SeaWork2.0.19 - 2026-08-07
New
Pinned agents grouped by project - The sidebar now shows your pinned agents grouped under their projects instead of in one flat list, and pinned agents also appear within their project group. Project groups stay sorted by your most recent activity.
Fixes
Excel and file previews open reliably - Files opened in the built-in preview (including Excel workbooks) no longer fail to load due to a cross-origin restriction.
Timeline loads more smoothly - Duplicate history requests are now coalesced, and the timeline no longer flickers or skips entries when you scroll quickly.
Clearer agent error messages - When an agent fails to start, the error shown now includes the real failure reason (for example a serialization error) instead of a generic “[object Object]”.
Slash commands list stays usable - Typing / in the command box no longer shows a “request failed” error for agents that don’t support command listing.
2.0.18 - 2026-08-03
Fixes
Windows installer closes cleanly when upgrading - Installing an update over an existing copy no longer gets stuck on a window that won’t close, and a failed upgrade restores the previous install.
2.0.17 - 2026-08-03
New
SVG chat previews - SVG files in chat now open with a visual preview instead of raw markup.
Pick up where another agent left off - Use /from-agent to restore context from another agent so you can continue the same work without re-explaining.
Sign in with email - Log in with an email address and a one-time code, alongside existing sign-in options.
Sign out on desktop - Sign out from the desktop app when you need to switch accounts or clear the current session.
Improvements
More reliable image messages - Incoming images are handled more stably, with less duplicate processing when the same image arrives more than once.
Fixes
Deleted conversations stay deleted - Removing a conversation no longer brings it back in the list.
Drafts no longer overwrite what you just typed - Restoring a saved draft no longer replaces fresh input you have already started writing.
Web conversation timeline stays out of the way - The timeline no longer covers chat content on web.
Verification code field focuses automatically - The one-time code input is ready to type into as soon as the sign-in screen appears.
Conversation timeline loads further back - The timeline now keeps loading older messages even past the locally kept history, so you can scroll further into long conversations.
Timeline waits for history before scrolling - Opening a long conversation no longer scrolls before the older messages have actually loaded.
2.0.16 - 2026-07-28
Fixes
Replace conflicting local skills and personas when installing from a group - If a shared asset conflicts with something you already have, you can replace the local copy instead of getting stuck.
Daemon starts more reliably with many agents - Cold-start memory use is lower, so the app is less likely to run out of memory when loading a large agent library.
Windows desktop packaging includes Starry CLI again - The Windows installer correctly bundles the Starry CLI binary after extract.
2.0.15 - 2026-07-23
New
Share personas and skills with a group - Create or join a group with an invite code, publish your agent personas and skills so members can use them, and install shared ones into your own setup in a tap. Manage members, roles, and invite codes from a new Groups section in Settings.
Improvements
Richer file previews - CSV and TSV files now display as tables, you can select a whole file at once, and code blocks and quotes in markdown previews have a hover copy button for grabbing just that content.
Your permission mode is remembered - When you change an agent’s permission mode mid-conversation, that choice carries over to your next session with the same agent, on both mobile and web.
Fixes
Permission mode changes take effect for Grok - Changing the permission mode for an agent that can’t switch mid-session, such as Grok, now reloads the session to apply the new mode instead of quietly keeping the old one.
2.0.14 - 2026-07-14
New
Easier agent mentions in nested workspaces - Agent mentions now search across multiple directory segments, making it easier to find the intended agent in deeper project paths.
More ways to learn and get started - Added guides for Codex CLI, Claude Code, remote control, and current model pricing on the Seawork website.
Improvements
Clearer product and remote-control pages - Expanded localized product pages and FAQs with clearer explanations of remote coding and supported tools.
2.0.13 - 2026-07-13
New
Grok Build as a first-class agent - Install and run Grok Build from Seawork alongside your other coding agents, with guided install and the same conversation workflow.
Improvements
Quieter conversation timeline - System-instruction messages are folded by default so the timeline stays focused on real turns, and timeline hover targets are easier to hit.
Multi-image editing - Image edit can take multiple source images; generate and edit jobs continue in the background so the UI stays responsive.
Faster, leaner sharing - Publishing a shareable page is quicker and more reliable.
Clearer file preview navigation - When viewing a file, a back control returns you to the conversation.
Fixes
Cleaner Codex replies - Citation markup is stripped from live messages and history so conversations read more cleanly.
More reliable agent install - Grok install verifies binary integrity before use, and install/cancel edge cases are harder to get stuck in.
2.0.12 - 2026-07-10
New
Share a conversation as a web page - Publish any session as a shareable HTML page with a slash command, optionally passing an instruction to scope what gets shared. Manage every page you’ve shared from a new section in Settings. Shared pages include a footer link to download Seawork, and the share link now appears right in the conversation timeline.
Bind multiple WeCom bots - Connect more than one Enterprise WeChat (WeCom) bot at the same time, each with its own agent and settings.
Improvements
Continue after a request error - When a request fails, a new “Continue” action lets you pick up the conversation from where it stopped instead of starting over.
Faster sharing - Publishing a shareable page is noticeably quicker.
Fixes
Shared pages keep your secrets out - Secrets are redacted before a page is published, and share links can be protected so private conversations stay private.
The app starts even with a bad saved schedule - A corrupted schedule file no longer prevents the daemon from starting; the bad entry is skipped instead of blocking everything.
WeCom settings survive upgrades - An older single-bot WeCom setup is now carried over automatically after updating, instead of being rejected on startup.
2.0.11 - 2026-07-08
New
Mention agents with # - Type # in the message box to mention another active agent working in the same directory. Your agent hands the request to the mentioned agent, waits for it to finish, and brings the result back into your conversation.
Improvements
Signed Windows installer - The Windows installer is now code-signed, so installing and updating no longer triggers unknown-publisher warnings.
Tidier image replies - Images generated by agents now appear as a compact thumbnail strip at the bottom of the reply instead of full-width blocks in the middle of the text.
Fixes
Safer Windows updates - Fixed a rare failure where updating on Windows could leave the app missing entirely. Updates no longer install silently while the app is closing, and if a previous version’s uninstaller is broken, the update stops with clear guidance instead of removing the existing installation.
File previews back in the new layout - Clicking a file in the new sidebar layout opens a visible preview again, including from the new-conversation screen.
Project menu works in the browser - The project actions menu in the sidebar now opens reliably on web.
Removing projects is reliable - Removing a project now consistently removes all of its workspaces, including archived ones.
Drafts survive retries - Retrying a failed message no longer overwrites what you were typing.
