shows how full the current conversation is getting, so you know when it is a good time to start a fresh chat for the best results.

Fixed

• Switching conversations is reliable. Messages no longer duplicate or linger from the previous chat when you move between conversations, including conversations started by a routine.
• Two routines at the same time both run. If you schedule two routines for the same moment, both now run instead of only one.
• The model picker shows live status. Sign-in status updates the instant a provider connects or disconnects, so the picker always reflects what is actually connected.
• Claude Opus 4.8 on its highest effort setting. Fixed an issue that could stop Opus 4.8 from running at its extra-high effort level.
• Recover from a stuck chat. If a chat got stuck mid-activity, Houston now repairs it on its own instead of staying frozen.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows x64: auto-update from 0.4.7 or later pulls you forward automatically. The MSI is still not OS-code-signed (SignPath integration still pending).
• Windows ARM64: there is no auto-update path from an emulated x64 install. Download the ARM64 MSI manually, uninstall the x64 build first, then install the ARM64 one.

Houston 0.4.15

Adds Claude's newest model and a more flexible OpenAI sign-in flow, plus a round of fixes for renaming agents and recovering from corrupt history.

Added

• Claude Opus 4.8. Anthropic's newest Opus is available wherever you pick a Claude model, for both new chats and existing agents.
• OpenAI device-code sign-in. Signing in to OpenAI now works in setups where Houston cannot open a browser for you, such as a remote engine, the web app, or a headless server. Houston shows a short code and a URL; open the URL on any device, paste the code, and you are signed in.

Fixed

• Renaming an agent or workspace works for every edge case. Renames that only change capitalization (for example, "Sales" to "sales") or that keep the same name go through cleanly now. The agent's folder location also stays in sync after a rename, so reveal-in-File-Manager and similar actions point at the right folder.
• Recover from a corrupt routine history. If a saved routine run history gets damaged, Houston now repairs it on its own so the routine can run again instead of getting stuck.
• Translated session-finished notification. The desktop notification that fires when an agent finishes its work now appears in your language.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows x64: auto-update from 0.4.7 or later pulls you forward automatically. The MSI is still not OS-code-signed (SignPath integration still pending).
• Windows ARM64: there is no auto-update path from an emulated x64 install. Download the ARM64 MSI manually, uninstall the x64 build first, then install the ARM64 one.

Houston 0.4.14

Builds on v0.4.13 with a round of reliability fixes (Windows sign-in, setup errors, starting missions, and a friendlier crash screen) plus a cleaner macOS installer.

Fixed

• Start a new mission from Mission Control. The "new mission" action works again from Mission Control.
• Automatic recovery from a corrupted conversation history. If a saved conversation gets corrupted, Houston now repairs it on its own instead of getting stuck when you resume.
• Clearer, translated setup errors. When Claude Code can't finish setting up, Houston now shows the real reason, in your language, instead of a generic failure.
• Readable startup errors instead of a white screen. If Houston hits a problem while launching, you now get a clear, copyable error with a way to report it, rather than a blank window. On older macOS versions you get an update prompt instead.
• Agent-created notes show up in Files. Markdown that an agent writes now appears in the Files tab right away.
• Notification clicks open the mission (Windows and Linux). Clicking a finished-session notification now takes you straight to that mission.
• No leftover background process after quitting (Windows and Linux). Closing Houston now fully shuts down its background engine.

Changed

• Cleaner macOS installer. The install window has a clear drag-to-Applications layout, and if you open Houston directly from the disk image it offers to move itself into Applications for you.
• Keyboard shortcuts live in Settings. The Commands list moved into Settings as a "Keyboard shortcuts" section.
• Reasoning effort is per agent. You can set reasoning effort per agent now, and it only shows for models that support it.
• Tidier sign-in dialog. The provider sign-in link is tucked behind a "reveal" toggle so the dialog stays clean.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows x64: auto-update from 0.4.7 or later pulls you forward automatically. The MSI is still not OS-code-signed (SignPath integration still pending).
• Windows ARM64: there is no auto-update path from an emulated x64 install. Download the ARM64 MSI manually, uninstall the x64 build first, then install the ARM64 one.

Houston 0.4.13

Hotfix on top of v0.4.12. Resolves a Composio sign-in issue from the previous release and adds a way to sign out of Composio from inside Houston.

Fixed

• One-time Composio sign-out on first launch. A previous release left some users signed into Composio in a state that couldn't be recovered from inside the app. v0.4.13 signs everyone out of Composio once on first launch so the next sign-in starts clean. Your connected apps stay connected on the Composio side, you just need to sign in again the next time you open the Integrations tab.

Changed

• "Sign out" button in Integrations. Top-right of the Integrations view when you're signed in. Confirms before signing out so it isn't a one-misclick mistake.
• Searchable category filter in Integrations. The category dropdown on "Browse all apps" is now sorted alphabetically, capped at a sensible height instead of stretching down the screen, and lets you type to filter (typing "mark" jumps to "Marketing").

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows x64: auto-update from 0.4.7 or later pulls you forward automatically. The MSI is still not OS-code-signed (SignPath integration still pending).
• Windows ARM64: there is no auto-update path from an emulated x64 install. Download the ARM64 MSI manually, uninstall the x64 build first, then install the ARM64 one.

Houston 0.4.12

First release on top of v0.4.11. Codex on ChatGPT accounts is unblocked, the OpenAI model picker no longer offers a model that doesn't exist, and a run of mission-control + onboarding + platform improvements ships in one cut.

