Cowork Os

CoWork OS is a local-first personal agentic OS and everything app for getting real work done.
Code, research, design web pages, create documents, work with spreadsheets and decks, run automations, and ask your agent for changes without switching context.

Getting Started · Composer Mentions · Message Box Shortcuts · Ask Inbox · Everything Workbench · Browser Workbench · Use Cases · Release Notes 0.5.44 · Platform Updates · Documentation · Changelog · Security · Contributing

Why CoWork OS?

• Personal Agentic OS — CoWork OS is still a local-first operating system for personal AI agents: tasks, memory, skills, providers, approvals, channels, devices, and automations are governed in one workspace.
• Everything Workbench — Chat, tasks, documents, spreadsheets, presentations, web pages, PDFs, and file outputs live in one local-first workspace. Generate an artifact, open it in the app, review or edit it, and ask for follow-up changes with fewer switches into separate office apps. Learn more
• Browser Workbench / Browser V2 — Test live websites and local apps in a visible in-app browser by default. The agent uses the same right-sidebar/fullscreen browser the user can see, with CDP-backed Browser V2 automation, accessibility snapshot refs, tabs, downloads/uploads, diagnostics, screenshots, annotation, and explicit opt-in for real signed-in Chrome/Edge control. Learn more · Architecture
• Coding, web design, and knowledge work — Use the same agentic workspace for repository work, frontend/product design, research, docs, spreadsheets, presentations, web pages, PDFs, and long-running operational tasks.
• Inbox Agent — local-first email workspace with Classic and Today modes, AI triage, an Ask Inbox sidebar with live mailbox-agent steps and hybrid evidence search, @Inbox main-composer routing, manual reply/reply-all/forward, editable AI drafts, sender cleanup, commitments, and background sync.
• Workflow Intelligence — Always-on cognition is now framed as one reviewable loop: Memory is the source of truth, Heartbeat schedules reflection, Reflection evaluates evidence internally, and Suggestions are the user-facing output.
• Digital Twin Personas — Pre-built opt-in persona presets for every role. Twins stay visible and activatable, but they no longer own heartbeat, workflow-intelligence, or memory runtime state.
• Zero-Human Company Ops — Configure CoWork OS as a founder-directed autonomous company shell with venture workspace kits, automation-profile-backed operator agents, strategic planner loops, and Mission Control ops monitoring.
• Managed Devices — Run and inspect tasks across saved remote machines from a dedicated Devices tab with connection controls, summaries, remote task feeds, and remote file attachment picking.
• Automations Control Center — A single Automations section now separates Core Automation, Triggered Work, Remote/Device Work, and optional Twin surfaces instead of blending them into one runtime.
• 34 LLM provider options — 13 built-in providers plus 21 compatible/gateway options, including Claude, OpenAI, Google, Ollama, AWS Bedrock, and OpenRouter. Bring your own keys or run local models.
• 17 messaging channels — WhatsApp, Telegram, Discord, Slack, Teams, Google Chat, Feishu/Lark, WeCom, iMessage, Signal, X, and more. Chat with your AI from anywhere.
• 44 MCP connectors — Pre-built enterprise integrations across CRM, support, productivity, analytics, and payments, with native connector support for the most common internal surfaces.
• Composer @ mentions — Type @ in the message box to pick Agents, configured Integrations, or Files. Integration chips keep icon+name rendering in the prompt and task history while sending soft routing metadata to the runtime. Learn more
• 19 bundled packs — including Mobile Development, Game Development, 5 Financial packs, and the CoWork Shortcuts pack for message-box workflow aliases, with 100+ pack skills/shortcuts, in-app Plugin Store for installing community packs, remote registry, and enterprise admin policies for organization-wide control.
• 147 built-in skills — Documents, code review, web search, image generation, cloud integrations, game development, mobile development, financial analysis, infrastructure-as-code, React/Next.js performance guidance, and more, including bundled workflows such as llm-wiki for persistent research vaults, manim-video for programmatic technical animation, kami for editorial PDFs and slide decks, react-best-practices for React workspace changes, and taste-skill for high-agency frontend design.
• Profiles & portability — run separate CoWork profiles with isolated app data, export/import complete profile bundles, and keep credentials, channels, skills, and sessions separated by profile.
• Active Context sidebar — Always-visible panel showing 44 available MCP connectors with branded Lucide icons and enabled skills, auto-refreshed every 30 seconds.
• Agent teams — Multi-agent collaboration with shared checklists, graph-backed runs, collaborative mode, multi-LLM synthesis, persistent teams, and ACP-targeted delegation for local or remote specialists.
• External agent orchestration — Discover ACP agents, delegate work to A2A-compatible remote endpoints, and keep remote calls under the same approval and policy model as other governed actions, with persisted ACP tasks, restart-safe resume, remote cancel support, and the shared orchestration graph as the runtime backbone.
• Think With Me mode — Socratic brainstorming that helps you clarify thinking without executing actions.
• Chat mode — Direct LLM chat with no tools by default, no step timeline, same-session follow-ups, chat-only streaming for supported providers, and a narrow read-only analysis exception for uploaded PDFs that need deeper document reading.
• Build Mode — Go from idea to working prototype with a phased canvas workflow (Concept → Plan → Scaffold → Iterate) and named checkpoints.
• AI Playbook — Auto-captures what worked from successful tasks and injects relevant patterns into future prompts. Repeated patterns auto-promote to governed, one-click-approvable skills via the Playbook-to-Skill pipeline.
• Evolving Intelligence — The layered memory runtime keeps curated hot memory prompt-visible by default, leaves archive recall off by default, and adds explicit search_sessions / memory_topics_load recall paths for recent runs and topical packs. Adaptive Style Engine learns your communication preferences from message patterns and feedback. Evolution Metrics dashboard quantifies improvement over time (correction rate, knowledge growth, style alignment).
• Operator Runtime Visibility — Task completion now shows what Cowork learned, semantic batch labels and follow-up completions stay visible, unified recall spans tasks/messages/files, persistent shell sessions preserve operator state, and model routing/fallback is visible in the UI and Mission Control.
• Chronicle (Desktop Research Preview) — Opt-in local recent-screen context for vague prompts like “what is this?”, “why is this failing?”, or “sync the latest draft”, with consent-gated passive capture, pause/resume controls, promoted screen_context evidence in Mission Control, and optional linked background memory generation. Documentation: docs/chronicle.md.
• Configurable fallback chains — LLM and web-search providers can run in an explicit ordered fallback chain, including Exa for search, immediate failover on retryable provider errors, provider/model-level failover visibility in the UI, and a configurable cooldown before retrying the primary LLM route.
• Provider-aware prompt caching — Prompt caching is on by default for supported routes. CoWork keeps stable system sections cacheable, keeps volatile turn context out of the cached prefix, uses Anthropic automatic caching when available, falls back to explicit Claude breakpoints on OpenRouter, and derives stable OpenAI-family cache keys for GPT routes such as Azure gpt-5.4 / gpt-5.4-mini.
• Usage Insights — Dashboard showing task stats, cost/token tracking by model, activity heatmaps, top skills, per-pack analytics, per-persona success/retry metrics, and task-result satisfaction signals.
• ChatGPT History Import — Import your full ChatGPT conversation history. CoWork OS instantly knows your preferences, past projects, and context — no cold start. All data stays encrypted on your machine and never leaves it.
• Computer use (macOS) — Native desktop control via screenshot, click, type_text, keypress, and related tools with helper-targeted Accessibility + Screen Recording permissions, single-session sequential execution, Esc to abort, screenshot-relative coordinates, and stale-captureId protection. Prefer browser and shell tools for web and repo work; computer use is routed as a last-resort lane in policy and planning guidance. Documentation: docs/computer-use.md.
• Security-first — Approval workflows, sandboxed execution, configurable guardrails, encrypted storage, and a verified suite of 4,932 automated tests across 390 test files.
• Structured guidance — In plan-mode flows, the agent can pause with short multiple-choice prompts instead of ambiguous free-text follow-ups.
• Runtime resilience — Main interactive tasks run without an implicit turn window, while explicit turn caps, lifetime safety fuses, context-overflow recovery, and safe path normalization keep long-running work moving without silent file drift.
• Local-first & BYOK — Your data and API keys stay on your machine. No telemetry. No middleman.