Cleaner conversation history - Older prompts no longer disappear when browsing long Codex conversations, and internal system messages no longer show up in the history.
2.0.10 - 2026-07-06
New
Migrate Codex conversations to Claude Code - A new “Migrate to Claude” action converts a Codex conversation into a Claude Code session with its history preserved, so you can continue the same work with a different agent.
Improvements
New sidebar layout by default - The updated desktop layout is now the default, with conversation context menus (rename, pin, archive, reload, delete), a dedicated section for pinned conversations, and project-level actions like rename, archive all, and reveal in Finder. If you previously chose the classic layout, your choice is kept.
Sub-agents stay out of the way - Agents spawned by tools no longer clutter the sidebar, command palette, or workspace tabs. They remain available in the Agents tab behind a “Show sub-agents” toggle.
Fixes
Smoother Windows updates - Updating on Windows now shows installer progress instead of appearing stuck, no longer asks for an install location again, closes running components cleanly, and reliably relaunches the app when done.
Fewer crashes for heavy users - Increased available memory to prevent crashes when managing hundreds of conversations or multiple integrations.
“Archive all” includes pinned conversations - Archiving all conversations in a project now also archives pinned ones.
Select All scoped to file previews - Pressing Cmd+A / Ctrl+A in a Markdown preview now selects only the file content instead of the whole page.
2.0.9 - 2026-07-02
New
SeaArchive integration - A new built-in plugin provides read-only access to your archived conversations through agent tools.
External plugin webview support - Plugins can now render custom web interfaces through an overlay panel.
Interactive agent questions - Claude agents can now ask questions and receive structured answers directly in the conversation.
Improvements
More reliable terminal reuse - The terminal panel validates that terminals are still alive before reusing them, preventing “not found” errors.
Better agent watchdog behavior - The watchdog monitors stderr activity to avoid incorrectly stopping active turns, and can self-heal from temporary fences.
Smoother reconnection - Messages sent while disconnected now wait for reconnection to complete before being delivered.
Windows installer diagnostics - The installer now logs detailed diagnostics when installation fails, and can recover from a broken previous uninstaller.
Fixes
Model picker works reliably - The model catalog no longer fails to load due to missing authentication tokens.
Accurate multi-byte text handling - Fixes crashes and corrupted output when processing multi-byte UTF-8 characters in agent streams.
Phone login completes on desktop - Desktop callback flow for phone-based login now completes successfully.
New agents get a default workspace - Creating a new agent no longer leaves the workspace field empty.
Canvas downloads work - Files generated in Canvas can now be downloaded without browser restrictions.
Lower memory use in desktop dev tools - Eliminates a console flood that could exhaust main process memory.
2.0.8 - 2026-07-01
Improvements
More capable built-in integrations - Plugin-based tools can now run background operations and bypass proxies when needed, and the Starry integration supports signing in and one-step setup.
Fixes
Fewer memory-related crashes - Fixes a connection resource leak that could slowly grow memory use during remote access, reducing crashes over long sessions.
Question cards always render - Questions from Claude now show as interactive cards even when they arrive without an identifier.
2.0.7 - 2026-06-30
New
More built-in integrations - Built-in tools like Canvas, SeaFlow, and SAC now run on a unified plugin host, and a new Starry integration lets agents install and list Starry skills.
Improvements
Smoother desktop sign-in - Desktop login is more reliable: the sign-in callback survives deep links, expired website sessions are cleared automatically, and login completes through a more robust flow.
New agent drafts are preserved - A draft you start for a new agent is kept when you navigate into settings and back.
Fixes
Lower memory use and fewer crashes - Long Codex turns and image-heavy conversations no longer accumulate unbounded memory, reducing background and window crashes.
Confirmation before deleting automation tasks - Deleting an automation task now asks for confirmation first.
WeCom reliability - Incoming files are kept within a bounded directory, and pending approvals are cleared when you run /stop or /new.
Reliable downloads across drives - Fixes a failure that could occur when saved files were moved across different drives.
2.0.6 - 2026-06-29
Fixes
Image attachments send reliably to Codex - Desktop image attachments are kept available until a message finishes sending, so Codex receives the images you selected instead of a text-only message.
Missing image files fail clearly - If an attached image cannot be read, SeaWork now surfaces the send failure instead of silently sending the message without the image.
2.0.5 - 2026-06-29
New
Built-in browser automation - SeaWork now includes a Browser panel for opening pages, picking page elements, and giving agents browser context they can act on directly.
Slash commands for built-in tools - You can now use /canvas, /seaflow, and /sac from the composer to open the right tool or steer the agent toward the matching capability.
New desktop workspace layout - Desktop users can opt into a project-grouped conversation layout with a clearer conversation list, workspace tools, and project defaults.
Improvements
Codex conversations recover more gracefully - Long or interrupted Codex turns now preserve more partial output and avoid duplicate replies when a stalled turn later finishes cleanly.
Browser automation is safer across windows - Browser actions stay tied to the correct page, reducing cases where an action could target the wrong browser view.
Canvas and SeaFlow usage is easier to understand - SeaWork now tracks panel usage more clearly so the experience can be improved around the tools people actually use.
Fixes
Daemon memory use is better bounded - Long-running desktop sessions are less likely to accumulate unbounded memory from internal tracking state.
Attachments stay with slash-command messages - Files and images attached to a message are preserved even when the text starts with a slash command that opens or steers a tool.
2.0.4 - 2026-06-26
New
Plugins page - Built-in and external tools are now organized as plugins with their own sidebar entry, including per-category toggles so you can enable or disable groups of built-in tools.
Infinite canvas - A new canvas (web and desktop) for laying out and editing images, with box-select editing and canvas-aware skills.
SeaFlow panel - A side panel that drives the SeaArt flow canvas directly inside SeaWork, with automatic login and canvas import/export.
Image generation and editing with gpt-image-2 - The image plugin can now generate and edit images using gpt-image-2.
File preview on web - Preview PDF, Office documents, media, and archive files directly in the web app.
OpenAI status indicator for Codex - Codex agents now show OpenAI service status, making it easier to tell upstream outages from local problems.
Login entry screens - New SeaWork login and desktop browser login entry points.
Improvements
Image attachments stay consistent across devices - Sent image attachments and their metadata are preserved reliably across reloads, restarts, and when the same conversation is opened on another device.
Fixes
Codex turns recover from false stalls - Force-failed Codex turns self-heal when a matching response arrives, and plan-approval prompts now surface reliably from the final message.