Fixed

• Codex sign-in, sign-out, and mission runs work again on ChatGPT accounts. codex login and codex logout were aborting with unknown variant 'xhigh' whenever a newer codex CLI had written model_reasoning_effort = "xhigh" into ~/.codex/config.toml. Houston now passes -c model_reasoning_effort=high to both, matching what the runner already did.
• Phantom gpt-5.5-codex model removed from the OpenAI picker. OpenAI never shipped that SKU; every call came back with a 400 "not supported when using Codex with a ChatGPT account" regardless of plan. Catalog now lists only gpt-5.5. Stored configs that point at the retired model fall back to gpt-5.5 automatically the next time the chat panel reads them.
• Mission panel stays open through the model picker and closes cleanly on Esc when the new-mission field is empty.
• Connect-phone copy is cross-platform in English, Spanish, and Portuguese, instead of using macOS-only wording on Windows.
• Plus icon and left-aligned model picker restored after the previous chat header refactor.
• Workspace index self-heals when corrupted, and writes use a per-call tmp filename so two concurrent writes can't clobber each other.
• Sessions no longer vanish or stick on "running" after an unexpected provider exit.
• Agent files browser renders correctly on Windows including empty folders.

Changed

• Per-agent provider + model storage. Workspace-level defaults are retired; each agent now carries its own provider and model in its .houston/config/config.json. Existing workspaces migrate idempotently on next launch.
• Create-with-AI for new agents. Type a sentence describing what you want the agent to do; the provider you pick generates a CLAUDE.md, suggests Composio integrations, and proposes a starting routine.
• Paste images straight from the OS clipboard into chat.
• Global keyboard shortcuts across mission control, board, and chat, plus a command-palette-style sidebar Commands menu and a chat attach menu.
• "Finder" renamed to "File Manager" throughout the UI so Windows users see terminology that matches their OS.
• Gemini provider paused in the picker while the engine adapter is hardened. Existing Gemini agents keep working; new agents can't select it until it returns.
• Settings sidebar shows the app version at the bottom so bug reports include it without users having to dig.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows x64: auto-update from 0.4.7 or later pulls you forward automatically. The MSI is still not OS-code-signed (SignPath integration still pending).
• Windows ARM64: there is no auto-update path from an emulated x64 install. Download the ARM64 MSI manually, uninstall the x64 build first, then install the ARM64 one.

Known limitations

• Windows MSI not yet OS-code-signed. SignPath integration still pending.
• "Continue with Microsoft" sign-in for Houston itself remains hidden while we reconcile the Supabase Azure provider with the production callback URL. Google sign-in is unaffected.
• Mobile companion is still beta. Expect occasional reconnect hiccups.

Houston 0.4.11

A hotfix release. 0.4.10 shipped a race condition that could leave your workspace index on disk with a few stray bytes at the end after two rapid actions on the same workspace, for example two provider changes back to back. Once that happened, Houston would fail to load workspaces and agents on next launch and drop you on the "start tutorial" screen with nothing reachable.

What this fixes

• Self-healing workspace index. On launch, Houston now detects the corrupted shape, recovers the valid part, and rewrites the file cleanly. If you have been stuck on the start tutorial after a recent update, just install 0.4.11 and relaunch. Your workspaces and agents come back automatically. No manual file editing, no data loss.
• No more race on the workspace index. The writer now uses a unique temp filename per save, so two concurrent updates can never stomp each other again.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.

If you are currently stuck on the start tutorial

Auto-update should still reach you because the update notification lives outside the broken workspace surface. If for any reason it does not, download the 0.4.11 DMG from the GitHub release page and drag it over /Applications/Houston.app. The new app reads your existing ~/.houston/ and repairs the file on first launch.

Known limitations

• If you manually edited ~/.houston/workspaces/workspaces.json to recover from this bug earlier, you do not need to do anything. Your file is already clean.
• The self-healing only catches the specific "valid JSON + trailing bytes" pattern this bug produced. A file that is mangled in a different way will still surface a real error so you can see it.

Houston 0.4.10

The big consolidation release. Native Windows ARM64 alongside the existing Mac universal and Windows x64 builds. A reimagined first-run that walks new users from cold app to a working assistant, then turns the resulting work into a reusable Skill and a daily Routine. Per-workspace WORKSPACE.md + USER.md context that gets injected into every agent's system prompt. A Linear-style keyboard layer with a command palette and arrow nav on boards. The ability to share an agent with a friend as a single .houstonagent file. Plus a stack of Windows reliability fixes from a month of live testing.

What's new

• Native Windows ARM64. First ever ARM64 MSI. Composio runs as a real aarch64 binary (built from the houston-windows-support fork with a bun:ffi fallback for Bun's missing TinyCC on Windows-on-ARM). PortableGit ships per arch and auto-extracts on first launch, and CLAUDE_CODE_GIT_BASH_PATH is exported globally so every Claude Code spawn site inherits it. No more "install Git for Windows" hand-holding; no more emulated x64.

• Share an agent with a friend. A new "Share with a friend" action on any agent row packages CLAUDE.md plus chosen skills, routines, and learnings into a single .houstonagent file. Optional Houston-driven anonymize pass redacts emails, paths, handles, and phones with side-by-side diffs you accept per item. Recipient opens New Agent, picks "From a friend", confirms name, color, and provider, and the agent shows up in the sidebar.

• Create a skill from scratch. A third tab on Add Skill next to Skills.sh and GitHub. Title auto-slugs, body is markdown, collision-aware.

• Per-workspace context: WORKSPACE.md + USER.md. Two markdown files at the root of each workspace feed every agent's system prompt at session start. WORKSPACE.md is about the company, product, and customers (shared by every agent in the workspace). USER.md is about you: role, goals, preferences. Edit them from new Settings tabs that mirror the agent Instructions UX exactly. Agents can write them through the normal file-write tool when you tell them something new about yourself.