Imported files with a byte-order mark parse correctly - Importing Codex/Claude config and canvas files saved with a BOM (common on Windows) no longer fails silently.
Windows file paths are handled correctly - The file explorer normalizes Windows drive-letter file paths properly.
Image editing is hardened - The image edit tool blocks SSRF and DNS-rebinding attempts and handles its results more reliably.
SeaFlow panel is more stable - The panel stays anchored while resizing the window, opens when invoked, and no longer creates a new canvas on every start.
2.0.3 - 2026-06-23
Improvements
Image-heavy Codex conversations are more efficient - Images sent to Codex are optimized before upload, reducing payload size and making multimodal turns less likely to hit large-request limits.
Long Codex runs recover more accurately - SeaWork now waits longer than Codex’s own stream-idle window before treating a quiet turn as stalled, reducing false timeout recovery during slow upstream streams.
Codex stall diagnostics are easier to capture - Desktop settings now include a verbose logging toggle for Codex sessions, with the saved preference applied consistently when agents start.
Image attachments survive reloads and restarts - User message images and their metadata are preserved more reliably across conversation reloads, daemon restarts, and fallback paths.
Fixes
Draft text is saved more reliably - Message drafts are flushed when switching conversations or closing input views, reducing cases where the latest typed text is lost.
Creating files and folders from the explorer is smoother - New file and folder actions now default to the workspace root when appropriate and avoid duplicate or stuck creation states.
Desktop startup handles stale local state more safely - Stale daemon lock files are cleaned up more carefully, so desktop startup is less likely to be blocked by a previous crashed process.
2.0.2 - 2026-06-22
Improvements
More reliable agent recovery during long runs - Codex sessions now better recognize upstream activity and recover when progress continues after a watchdog timeout, reducing cases where a conversation appears stuck.
Claude tools are available more consistently - Claude sessions now load SeaWork’s built-in tools by default and preserve local tool access when latency proxy settings are enabled.
Voice mode works better behind proxies - Speech transcription and playback now honor configured proxy settings, with safer handling for proxy credentials in logs.
Desktop releases install more reliably - Desktop and Android release builds now use the expected native packages and download mirrors more consistently across Mac, Windows, Linux, and Android.
Fixes
Codex permissions and modes stay in sync - Full-access sessions now clear pending permission prompts correctly, and selected modes persist more reliably across agent starts.
Headless daemon authentication is more reliable - Headless daemon starts can use saved login credentials for gateway requests without requiring a separate environment setup.
Desktop update failures are less disruptive - Background update download checksum failures no longer crash the desktop app.
Large conversation streams use bounded memory - Very long conversation streams are capped more carefully to reduce memory pressure during extended sessions.
2.0.1 - 2026-06-18
Improvements
Consistent download naming - Desktop download files now use the SeaWork name consistently across Mac, Windows, and Linux.
2.0.0 - 2026-06-18
What’s new
Skill Hub brings agent skills into the app - Browse a curated skill catalog, search and preview skill docs, test-run skills, and install or update them directly from SeaWork. Install links from the web now open straight into the app.
Browser extension for sending pages to agents - A new companion extension can send the current page URL and selected text to a local SeaWork agent, with guided setup from the desktop app.
Richer media generation in conversations - Agents can now create images, videos, audio, and 3D assets directly in chat, with generated results rendered inline instead of returned as local file paths.
Clearer usage limits - The quota view now shows multimodal balance and usage windows with progress bars. When a window limit is reached, SeaWork shows a clearer reset countdown in the conversation.
Search conversations and files - Cmd/Ctrl+F now works inside conversations and workspace files, with highlighted matches and keyboard navigation through results.
Older history loads as you scroll - Long conversations can now load earlier messages on demand, making extended agent runs easier to review.
Improvements
More dependable Codex recovery during long work - Codex sessions recover more reliably from interruptions, stream reconnects, and quiet turns that finish shortly after the watchdog threshold.
Safer workspace links and Skill Hub installs - File links and Skill Hub installs are handled more carefully, including paths, external links, and install states.
Quieter desktop crash reporting - More benign helper-process, log-write, and environment-related failures are treated as warnings instead of app crashes.
Fixes
Search keeps focus while navigating matches - Conversation search no longer loses focus on repeated Enter, and Escape closes search more consistently.
Generated media workflows complete more reliably - SeaArt generation now handles async task completion and status aliases more accurately.
1.0.38 - 2026-06-15
What’s new
Smoother, more tactile app interactions - SeaWork now feels more responsive across common workspace actions, with more polished motion, feedback, and visual layering while navigating and working with agents.
Claude support is ready for the newer agent runtime - Desktop builds now package the newer Claude agent runtime correctly, so Claude sessions can start reliably after upgrading.
Improvements
More dependable Codex recovery during long work - Codex sessions now stay alive more reliably during permission waits, interruptions, long-running operations, and large local outputs.
Better workspace file discovery - File mentions and workspace search handle symlinked folders more safely, including links that point outside the current workspace.
Safer scheduled agent defaults - Automation-created agents now wait for preferences to load and validate saved mode or reasoning settings before starting.
Fixes
Plan mode stays read-only for Codex - Codex agents now preserve the expected read-only sandbox while planning, reducing the chance of accidental edits before implementation is approved.
Desktop crash reports are quieter and more accurate - SeaWork filters more environmental helper-process crashes and shutdown noise, while preserving useful diagnostics for real failures.
Doctor network checks better identify certificate problems - Doctor reports untrusted-root TLS failures more clearly, especially on older Windows systems.
1.0.37 - 2026-06-13
What’s new
Edit plans before an agent starts implementing - When an agent proposes a plan, you can revise the plan text before approving it, making it easier to steer work without starting a new conversation.
Image attachments are easier to keep and revise - User-uploaded images are preserved more reliably across refreshes and device handoffs, including when editing earlier messages.
Improvements
Clearer request-failure guidance - Common account, network, and upstream request failures now show more specific hints in the conversation, making it easier to decide whether to retry, check connectivity, or follow up with support.
More reliable Codex session recovery - Codex sessions now handle goal commands, interruptions, retries, and local state more consistently, reducing cases where a conversation appears stuck or loses runtime context.
Update prompts are safer on desktop - After an update replaces the running app bundle, SeaWork now asks for a relaunch so the desktop app and daemon stay on the same version.
Fixes
Bug reports include better failure context - Reports for failed turns now retain more useful diagnostics while filtering noisy log lines, so support has the right details without extra back-and-forth.
Normal replies stay separate from error cards - Assistant messages are no longer mistaken for system errors, while real request failures still show the dedicated error card.
Website copy is clearer - The public website no longer describes SeaWork as open source.
1.0.36 - 2026-06-12
What’s new
Reusable prompt injection for every provider - You can now set a global prompt for all agents and add project-specific prompts from the project menu, so new agents can consistently inherit your preferred instructions.
File edits stand out in conversations - Tool calls that edit or create files are now visually highlighted, including inside collapsed tool groups, making code changes easier to spot while reviewing an agent run.
Fixes
Codex local state is preserved more broadly - Managed Codex homes now keep future runtime files and locked Windows state files instead of deleting unknown root entries during startup.
1.0.35 - 2026-06-12
Fixes
Desktop startup recovers from stale local daemons - SeaWork now clears an old daemon process that is still holding the local port before starting a replacement, so the desktop app is less likely to get stuck after a restart.
Codex local state is preserved more safely - Managed Codex homes now keep their runtime database, plugins, and root state more reliably while cleaning up stale legacy aliases.
1.0.34 - 2026-06-11
What’s new
Richer Enterprise WeChat media workflows - Enterprise WeChat messages can now carry inbound images into SeaWork conversations, and agents can return richer media responses without exposing local file paths.
Built-in file copying from desktop views - File previews and file explorer rows now make it easier to copy file contents, including empty text files, and Windows reveal actions handle paths more reliably.
Automation helpers for development workflows - New guarded automation helpers make it easier to run issue work, pull request checks, and review loops from SeaWork-managed agents.
Improvements
More reliable Codex retries and recovery - Codex turns keep retry ownership more accurately, recover better when retry output arrives in unusual shapes, and avoid stale turn state after interruptions.
More dependable model lists - Provider model lists now fail closed with clearer empty states instead of blocking the rest of the app when a catalog cannot be loaded.
Better issue tracker host handling - Git forge hosts are normalized more consistently, including SSH aliases, so linked pull requests and issue tracker settings resolve more predictably.
Fixes
Normal assistant replies no longer render as errors - Regular assistant messages are preserved as chat content, while legacy system-error messages still show the dedicated error card.
Failed turns are easier to diagnose - Turn failures now stay visible in bug report logs and are represented with dedicated error items instead of being lost in generic assistant output.
Quieter Windows crash reporting - Benign Windows GPU helper restarts are no longer reported as SeaWork crashes.
1.0.33 - 2026-06-10
What’s new
Delete sessions directly from desktop views - You can now remove sessions from desktop rows and workspace tabs, with a confirmation step before destructive actions.
Faster WeCom bot setup from the workspace - WeCom bot setup is easier to reach and can now bind to an existing agent more smoothly.
Improvements
Better handling for file links outside the workspace - Absolute file links now preview more reliably, even when they point outside the current project.
More reliable model and session recovery - Saved sessions handle stale model information more safely, and DeepSeek sessions recover more gracefully from active-turn conflicts.
Smoother navigation and startup flow - Initial prompts, session navigation, and linked pull request selection now behave more predictably in day-to-day use.
Fixes
Bug report fallback works again - When the richer report flow is unavailable, SeaWork now falls back correctly instead of leaving reporting stuck.
Clearer failed-send feedback - Failed message sends now surface clearer diagnostics, making it easier to understand what went wrong.
Settings text is more complete across languages - The settings library section is now localized more consistently.
1.0.32 - 2026-06-09
What’s new
Quote assistant replies when composing - You can now select text from an assistant message and carry that quote into the composer, making follow-up questions and corrections easier to anchor to the exact response you mean.
Improvements
Clearer quoted context in conversations - Quoted assistant text is highlighted and threaded into the draft flow more consistently, so it is easier to see what context will be sent with your next message.
1.0.31 - 2026-06-09
What’s new
Enterprise WeChat bot integration - You can now connect SeaWork to Enterprise WeChat, bind the bot through a QR flow, and receive streamed smart bot replies from your configured workspaces.
Shared workspace memory for Codex and DeepSeek - Codex and DeepSeek sessions can now receive the relevant Claude project memory index for the same workspace, making cross-agent handoffs keep more of the context you already built up.
Improvements
More reliable Codex sessions on Windows and long threads - Codex now avoids Windows-only symlink permission failures, recovers from stuck compaction, and keeps its first-run managed home state stable.
Safer model selection - New sessions fall back to provider defaults when a saved model is no longer available, so stale model choices no longer block agent creation.
Clearer Enterprise WeChat setup - First-user binding, allowlist controls, loading states, and workspace selection now behave more predictably while configuring the bot.
Fixes
Conversations stop showing stale busy states - Completed or failed turns now update the conversation state more reliably instead of leaving the UI looking like the agent is still working.
Account and billing errors are easier to understand - Insufficient-balance errors from newer gateways now show the right user-facing hint.
Bug reports are more resilient - Drafts and screenshots are preserved more reliably while moving through the app, so reports are less likely to lose context before submission.
Desktop release skills are bundled reliably - Desktop builds include the bundled skills expected by agent sessions after install.
1.0.30 - 2026-06-07
What’s new
Built-in image generation skill for Codex — Codex agents now ship with a bundled image-generation skill, available on both Windows and macOS, so you can generate images directly from a Codex session without extra setup.
Manual context compaction for Codex — You can now run /compact in a Codex conversation to compact the context on demand, the same way Claude already supports, keeping long sessions responsive.
Fixes
DeepSeek runs connect to the right gateway — DeepSeek sessions no longer fail with a 404 by sending requests to the wrong endpoint; model traffic is now reliably routed to the correct gateway.
More reliable agent connections — SeaAgent sessions now resolve their gateway credentials correctly, fixing connection errors at startup.
Smoother in-chat image viewing — The image viewer has clearer close controls and previews are sized more consistently in the conversation.
1.0.29 - 2026-06-06
What’s new
Inline image previews in chat — Images your agent generates now render directly in the conversation, and SeaWork routes image generation through its own image gateway for more reliable results.
Improvements
Long Codex conversations stay healthy — Codex sessions now auto-compact as a thread grows, so very long conversations no longer get stuck repeatedly failing. Turn timeouts pause during compaction, and a Codex session reconnects automatically if its backend exits mid-run.
File mentions in new-session drafts — You can now reference files with @-mentions while drafting a brand-new session, not just inside an active conversation.
Conversation history stays visible offline — A Codex agent’s past timeline is restored and shown even when you’re temporarily disconnected, instead of appearing empty.
Fixes
Smaller UI polish — The agent stop button icon is now properly centered, and the desktop app no longer logs benign GPU-process shutdown messages as errors.