• Reimagined first-run: Try, Skill, Routine, Summary. The day-plan mission turns into a reusable Skill (plan-my-working-day) plus a daily Routine that calls that Skill at the time you pick, ending on a celebration screen that lists your new Skill and Routine. Same chat surface throughout, Skip always available, defaults to your connected Google or Outlook toolkits.

• Keyboard shortcuts + command palette. Linear-style power-user layer:

  • ⌘K command palette (agents, recent missions, top-level actions)
  • ⌘N new mission scoped to the current board
  • ⌘M Mission Control
  • ⌘[ / ⌘] cycle agents
  • ↑↓←→ move across the kanban grid
  • ? cheatsheet

  Selected mission card picks up the comet-glow treatment and hydrates straight into the chat composer ready to type.

• Two-model picker + integration-style provider cards. Anthropic: Opus + Sonnet (default Sonnet). OpenAI: Codex via your ChatGPT plan, default model reverted from gpt-5.5 to the stable lineup after a rejection-card surfaced "GPT-5.5 not in your plan" cleanly instead of failing silently. Provider cards now match the Composio integration style: one click to sign in, one click to sign out.

• Per-mission model selection persists. Picking a different model on one mission no longer cascades to siblings. Each mission stores its own provider + model in activity.json.

• Settings: 3 sections + Report bug. Workspace (name, language, timezone, theme), AI provider, Connect phone, plus a Report bug section that submits straight into Linear and surfaces the returned issue ID inline.

• Privacy + Terms pages on the website. Linked from every footer.

Reliability

A month of live Windows testing surfaced a long list of issues:

• Windows ARM64 claude-code now auto-installs instead of failing with "unknown host platform".
• Windows first-launch crash loop caused by PortableGit extraction edge cases is gone; the engine extracts the SFX natively before any claude.exe spawn.
• Windows OAuth sign-in now completes via implicit-flow callback in addition to PKCE (PKCE alone failed silently on some Microsoft tenants).
• Single-instance Tauri so the OAuth deep-link reaches the already-open Houston window instead of spawning a fresh process that doesn't know about the original sign-in.
• Persistent session storage with surfaced storage errors and a "Send logs" CTA when things go sideways.
• Skip-tutorial escape hatch for accounts that get stuck mid-flow.
• "A local tool failed to start" on user-clicked Stop is gone. The engine no longer misreports a graceful cancellation as a crash.
• Stale codex global config no longer kills sessions. The engine passes -c model_reasoning_effort=medium explicitly so an xhigh left in your ~/.codex/config.toml from a newer CLI cannot blow up the session.
• Bundled codex bumped to 0.130.0, which is the version gpt-5.5 requires server-side.
• No more 5 to 10 second "frozen" feeling at session start. The parser now anchors on codex's turn.started event instead of waiting for the first reasoning token.
• Claude 401 errors stop showing "Something got stuck" and surface as a real re-auth card.
• Auth gate updates on sign-in. Session is now written into the React Query cache after setSession so the gate flips immediately.
• Onboarding dead-zones killed: missing providers.json keys, the Activity empty-state flash, the tutorial waiting-state hang.
• Windows clickable file links + Markdown preview in the New files section.
• Light / dark mode: hardcoded color values across the UI replaced with semantic CSS tokens so dark mode actually looks dark.
• PostHog instrumentation for growth (user_signed_in, user_signed_out), activation (onboarding_started, onboarding_completed), and engagement (session_completed), all scoped per workspace and agent.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows x64: auto-update from 0.4.7 or 0.4.9 should pull you forward automatically. The MSI is still not OS-code-signed (SignPath integration is on the roadmap). One SmartScreen "Run anyway" prompt on first install only.
• Windows ARM64: this is the first ARM64 build. There is no auto-update path from an emulated x64 install. Download the ARM64 MSI manually, uninstall the x64 build first, then install the ARM64 one. Future ARM64 to ARM64 updates apply automatically.

If you are on 0.4.5 or later (Mac / Windows x64)

Houston should detect 0.4.10, download it, install it, and relaunch into the new version on its own.

If you are on 0.4.2 or earlier (Mac)

Auto-update will not pull you forward from those builds. Quit Houston, drag a fresh DMG over /Applications/Houston.app, and reopen. After that one manual step you are back on the healthy updater path and future versions apply automatically.

Known limitations

• Windows MSI not yet OS-code-signed. SignPath integration still pending. One SmartScreen "Run anyway" click on first install only, both x64 and ARM64.
• "Continue with Microsoft" sign-in for Houston itself remains hidden while we reconcile the Supabase Azure provider with the production callback URL. Google sign-in is unaffected. The Microsoft 365 + Outlook tutorial path is unaffected.
• Mobile companion is still beta. Expect occasional reconnect hiccups. Keep your Mac awake while using it.

Houston 0.4.9

A polish pass on the first-run tutorial. The "Plan my next working day" mission now runs noticeably faster, works for Microsoft accounts, and finishes on a full-bleed report screen instead of an easy-to-miss card in the chat split. No engine changes, no breaking changes. Drop-in on top of 0.4.8.

What's new

• Provider choice on the tutorial mission. The agent now asks "Google or Microsoft?" and binds Gmail + Calendar or Outlook + Outlook Calendar accordingly at runtime. Microsoft 365 users are no longer pushed through a Google-only flow.
• Tutorial mission is faster. Instead of dispatching two subagents and waiting on each, the agent reads your calendar and your 20 most recent emails inline as parallel tool calls in the same turn. Reply drafts are created the same way: three drafts, one turn, three parallel tool calls. End-to-end the AHA reaches "drafts ready in your inbox" in roughly half the time of 0.4.8.
• No more Sheets step + no follow-up email back to you. The 4-tab Google Sheet and the post-mission email are gone. The agent now produces a single structured chat reply with bold "Don't miss / Your day / Worth replying / Your commitments" sections, in your language, and leaves the actual reply drafts sitting in your inbox where you'd reply from anyway.
• Full-bleed success screen. The agent's final report renders on its own screen with a sticky-footer "Enter Houston" CTA, instead of a card inside the chat split. Much harder to miss the next-step button.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows: auto-update from 0.4.7 or 0.4.8 should pull you forward automatically. The MSI is still not OS-code-signed (SignPath is still pending). One SmartScreen "Run anyway" prompt on first install only.