1.0.28 - 2026-06-05
What’s new
Run stats at the top of every conversation — A new stats bar shows token usage, cache hit rate, and balance for the current run, so you can keep an eye on consumption without leaving the conversation.
Improvements
Easier model picking on new sessions — The new-session draft now highlights your current model and restores the ability to favorite or unfavorite a model right from the hero card, making it quicker to find and reuse the models you use most.
Steadier run statistics — Token counts now stay stable while a turn is in progress, cache hit rate and total token math are corrected, and balance figures are formatted safely even when the source value is not a clean number.
Broader model catalog support — SeaWork now accepts additional OpenAI-style model catalogs, so more provider model lists load correctly.
Fixes
Codex permission requests recover and behave consistently — Pending Codex permission prompts that were orphaned mid-run are now recovered, and responding to a permission request is idempotent, so a run no longer gets stuck waiting on a prompt or double-handling the same approval.
One-click desktop bug reports work again — Reporting an issue from the desktop app once more opens the one-click flow and attaches diagnostic logs, instead of dropping you into a blank manual form.
Quieter crash reporting — Crash intake now filters out reports from unrelated helper processes that were mis-attributed through the OS crash handler, keeping the report queue focused on real SeaWork crashes.
Accurate failed-request ids — A failed request is no longer mislabeled with a previous request’s id, so the id shown on a failed turn now points at the request that actually failed.
Desktop no longer crashes on update download failures — A background update download that returns a 404 or other HTTP error is now treated as a recoverable update problem instead of crashing the app, so a temporarily unavailable download cannot take the whole desktop app down.
1.0.27 - 2026-06-03
What’s new
DeepSeek CodeWhale provider support — SeaWork can now run DeepSeek CodeWhale agents, including one-click desktop setup, provider model discovery, session resume, and clearer provider availability states.
Better failed-turn reporting — Failed agent turns now show the upstream request id when available and offer clearer retry and report actions, making support cases easier to diagnose.
Improvements
More reliable provider and model loading — Model lists now load through a shared snapshot path with better stale-data handling, clearer catalog failure messages, and faster refresh behavior after workspace changes.
Smoother message input handling — Message inputs resize more reliably, preserve clipboard and selection behavior on web, keep queued prompts out of the visible timeline, and append queued prompts consistently when a run is ready.
Faster workspace status refreshes — Workspace git state and pull request metadata are cached and refreshed after relevant git mutations, reducing repeated background work while keeping sidebar status current.
Fixes
Clearer desktop startup failures — Desktop startup now surfaces daemon port conflicts directly and reduces crash-report noise from known startup, native runtime, and install-path failures.
Crash reports ignore unrelated helper processes — Crash intake now filters out reports from unrelated helper binaries that were mis-attributed through Crashpad, keeping the issue queue focused on SeaWork crashes.
Enterprise TLS failures are easier to act on — Gateway failures caused by enterprise TLS or untrusted root certificates now produce clearer remediation guidance instead of generic provider errors.
Claude and Codex runs recover from more edge cases — SeaWork now handles early stream EOFs, mid-turn stream failures, cancelled metadata generation, and Claude SDK exit races without wedging the active turn.
DeepSeek streaming is more resilient — DeepSeek sessions now handle provider environment overrides, malformed SSE frames, runtime routing, start-turn races, and final text fidelity more reliably.
Symphony workflow state is safer — Symphony now preserves live sessions more carefully, syncs labels more safely, keeps GitHub issue closing under GitHub’s native flow, and avoids cleaning worktrees before linked pull requests are done.
1.0.26 - 2026-06-01
What’s new
Shared provider environment settings — Provider-specific environment variables can now be configured from SeaWork and reused consistently across Claude and Codex runs, making it easier to keep local agent setup aligned.
Improvements
More reliable session startup and resume — SeaWork now gives busy daemons more time to return workspace and agent directories, clears stale create-flow state after reconnects, and finds Claude resume transcripts more reliably across project directories.
Codex resume keeps shared MCP tools attached — Resumed Codex sessions now keep the shared Library MCP configuration available, so tools and skills remain visible after returning to an existing session.
Lower CI overhead — Project CI now uses the self-hosted M2 runner for more checks and avoids publishing npm packages from release-candidate tags.
Fixes
Fewer noisy crash reports — Known non-actionable disk-full, missing Swift compiler, and native runtime quarantine failures are filtered more accurately so real reports are easier to spot.
Native speech startup is safer on quarantined platforms — SeaWork skips preloading native speech runtime paths when a platform has been quarantined, reducing startup failures from known-bad native packages.
1.0.25 - 2026-05-28
What’s new
Scheduled agent runs — A new schedule management screen lets you set up recurring agent runs, see upcoming and completed runs, trigger a run on demand, and review schedules with localized time labels that match your timezone.
Shared MCP servers and skills across Claude and Codex — MCP server and skill configurations are now centrally managed and shared between Claude and Codex, so a single setup applies across both agents instead of having to configure each one separately.
Improvements
More reliable Codex turn handling — Cancelling or interrupting a Codex turn now tears down cleanly without leaving the chat stuck, and very fast incoming turns are no longer dropped during startup. Approval prompts from Codex are bridged correctly so turns no longer stall waiting for a response that never arrives.
Available models reflect your account — The list of usable models is now loaded from your signed-in account configuration, so the picker matches what your gateway actually allows.
Fixes
Agents recover after watchdog failures — When the agent watchdog has to interrupt a stuck run, the next message now starts a fresh turn instead of getting blocked behind the failed one.
Clearer messages when the disk is full — Running out of disk space now surfaces a clear “no space left” message instead of looping through crash reports.
Quieter crash reports when the port is in use — When the daemon can’t start because another SeaWork is already running on the port, you now see a clean port-in-use message instead of an automatic crash report.
Mobile chat input no longer overflows — The workspace message input on mobile stays inside the screen instead of spilling past the edge on narrower devices.
Windows terminals survive process launch failures — Transient Windows process-creation failures in the terminal no longer take the daemon down with them.
Schedules evaluate in your local time — Cron expressions for scheduled runs now run on your local clock instead of UTC, so a schedule set for “9am” fires at 9am where you are.
1.0.23 - 2026-05-23
What’s new
Faster workspace resume — Workspaces with many saved agents now load much more quickly, so returning to SeaWork is less likely to hit resume timeouts.
Live Codex streaming in chat — Codex responses now appear progressively in the chat instead of waiting for the full turn to finish.