If you are on 0.4.5 or later

Houston should detect 0.4.9, download it, install it, and relaunch into the new version on its own.

If you are on 0.4.2 or earlier (Mac)

Auto-update will not pull you forward from those builds. Quit Houston, drag a fresh DMG over /Applications/Houston.app, and reopen. After that one manual step you are back on the healthy updater path and future versions apply automatically.

Known limitations

• Windows ARM64 not yet shipped. v0.4.9 is x64-only on Windows.
• Windows MSI not yet OS-code-signed. SignPath integration still pending. One SmartScreen "Run anyway" click on first install only.
• "Continue with Microsoft" sign-in for Houston itself remains hidden while we reconcile the Supabase Azure provider with the production callback URL. Note: this is the Houston app login, separate from the Microsoft 365 + Outlook tutorial path that this release adds. Google sign-in is unaffected.
• Mobile companion is still beta. Expect occasional reconnect hiccups. Keep your Mac awake while using it.

Houston 0.4.8

The "tutorial demo that actually lands" release. The first-run flow now opens with a single, repeatable mission, "Plan my next working day", that exercises Gmail, Calendar, and Sheets in parallel and finishes with a 4-tab Google Sheet plus a follow-up email. No more morning-only chips that felt broken at evening demos. The success card has been re-skinned with the same running-glow comet border the Activity Board uses, so the next-step CTA actually gets clicked. Plus a real fix for the silent zero-result you'd see when installing certain skill repos.

What's new

• Single, parallel-tools tutorial mission. The 6 personal-assistant chips ("morning brief", "meeting prep", "weekly recap", etc.) are gone. The new flow uses one "Plan my next working day" mission that runs two subagents in parallel against Gmail + Calendar + Sheets, cross-references their results for the "I didn't notice this" moment, and produces a real 4-tab Google Sheet (Schedule / Replies needed / Commitments / Don't miss) plus a sent follow-up email. Re-runnable any day, every day, regardless of what time you open the app.
• Comet-glow success card. The end-of-tutorial card now wears the same animated card-running-glow border the Activity Board uses for in-flight missions. People were missing the success card and the "start the tour" CTA on it; the comet treatment fixes that.
• Repos with a single root SKILL.md install correctly. Skills like refero_skill (underscore in the repo name) or any repo whose canonical skill is at the root would silently install zero skills with no error. Now they install. Skill authors' declared name: in frontmatter is preserved as the slug; we no longer overwrite it with a slugified version of the repo name.
• Skill install failures are visible. Errors from the install path no longer get swallowed by an internal tracing::warn! and turned into "installed 0 skills". They propagate up and render in the dialog as real error copy, so when something fails you can actually see why.
• Download for Windows on the website. The marketing site has a "Download for Windows" button next to the Mac download in the hero and pricing card. Clicking opens a beta-disclosure modal with a button to join the Windows waitlist (same form as the Mac one). Doesn't change the app itself, but it's how we'll size Windows demand in PostHog.

Reliability

• No silent failures, codified. Houston's CLAUDE.md now bans the patterns that produced the "installed 0 skills" bug, namely catching errors with tracing::warn! and returning Ok, plus let _ = x.await for ops that can fail. New code in the engine has to surface failures through ? to the route layer. We're going to be auditing the rest of the codebase against this rule over the next releases.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows: auto-update from 0.4.7 should pull you forward automatically. The MSI is still not OS-code-signed (SignPath is still pending). One SmartScreen "Run anyway" prompt on first install only.

If you are on 0.4.5 or later

Houston should detect 0.4.8, download it, install it, and relaunch into the new version on its own.

If you are on 0.4.2 or earlier (Mac)

Auto-update will not pull you forward from those builds. Quit Houston, drag a fresh DMG over /Applications/Houston.app, and reopen. After that one manual step you are back on the healthy updater path and future versions apply automatically.

Known limitations

• Windows ARM64 not yet shipped. v0.4.8 is x64-only on Windows.
• Windows MSI not yet OS-code-signed. SignPath integration still pending. One SmartScreen "Run anyway" click on first install only.
• "Continue with Microsoft" sign-in remains hidden while we reconcile the Supabase Azure provider with the production callback URL. Google sign-in is unaffected.
• Mobile companion is still beta. Expect occasional reconnect hiccups. Keep your Mac awake while using it.

Houston 0.4.7

The reliability follow-up to 0.4.6. Everything that landed last week, the tutorial, the settings rail, the integrations sign-in card, gets hardened against real-world flakiness here. Connect cards no longer get stuck. The skills marketplace stops freezing on a single search. The guided tour now actually walks you through the workspace tabs instead of just pointing at them. Plus a few coming-soon provider cards in the brain picker so you can see what's next.

What's new

• Connect cards in chat actually finish. The previous spinner could get stuck for minutes when Composio's consumer-connections endpoint lagged or returned a different slug shape than the one we expected. Three things fixed this together: the engine now pushes a ComposioConnectionAdded event the moment it sees the connection on the backend (no more polling roulette), every visible "not connected" card now refreshes itself on focus and on a 10s heartbeat (catches connections you make outside the chat, in another agent, or on the CLI), and 6 seconds after you click Connect the spinner becomes an "I've connected" button you can click to force a refresh. No more "stuck until restart".
• Skills marketplace is responsive again. Searching skills.sh would freeze the dialog any time the network call took longer than a keystroke, because every search shared one mutex. Each query now runs independently. Opening the dialog uses a separate cached "popular" feed so your first keystroke is never blocked behind a network call. Errors are now typed and explained in plain English (rate-limited / offline / repo private / repo missing / malformed skill) instead of "something went wrong" plus a console log.
• Guided tour walks you through the tabs. Each tour step now actually switches to the matching tab (Activity, Routines, Files, Job Description, Mission Control) before spotlighting it, so you see what the tab looks like, not just the trigger that opens it. Two new steps: spotlight the "add agent" button, then open the New Agent dialog and spotlight the store grid so you can see what installing an agent looks like. New "Back" button so you can step backward, and a closing "Now go build something amazing" outro.
• Integrations provider is now anonymized in the UI. Onboarding, the integrations tab, and the inline chat sign-in card now say "integrations provider" instead of "Composio" (Composio is mentioned only as the default provider). Backend, agent system prompts, internal identifiers, and the bundled CLI are unchanged, this is purely user-facing copy. We expect to add more integration providers in the future.
• More providers visible in the brain picker. Gemini, SubQ, DeepSeek, and MiniMax show up as "Coming soon" cards alongside OpenAI and Anthropic in the onboarding brain mission and the settings provider picker. Brand SVGs for the real ones, a monogram for SubQ. Cards are non-clickable until those providers are wired up.

Reliability

• Search dialog seeds without blocking your keystrokes. Opening the skills dialog used to fire a fake search("ai") that competed with whatever you typed for the same lock. Removed. Popular feed is now cached for 24h and renders instantly.
• AbortController everywhere in skills. Each keystroke aborts the prior in-flight request, so racing search calls don't paint stale results.
• Stable error tags on the wire. The engine now exposes typed error.kind strings (e.g. repo_private, github_rate_limited, skill_malformed) so the UI can map specific failures to actionable copy without parsing message strings.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows: auto-update from 0.4.6 should pull you forward automatically. The MSI is still not OS-code-signed (SignPath ships in a follow-up); first-time installers will see one SmartScreen "Run anyway" prompt.

If you are on 0.4.5 or 0.4.6

Houston should detect 0.4.7, download it, install it, and relaunch into the new version on its own.

If you are on 0.4.2 or earlier (Mac)

Auto-update will not pull you forward from those builds. Quit Houston, drag a fresh DMG over /Applications/Houston.app, and reopen. After that one manual step you are back on the healthy updater path and future versions apply automatically.

Known limitations

• Windows ARM64 not yet shipped. v0.4.7 is x64-only on Windows.
• Windows MSI not yet OS-code-signed. SignPath Foundation integration is still pending. One SmartScreen "Run anyway" click on first install only.
• "Continue with Microsoft" sign-in remains hidden while we reconcile the Supabase Azure provider with the production callback URL. Google sign-in is unaffected.
• Mobile companion is still beta. Expect occasional reconnect hiccups. Keep your Mac awake while using it.

Houston 0.4.6

The "first run that feels like a game" release. Houston now opens with a guided tutorial that gets a brand-new user from cold app launch to a real, working assistant in under a minute, no form-filling, no docs. This is also the first cut that ships natively on Windows, the first one with a real Settings rail, the first one with a Sign-out button on connected providers, and the first one with the Outbound cold-email agent in the store.

What's new

• Video-game-style first-run tutorial. Replaces the old form-based onboarding with a guided flow: name your assistant, sign in to your AI provider, sign in to Composio so the agent has tools, then run a real chat against it. The "Try it" mission isn't a demo, it kicks off a real mission via the same chat surface used everywhere else, so the conversation lives on the Activity board and survives once the tutorial is done. Predefined chips (morning brief, meeting prep, weekly recap) make the first message zero-effort.
• Inline Composio sign-in card in chat. When the agent tries to call a tool that needs Composio and you're not signed in, it posts a connect card directly in the conversation. One click, you're authenticated, the agent retries. Mirrors the existing per-toolkit connect-card pattern so users never get bounced out to Settings to recover.
• Workspace UI tour. After graduation, a coachmark tour walks you through the agents area, tabs (Activity, Routines, Files, Job Description), Mission Control, Integrations, and the new "Guide me" pill in the tab bar. Replayable any time from that pill.
• Settings has a proper sidebar. The long-scrolling settings page is gone. Account, Workspace, Provider, Appearance, Language, Timezone, Connect phone, and Danger zone each live in their own section with one-click left-rail navigation.
• Sign out from Codex and Claude Code. Connected provider cards now have a "Sign out" link. It runs the official CLI logout command (claude auth logout clears the macOS Keychain entry, codex logout revokes the ChatGPT refresh token server-side), then re-surfaces the login flow inline.
• Outbound agent in the store. A new agent that runs the LinkedIn to Apollo to Instantly cold-email pipeline end-to-end. Find leads on LinkedIn, enrich via Apollo, send via Instantly, all from one chat.
• Connect phone moved into Settings. The phone-pairing QR, connection state, and the disconnect-all-phones reset now live together inside Settings under a BETA pill, instead of hanging off the global sidebar.
• Improved agent sidebar. Sidebar item rows pick up the work refactor from the new SidebarSectionNav, with cleaner spacing and active-state treatment.