Gateway typing speed test script — A local diagnostic script now makes it easier to compare direct gateway streaming with SeaWork agent response timing.
Improvements
More reliable Codex setup — Codex provider setup now keeps the SeaWork gateway configuration aligned more consistently and includes better fallback behavior for model lists and skill diagnostics.
Cleaner Codex skill handling — Desktop skill diagnostics now respect custom Codex homes and report fixable skill issues more accurately.
Better development isolation — Local desktop and worktree development now keep Portless names and agent auth isolated more reliably across checkouts.
Fixes
Loopback desktop connections work more reliably — Desktop and web clients can now connect through loopback origins more flexibly, reducing false origin rejections in local setups.
Fewer noisy crash reports — Known non-actionable crash noise is filtered more aggressively so real issues stand out.
Workspace metadata stays responsive — Workspace descriptor building is more parallel and skips archived agents where possible, reducing background stalls in large workspaces.
1.0.22 - 2026-05-20
What’s new
Desktop file attachments — Desktop chats now accept arbitrary dropped or picked files, preserve image previews more reliably, and support larger inline message images.
New File from the file explorer — The file explorer context menu now includes a New File action with safer optimistic updates and clearer create-state handling.
Richer command autocomplete — Slash commands and agent shortcuts are easier to find with fuzzy matching and better draft preservation while switching contexts.
Sign out from desktop — Desktop users can now sign out and clear the local auth file directly from the app.
Improvements
More reliable agent startup checks — Codex and Claude availability checks now honor provider environment settings more consistently across the daemon, desktop app, and doctor diagnostics.
Better response progress visibility — Agent turns now surface more detailed status and latency information, including model, tool, and token timing when debug data is available.
Smoother desktop layout behavior — Sidebar, workspace tab, and desktop list layouts now handle narrow windows and overflow states more predictably.
Fixes
Agent turns recover after startup failures — Failed turn starts now clean up pending state correctly so later messages do not get stuck behind a fake active run.
Chat scrolling is less jumpy — Manual scrolling now detaches auto-follow more reliably, preventing the chat from snapping back to the bottom while reading earlier output.
Message images survive reconnects and refreshes — Locally attached images are rehydrated more reliably so sent messages keep their previews.
Desktop crash reports are quieter — Additional known external-plugin and native shutdown noise is filtered so actionable crash reports are easier to identify.
1.0.21 - 2026-05-15
What’s new
SeaWork naming throughout the app — User-visible product naming now consistently uses SeaWork across the desktop app, CLI paths, install checks, and packaged desktop builds.
Multiple desktop windows — The desktop app now supports opening additional windows from the File menu, making it easier to monitor or control more than one workspace at a time.
Raw markdown view in the file pane — Markdown files can now be toggled between rendered preview and raw source view directly inside the file pane.
Built-in workflow shortcuts — New workflow actions help start feature work, pick issues, and follow pull request review activity from inside SeaWork.
Create folders from the file explorer — The file explorer context menu now includes a New Folder action with safer empty-name handling, root-level support, and cross-workspace safeguards.
Improvements
Workspace file filtering — File search and file panels now respect VS Code-style ignore patterns and workspace ignore settings more consistently, with safer refresh behavior after reconnects.
File previews stay fresh while open — The file preview pane now refreshes while focused without triggering unnecessary window-focus refetches elsewhere.
Latency debugging controls — The desktop latency debug toggle now coordinates daemon restarts more carefully, shows failures in the UI, and keeps the proxy setting aligned with the toggle state.
Updater install handoff is more reliable — Desktop update downloads now keep the install handoff unblocked so users can complete the update flow more consistently.
Fixes
Interrupted Codex turns no longer duplicate answers — Cancelling or interrupting a Codex turn now preserves the cancel marker across completion and failure paths so stale output does not get appended again.
Desktop crash reports are less noisy — Benign renderer and GPU child-process shutdowns are filtered more accurately so real crashes are easier to spot.
Windows transient rename failures are retried — Temporary Windows filesystem rename failures now retry instead of crashing the daemon.
Linux desktop release packaging is fixed — Linux desktop builds now wrap the current SeaWork executable name correctly, so release packaging completes across all desktop platforms.
1.0.20 - 2026-05-13
What’s new
Automatic crash collection — Seawork now auto-collects crashes from the desktop app, daemon, and supervisor, with a first-launch toast that explains what is gathered and a runtime opt-out. Crash reports include a stdio tail and are deduped by minidump fingerprint so the same failure isn’t reported twice.
One-click copy on code blocks — Code blocks in chat now show a copy button that copies the contents to the clipboard, with layout adjusted so the button never overlaps the code.
Per-turn latency breakdown for debug mode — When debug mode is on, each agent turn now records a latency breakdown so you can see where time was spent during a response.
Improvements
Auto-updater diagnostics for hot updates — The desktop auto-updater now exposes diagnostics that make hot-update issues easier to investigate, persists the installer path correctly across the download/install handoff, and surfaces check failures to the UI instead of failing silently.
Smarter @mention routing — Inserting a file mention now targets the agent you were last focused on in the current workspace, and gracefully falls back when no agent tab is open instead of routing to a stale global focus.
File explorer handles symlinks safely — Directory symlinks inside a workspace are now expandable folders, cyclic self/ancestor symlinks are skipped, and symlink targets outside the workspace are no longer leaked into agent context.
Fixes
File pane no longer triggers a black-screen rerender — A React #185 infinite-update loop caused by a zustand selector returning a fresh array on each render has been fixed for the file-pane and file-explorer-pane.
External editor resolves file paths correctly — Editor launches now resolve relative paths to absolute, fall back to a workspace-only view when the target file no longer exists, gracefully handle filesystem-root workspaces, and treat the win32 workspace containment check as case-insensitive.
Supervisor waits for crash handlers — The daemon supervisor now awaits the crash handler before exiting in the production path, so unhandled rejections surface a crash report instead of disappearing.
1.0.19 - 2026-05-12
Improvements
Open containing folders from desktop tabs — Desktop file tabs now include an Open Containing Folder action when the platform supports it, making it quicker to jump from Seawork back to the local file location.
Richer bug report diagnostics — Bug reports now capture more daemon runtime context and redact sensitive fields more carefully, giving maintainers better evidence without exposing private data.
Fixes
Plan updates stay consistent after reconnecting — Plan tool updates now line up more reliably between restored state and live streaming updates, avoiding duplicate or mismatched plan entries.
Windows root folders open correctly — Opening containing folders now handles Windows drive roots without producing invalid parent paths.
Internal workspace summaries use isolated temporary directories — Background summary generation now runs from a temporary directory so it does not interfere with the user’s workspace.
Desktop balance checks use the dedicated balance endpoint — The desktop app now checks account balance through the intended endpoint instead of the registration API.
1.0.18 - 2026-05-11
What’s new
Preview HTML files in Seawork — HTML files can now be opened directly inside the app, with safer local preview handling so generated pages are easier to inspect without leaving the workspace.
Workspace-wide file search — The file explorer now supports full workspace search with gitignore-aware results, better refresh behavior, and smoother navigation into matching folders.
GitHub and GitLab tracker tabs — Workspaces can now connect to GitHub or GitLab issues and pull requests, with token setup, remote detection, and a dedicated tracker view inside the workspace.
Pull request review notifications — Seawork now surfaces new PR review comments and keeps unread review summaries visible, helping you catch requested changes without manually checking the browser.
Improvements
Faster worktree archiving — Large worktrees archive much faster and cleanup failures are handled more carefully.
Resend messages from chat history — User messages can now be resent directly, with draft images cleared correctly so retries do not carry stale attachments.
Insert file references from the file pane — File selections can be inserted into chat as mentions with line anchors, making it easier to point agents at the exact context they should use.
More reliable Codex session handling — Codex resume behavior, project-scoped skill shortcuts, and full-access approval settings are preserved more predictably across sessions.
Fixes
Workspaces stay archived when deleted — Background refreshes no longer revive projects that were intentionally removed or archived.
Worktree changes stay visible — The Changes tab remains available after creating a worktree.
UI remains usable during refresh failures — Temporary authoritative revalidation failures no longer block the interface.
Local HTML preview boundaries are stricter — Preview requests are constrained to trusted workspace state and handle malformed paths more defensively.
1.0.17 - 2026-05-09
What’s new
Migrate Claude Code sessions to Codex — Existing Claude Code conversations can now be migrated into Codex from the app, including resume-list discovery, workspace targeting, and source-session archiving so the old and new sessions stay clear.
In-app community forum — Feedback and community posts are now available inside Seawork, with localized identity handling and a My Feedback view that opens entries in the app.
Live agent progress and timing — Agent turns now show the current step while work is streaming and include duration details, making long-running sessions easier to follow.
Custom worktree location per project — Projects can use their own parent directory for generated worktrees instead of relying on one global location.
Symphony workflow package — Seawork now includes a workflow orchestrator package for coordinating GitHub issue work through local Seawork agents.
Improvements
More reliable migration flows — Migration now handles Codex auth setup, deep links, retargeting, duplicate sessions, workspace scoping, and repeated migration attempts more predictably.
Better desktop context menus — Desktop chat content now supports right-click Copy actions, including image and link copy options.
Safer message sending — Failed queued sends now roll back the optimistic message and restore the draft instead of leaving the conversation in a confusing state.
Cleaner local development isolation — Worktree development now keeps daemon state, ports, and desktop bundle targets isolated from the production app and other worktrees.
Fixes
Windows agent discovery is more robust — Seawork now prefers real Windows command shims over shadowed bare-name entries and preserves PATH order when resolving agent binaries.
Bug report drafts are preserved — In-progress bug report fields now survive navigation or refreshes before submission.
File links handle line numbers more consistently — Chat file paths now strip :line suffixes reliably while still honoring explicit #L<n> anchors.
1.0.16 - 2026-05-06
What’s new
Sessions group by working directory — Agent sessions are now grouped by their working directory, with git subdirectories treated as part of their parent repo. The sidebar stays focused on projects instead of fragmenting across every nested folder you’ve ever opened.
Dictation can use a custom OpenAI-compatible STT endpoint — Speech-to-text now honors a configurable base URL and language hint, so you can point it at a self-hosted Whisper, a regional OpenAI mirror, or any other OpenAI-compatible provider, and get accurate recognition for non-English speech.
Local-only settings — Settings can now be marked local-only so they don’t sync between devices — useful for machine-specific paths, keys, or preferences.
Refreshed default agent models — The out-of-the-box model choices for new agents have been updated to the latest generation.
Improvements
Sidebar disambiguates folders that share a name — When two open projects have the same basename (e.g. two different frontend/ directories), the sidebar now shows enough path context to tell them apart instead of looking like duplicates.
Workspace removal is precise — Removing a workspace now targets the exact workspace ID instead of matching by path, so workspaces that share a directory no longer disappear together.
Malformed model override no longer breaks the daemon — A bad agents.models override in your config is silently ignored on load instead of crashing the daemon, so a stray edit can’t lock you out of the app.
1.0.15 - 2026-04-30
Fixes
Daemon no longer crashes on Windows when speech features can’t load — On systems where the bundled speech-to-text and text-to-speech native modules fail to load (missing Visual C++ runtime, antivirus quarantine, etc.), the daemon now starts successfully without those features, instead of crashing the whole app on launch. Voice features remain unavailable until the underlying issue is fixed, but the rest of Seawork works.
Behind the scenes
Website automatically updates after a stable release — Previously, the seawork.ai download page could stay pinned to the previous version after a release shipped. The release pipeline now triggers the website deploy explicitly so /download always reflects the latest stable build.
1.0.14 - 2026-04-30
What’s new
Chinese (Simplified) translations across the app — The mobile and desktop UI now ships with full zh-CN translations and follows your system locale automatically.
File-level diff inside agent turns — Each agent turn now lists the files it touched at the bottom, with insertion/deletion stats; click any file to expand its diff inline without leaving the conversation.
Quieter tool-call timeline — Consecutive tool calls of the same kind are collapsed into a single line with a running summary, so long agent turns stop drowning the screen in tool spam.
Bug reports from the splash error screen — When the local daemon fails to start, the splash window now captures the crash and lets you submit a bug report straight from the error UI — no log digging required.
Codex migration helper — A guided flow walks existing Codex CLI users through importing their config so the agent works on first launch.
Per-model thinking budget — Models that support extended thinking now expose configurable thinking effort, so you can trade latency for depth on the agents that benefit from it.
Improvements
Chinese IME no longer sends mid-composition — Pressing Enter to confirm a Chinese candidate no longer fires the message; Enter only sends after composition ends.
Smoother conversation scrolling — Several scroll-anchor races have been fixed: expanding turn cards no longer jitters, expand/collapse animations don’t fight the virtualizer, and the view stays pinned where you left it.
File explorer no longer flashes “folder not found” after rename — Renames apply cleanly without the transient error state.