Native Windows build

• Houston for Windows ships as a signed-by-minisign MSI installer alongside the Mac DMG. Same agent, same engine, same chat surface. Composio, codex, and claude-code all run as native Windows executables. No WSL, no Docker, no terminal anywhere in the install or first-run path.
• In-app auto-update on Windows. Once you're on a Houston for Windows build, future versions arrive the same way they do on Mac: a notification in the app, one click to install, automatic relaunch.
• Windows first-run hardening. Sign-in opens the OAuth URL through cmd /C start instead of macOS open (Windows users could not get past the sign-in screen before). The engine spawns with CREATE_NEW_PROCESS_GROUP so the parent app's console events stop killing the subprocess every couple of minutes (STATUS_CONTROL_C_EXIT). The bundled CLI resolver knows the MSI install layout, so claude-code and composio are found instead of silently skipped. Composio's first-time login writes your API key to the Windows Credential Manager via the same FFI-against-native-APIs pattern Houston uses on the macOS Keychain. No plaintext on disk.
• Project moved to the gethouston org. github.com/ja-818/houston is now github.com/gethouston/houston. The old URL still redirects, so existing clones, forks, and updater endpoints keep working.

Reliability

• Sessions no longer die at the first message when an unrelated copy of Claude Code lives elsewhere on your computer. Houston now talks directly to the version it manages itself, which is always kept in sync with what Houston was tested against. Previously, an older claude install on the same machine could intercept the call and reject newer flags, surfacing as a vague "app error."
• Auth errors are actually visible. Sign-in surfaces post-callback OAuth failures inline with human-readable messages (the common AADSTS codes plus Supabase identity-collision errors are mapped to actionable sentences) instead of silently logging.
• Updater uses the same minisign signing key for both OSes. The signing key controls who can ship updates to existing Houston installs, regardless of platform.

Before you upgrade

• Mac: quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.
• Windows (first install): the MSI is not yet OS-code-signed (SignPath integration ships in a follow-up). You'll see a SmartScreen warning the first time. Click "More info" then "Run anyway." Once installed, future auto-updates run from a trusted parent process and don't show the warning.

If you are on 0.4.4 or 0.4.5 (Mac)

Houston should detect 0.4.6, download it, install it, and relaunch into the new version on its own.

If you are on 0.4.2 or earlier (Mac)

Auto-update will not pull you forward from those builds. Quit Houston, drag a fresh DMG over /Applications/Houston.app, and reopen. After that one manual step you are back on the healthy updater path and future versions apply automatically.

Known limitations

• Windows ARM64 not yet shipped. v0.4.6 is x64-only on Windows. ARM64 (Surface Pro X, etc.) is on the roadmap once we have demand to justify the second build pipeline.
• Windows MSI not yet OS-code-signed. SignPath Foundation integration is the next Windows-side PR. Until then, expect one SmartScreen "Run anyway" click on first install only.
• "Continue with Microsoft" sign-in is hidden for this release while we reconcile the Supabase Azure provider with the production callback URL. Google sign-in is unaffected.
• Mobile companion is still beta. Expect occasional reconnect hiccups. Keep your Mac awake while using it.

Houston 0.4.5

A reliability and polish release. Chat now recovers gracefully when a CLI provider misbehaves, attachments render in the conversation, the sidebar stays usable for agents with long names, and the language across Skills surfaces is consistent with the rest of the product.

What's new

• Skills surfaces now speak "Actions". The cards you used to see in Skills are the same ones you see elsewhere in the app, with consistent copy. Browsing, adding, and inspecting community Actions feels like one product, not two.
• A "+" button in the Running column. You no longer have to leave the board to start a mission. The Running column has its own footer button that opens New Mission with the composer already focused.
• Composer focus on new missions. Open New Mission and your cursor is already in the message box. Type, send, done.
• Attachments render inline. When you or an agent share a file in chat, you see it as a proper attachment card in the conversation, not a raw filename.
• Bug reports go straight to Linear. Send a bug from inside Houston and it lands as a ticket the team triages, with the diagnostics already attached.
• Sidebar controls stay visible for long agent names. The follow-up chip and the 3-dot menu no longer get pushed off-screen when an agent has a long name.

Reliability

• Recovery from malformed Claude requests. A specific class of provider error used to leave the chat stuck. Houston now surfaces a recovery card so you can retry without restarting the agent.
• Tool runtime errors show up as recovery cards. When a tool fails mid-mission, you see a card explaining what happened and offering a retry instead of a silent stall.
• Anthropic reconnect only fires when auth is confirmed. Houston no longer flashes reconnect prompts during the brief window before sign-in finishes resolving.
• Community skill search is hardened. Edge cases that used to throw or return empty when searching the community library are handled.

Before you upgrade

• Quit Houston before installing. macOS does not always replace a running app cleanly. If in doubt, remove /Applications/Houston.app and drag the new copy in.

If you are on 0.4.3 or 0.4.4

Houston should detect 0.4.5, download it, install it, and relaunch into the new version on its own.

If you are on 0.4.2 or earlier

Auto-update will not pull you forward from those builds. Quit Houston, drag a fresh DMG over /Applications/Houston.app, and reopen. After that one manual step you are back on the healthy updater path and future versions apply automatically.

Known limitations

• Windows support is still in progress. Engine compiles for Windows, signing and CI are next.
• Mobile companion is still beta. Expect occasional reconnect hiccups. Keep your Mac awake while using it.

Houston 0.4.4

A maintenance release that validates the in-app updater path now that the relaunch fix from 0.4.3 is out in the wild. Also includes a small sidebar polish so you can tell at a glance whether the update applied.

What's new

• Sidebar follow-up chip is bigger and lives where it belongs. The "needs you" badge in the agent sidebar now sits flush right (the same slot the menu used to own), and it's a wider oval so the count is easier to read across the room. The 3-dot menu still appears in the same spot on hover, so renaming and deleting an agent works the way it did.