Daemon works on older systems — A daemon-side git rev-parse flag that required Git 2.31+ has been replaced with a portable form that works on older systems.
Fixes
Settings text fields no longer crash the app — Adaptive text inputs rendered outside a BottomSheet are now safe.
Daemon crash diagnostics on Windows — Daemon startup crashes are now captured to a synchronous stdio log instead of disappearing into pino’s async buffer, making EDR/antivirus kills and native module failures actually diagnosable.
1.0.13 - 2026-04-28
What’s new
Self-healing environment doctor — A built-in diagnostic agent now detects missing or misconfigured pieces (codex/bun installation, registry mirrors, PATH gaps) and fixes them automatically when possible, with a clear report when it can’t.
Provider-unavailable hint on the draft screen — When a coding agent provider can’t reach its model, the draft form now surfaces a diagnostic card instead of silently failing, so you know to check your network or auth.
Improvements
Codex install on Windows works behind corporate registries — First-launch agent installation falls back through known good registry mirrors when corporate networks block the default, and no longer inherits a stale BUN_CONFIG_REGISTRY that pointed at an unreachable mirror.
Codex runs even without system Node — When system node is missing on a fresh machine, codex now uses the Node bundled inside the desktop app, so it works out of the box.
Sub-agents stay out of the Agents tab — Spawned sub-agents are no longer mixed in with top-level agents, keeping the list focused on what you started.
Generated image paths fixed on Windows — Codex-generated image paths now use the right slash flavor on Windows, so previews and attachments resolve.
Drag-and-drop file paths into the terminal — Dropping or pasting a desktop file URL into the terminal now produces a clean path on every platform.
Settings has a Back button again — A missing back button could trap you in Settings; navigation is restored.
Fixes
Workspace archive is no longer fragile — Archiving an already-deleted workspace no longer errors; the operation is now idempotent.
Bug reports include the reporter email in plain text — Easier to triage; no more digging through hidden fields.
1.0.12 - 2026-04-25
Improvements
Bug reports now route to the new intake service — Bug reports submitted from the app are sent to the updated intake endpoint.
1.0.10 - 2026-04-24
What’s new
Resume existing agent sessions — Pick up where you left off. Persisted agent sessions are now listed in the UI and CLI, scoped to the current workspace, so you can resume a prior conversation without starting a new agent.
Document attachments from file paths — Attach documents to messages by file path. The composer and picker now accept document types alongside images, and web uploads no longer silently drop files.
Open local files from chat — Click a local file link in chat and it opens in the workspace tab. Markdown files render inline in the workspace preview.
Rename agent tabs — Right-click an agent tab to rename it. Press Escape to cancel.
Codex image output in timelines — When Codex produces a native image output, it renders directly in the agent timeline.
Gemini models via Claude Code — Gemini gateway models now route through Claude Code for consistent tool-use behavior.
SeaArt multi-modal generation — The SeaArt CLI now supports video, audio, and 3D generation in addition to images.
Smarter @mention autocomplete — Filenames with spaces are escaped with a backslash instead of being wrapped in quotes, the @ prefix is preserved when selecting a match, a trailing space is added after the filename, and the menu falls back to the session’s first agent when nothing is focused.
Seawork Doctor — deeper diagnostics — Doctor now spawns actual probes of your installed Claude and Codex binaries to catch configuration issues before you hit them in an agent. Per-agent model selection, a 60-second probe timeout, and plain-English remediation text for detected root causes.
Fixes
- Composer no longer freezes when typing messages that begin with
@"..."— fixed an infinite loop in the @mention parser. - Desktop macOS directory picker lets you create new folders again.
- Desktop auto-updater manifests now point at the correct release asset paths (fixes silent update failures).
- Desktop Firebase analytics events report correctly again.
- Resumed Codex threads no longer stall waiting for a model selection.
- Android builds without
google-services.jsonskip the Firebase plugin instead of failing. - Resume session list hides internal metadata sessions and refreshes when agents change.
- staruniongame.com is supported as a known project host.
1.0.7 - 2026-04-15
What’s new
Redesigned landing page — The home screen is now a clean Codex-style workspace launcher. Open a project directly from the landing page and get straight into your agent without extra taps.
Message history navigation — Press the up/down arrow keys in the input bar to scroll back through your previous messages, just like a terminal shell.
Git diff stat in toolbar — The toolbar now shows a compact diff stat (files changed, insertions, deletions) alongside the source control button, so you can see the scope of changes at a glance.
Better default settings for new users — New installs start with the Zinc theme and queue-mode message sending — better defaults out of the box.
SeaArt CLI integration — Activate and manage SeaArt directly from the agent sidebar.
Agent deep-link authentication — Agents can now authenticate via deep link, enabling smoother OAuth flows for connected services.
Fixes
- Fixed project picker and auto-workspace selection on the landing page
- Fixed the source control button and diff stat merging into a single consistent button
- Improved permission card button visual hierarchy (Deny/Accept)
- Fixed IME composition handling and line-wrap formatting in the input bar
- Enriched daemon PATH in packaged macOS app so tools resolve correctly
- Model list now loaded dynamically from the Seawork API with robust fallback
1.0.0 - 2026-04-13
Seawork 1.0 — Your dev environment, in your pocket
We’re shipping Seawork 1.0.
Seawork is a mobile and desktop app for monitoring and controlling your AI coding agents from anywhere. Claude Code and Codex — watch them think, steer them when they go off track, and approve permission requests without running back to your desk.
Your code never leaves your machine. Seawork connects directly to your local development environment over an end-to-end encrypted channel. There’s no cloud intermediary storing your code, no agent logs sent to our servers. Just a secure pipe between your phone and your machine.
What’s in 1.0
Agent monitoring — Live output streaming, token usage, cost tracking, and tool call inspection across all running agents.
Full agent control — Send messages, approve or deny permission requests, stop agents, and branch into new tasks — all from your phone.
Multi-agent workspace — Run multiple agents in parallel, each in their own workspace. Switch between them instantly.
Remote access — Connect from anywhere over the Seawork relay. E2E encrypted with zero-knowledge architecture. Your machine, your keys.
Supported agents — Claude Code and Codex CLI. Runs the native agent harness exactly as you’d run it locally — your config, your MCP servers, your credentials intact.
Desktop app — macOS, Windows, and Linux. Includes a built-in daemon so there’s nothing extra to install.
Mobile app — iOS and Android. Available on the App Store and Google Play.
We built Seawork because we kept walking away from long-running agent sessions and coming back to messes. Now we don’t walk away. We watch, steer, and ship from anywhere.