Why this release

The 0.4.2 build shipped with a broken in-app updater. The fix landed in 0.4.3, but anyone still on 0.4.2 cannot pick it up over the air. This 0.4.4 release exists to confirm the fixed updater is healthy on 0.4.3 installs and that future updates apply cleanly without manual reinstalls.

If you are on 0.4.3

Houston should detect 0.4.4, download it, install it, and relaunch into the new version on its own. You will see the sidebar chip change shape, that confirms the auto-update worked end-to-end.

If you are on 0.4.2 or earlier

Auto-update will not pull you forward. Quit Houston, drag a fresh DMG over /Applications/Houston.app, and reopen. After that one manual step you are on the healthy updater path and 0.4.5 onward will apply automatically.

Before you upgrade

• Quit Houston before installing. macOS does not always replace a running app. If in doubt: rm -rf /Applications/Houston.app and drag the new copy in.

Known limitations

• Windows support is still in progress (engine compiles for Windows, signing and CI are next).
• Mobile companion is still beta. Expect occasional reconnect hiccups. Keep your Mac awake while using it.

Houston 0.4.3

A reliability and polish release. Search across all your missions, see live agent activity right in the sidebar, queue follow-ups while an agent is mid-run, attach files with confidence, and a long list of small rough edges is gone.

What's new

• Search across every mission. A search field in Mission Control and the per-agent board scans titles first, then falls back to mission text and full conversation history. Translated empty/loading states, fast even on large workspaces.
• Mission search in the agent header. The search box now lives in the agent header next to the title (not buried in a menu), one keystroke from anywhere in the per-agent board.
• See agent activity in the sidebar. Each agent in the sidebar now shows whether it's idle, working, or waiting on you, so you can tell at a glance which agent is busy without opening it.
• Queue follow-ups mid-run. Type a follow-up while an agent is still working and Houston now batches it into one combined next turn instead of dropping it. Stop/Esc still works, and you can run parallel sessions in the same folder.
• Better attachments. Attachments now stream through binary upload sessions, large selections chunk cleanly, unsupported files are rejected with a clear message, and your composer draft is preserved if something goes wrong with the upload.
• Updater that actually relaunches. When Houston downloads a new version, the install + relaunch handoff is more reliable, no more orphaned splash window or "did it update?" guessing.
• Phone connection survives restarts. "Connect phone" pairings used to drop after a Mac sleep or quit. They now persist across restarts, so you don't have to re-scan the QR code every morning.
• OpenAI reconnect, no flapping. When Codex needed to reconnect to OpenAI, the UI sometimes flickered between connected and disconnected for a few seconds. The state machine is steady now.
• Codex recovers when its session file is gone. If the underlying Codex resume rollout file disappears or got corrupted, Houston now falls back gracefully instead of getting stuck on "resuming...".
• Codex config stays your config. Houston no longer rewrites your ~/.codex/config.toml behind your back. If you customize it, it stays customized.
• Composio integrations refresh on connect. Connecting a new Composio integration (Gmail, Slack, etc.) used to need a manual reload to show up in the agent. The list now refreshes the moment the OAuth window closes.
• Provider model picker scrolls. The model menu in agent creation no longer overflows the screen for providers with long model lists. It scrolls inside the menu.
• Cleaner Action picker. Categories that were empty or showed only one Action no longer get their own tab, so the tab bar in the Action picker doesn't waste space.
• Mission titles match your provider. The auto-summarized mission title is now generated by the same provider your agent is using, so styling and tone stay consistent.
• Bug reports go through cleanly. The Report bug button now sends through a native webhook path that doesn't get blocked by Tauri's network sandbox, so reports actually reach us.
• Tighter agent prompts. Houston's bundled agent prompts had drifted into long, repeating instructions. They've been pruned, agents stay on-character with less context overhead and quicker first responses.
• Skill search degrades gracefully. When skills.sh hits its search rate limit, Houston now backs off and shows cached results instead of erroring.

Behind the scenes

• Public changelog. gethouston.ai/changelog/ now shows release notes for every version, pulled live from GitHub. Linked from the top nav and footer.

Before you upgrade

• Quit Houston before installing. macOS doesn't always replace a running app. If in doubt: rm -rf /Applications/Houston.app and drag the new copy in.

Known limitations

• Windows support is still in progress (engine compiles for Windows, signing + CI are next).
• Mobile companion is still beta. Reconnects are now durable, but expect occasional hiccups. Keep your Mac awake while using it.

Houston 0.4.2

A polish release on top of 0.4.1. Cleaner chats, agents that actually use what they've learned, and a Legal agent rewritten so a non-technical founder reads the picker and instantly knows what every Action does.

What's new

• Cleaner chat. Long stretches of process detail (file reads, tool calls) now collapse into a single tidy block by default. Tap to expand when you want the full trace, ignore otherwise. The chat reads like a conversation again.
• Agents remember what they learned. Every agent now picks up its LEARNINGS.md automatically on every turn, so corrections stick across sessions instead of being forgotten the next time you open the app.
• Better chat summary. The "what changed this turn" summary now correctly attributes files generated by your agent and stops mislabeling external edits. Generated files surface in the summary instead of getting lost.
• Houston speaks human, by default. Every bundled agent now follows strict rules for talking to you: no file paths, no JSON, no command-line jargon, no narrating its own steps. Plain sentences, only what you need to act. If your agent ever slips into technical talk, that's now a bug we can fix.
• Legal agent, fully rewritten (v0.2.1). All 40 Legal Actions renamed and rewritten in founder voice. The picker now reads "Is this name free?", "Answer a customer data request", "Calculate my Delaware franchise tax", instead of cryptic slugs. Form labels match. The actual legal advice is unchanged, only the wrapping.
• Store updates clean up after themselves. When a bundled agent ships an update that renames Actions, your existing copy migrates to the new names automatically instead of accumulating duplicate old + new entries in your picker.
• Action selection sticks. When you pick an Action, it stays selected in the composer with a clear chip showing your choice, instead of getting dismissed on the next render.

Before you upgrade

• Quit Houston before installing. macOS doesn't always replace a running app. If in doubt: rm -rf /Applications/Houston.app and drag the new copy in.
• Existing Legal agents installed from the Store will auto-migrate to v0.2.1 names on first launch. No action needed on your side.

Known limitations

• Windows support is still in progress (engine compiles for Windows, signing + CI are next).
• Mobile companion is still beta. Expect occasional reconnects. Keep your Mac awake while using it.

Houston 0.4.1

Big release. Sign in with Google, every Mac runs the same installer, the app speaks Spanish and Portuguese, your phone can manage Houston, and the major CLIs your agents use ship preinstalled. No terminal, no extra setup.

What's new

• Sign in with Google or Microsoft. Houston now has accounts. One click, native OAuth, your session lives in macOS Keychain (never in plain files). Sign out from the sidebar or Settings.
• Houston Store with curated agents. New Store tab ships with team-built agents you can install in one click, with built-in update checks. The store modal is simpler and the GitHub-import detour is gone, so creating an agent feels like adding an app.
• Actions everywhere. Skills are now called Actions and have a real UI. The empty chat shows up to 3 featured Actions, the picker has tabs by category and a Featured row, and starting an Action takes over the composer with a clean labeled form. The conversation shows your Action as a card with all the values you filled in.
• Empty board, one tap to start. Open an agent with nothing on its board and the New Mission picker pops automatically. No more staring at an empty kanban wondering where to click.
• Codex and Composio come bundled. Both CLIs ship inside the signed Houston app, so non-technical users get them preinstalled. Claude Code installs itself on first launch (one tap, no terminal). The download is bigger but the setup is one step.
• Friendly reconnect card when a CLI logs out. When Claude or Codex drops auth in the middle of a session, you now see a labeled card with a one-click reconnect instead of cryptic terminal noise in the chat.
• Functional bug reporter. When something goes wrong the toast says "Houston, we have a problem!" and the Report bug button actually sends the report straight to our team in Slack, with your email if you're signed in.
• Español y Português. Pick your language on first launch. Everything in the app, onboarding, sidebar, Mission Control, agent tabs, settings, dialogs, errors, flips live when you switch languages. Change it any time in Settings.
• One download for every Mac. The installer is now a universal binary, the same .dmg runs on Apple Silicon and Intel. No more guessing which one to grab.
• Mobile companion (beta). Tap "Connect phone" in the sidebar, scan the QR code with the Houston iOS app, and you can manage Houston from your phone. Your Mac stays the source of truth, the phone is a window into it. Action cards render on the phone too.
• Security disclaimer on first launch. A short, plain-English note about what AI agents can do on your behalf. Read once, accept, never see it again unless the terms change.
• Cleaner card actions. Rename, Delete, and Move to done are now three matching icon buttons with instant tooltips that say exactly what each one does.
• Mission Control caught up with the per-agent board. Renaming a card works there, opening a conversation actually shows chat history, and starting a New Mission lets you pick which agent runs it without leaving the page.
• Better link previews. Sharing gethouston.ai on WhatsApp, Slack, or Twitter now shows a real social card.

Before you upgrade

• Quit Houston before installing. macOS doesn't always replace a running app. If in doubt: rm -rf /Applications/Houston.app and drag the new copy in.
• First launch is slightly slower while Claude Code installs itself in the background. You'll see a small status indicator.
• If you had accepted the disclaimer in a pre-release build, you'll be prompted once for the final copy. Takes 10 seconds.

Known limitations

• Windows support is still in progress (engine compiles for Windows, signing + CI are next).
• Mobile companion is beta. Expect occasional reconnects. Keep your Mac awake while using it.
• A handful of rare error messages from Claude or Codex still pass through in English even when you're running in Spanish or Portuguese.
• If you abort the Google sign-in popup on the wrong browser profile, just click the button again to retry.

Houston 0.4.0

First release on the new engine architecture.

What's new

• Product prompts now live in the app, engine is a standalone platform binary
• Data consolidates under ~/.houston/ (workspaces move out of ~/Documents/Houston/ automatically on first launch)
• Real-time reactivity restored — the board updates live when agents write files
• Race-free engine startup + sidecar signing + notarization verified in CI
• Drag-and-drop file attachments fixed
• Onboarding detects your Claude/Codex CLI from any standard install location
• Routines tab auto-detects your timezone (change it in Settings)

Before you upgrade

• Quit Houston first. macOS doesn't always replace a running app.
• Your workspaces will move from ~/Documents/Houston/ to ~/.houston/workspaces/ on first launch. The old location stays as a backup.
• If you have custom scripts pointing at the old path, update them.

Known limitations

• Claude and Codex CLIs must still be installed separately. Bundled-CLI install is coming in a later release.

See the assets to download and install this version.

See the assets to download and install this version.

See the assets to download and install this version.

Houston v0.2.0 — First Release

The framework for building AI agent desktop apps. Rebranded from Keel & Deck.

What's in the box

11 React packages (@houston-ai/*): core, chat, board, layout, workspace, skills, connections, events, routines, review, memory

8 Rust crates (houston-*): sessions, db, tauri, channels, events, scheduler, memory, skills

Houston App: Unified desktop app with experience system — host multiple AI agent experiences in one app

Experience System: Three tiers — JSON-only manifests, custom React bundles, custom Rust capabilities

Quick Start

npx create-houston-experience my-agent
cd my-agent
pnpm install
pnpm tauri dev

Links

• npm: @houston-ai
• crates.io: search houston-