Ideas & Media

Stable workflow entry points for the newest high-impact capabilities.

• Ideas panel — curated launchpad of pre-written workflow prompts and capability-aware starting points, with deep links into common tasks.
• Research vaults (llm-wiki) — first-class workspace-local knowledge bases inspired by Andrej Karpathy's LLM Wiki concept, with deterministic raw-source capture, Obsidian-friendly notes, filed-back outputs, vault search, and vault-health analysis. Learn more
• Everything Workbench — generated documents, spreadsheets, decks, web pages, PDFs, and previews share the same artifact model: task-feed card, sidebar open, fullscreen workspace, follow-up composer, and refresh after the agent finishes the requested edit. Learn more
• Document artifacts — task-created Word-style files render as compact artifact cards. .docx opens directly in an editable right-sidebar document surface with Google Docs-style controls, save, copy, external-open, fullscreen mode, and the same functional follow-up composer used by spreadsheet artifacts. .doc, .rtf, .odt, .ott, .pages, and related Word-style formats are recognized with best-effort preview or external-app/folder actions. Learn more
• Designed editorial documents — bundled kami skill for resumes, one-pagers, white papers, letters, portfolios, diagrams, and slide decks with workspace-local source scaffolding and PDF/PPTX export helpers. Learn more
• Format-aware file preview popup — clicking any file link in chat opens a unified preview modal that adapts to the format. Built-in support for HTML, Markdown, code (with highlight.js syntax colors), JSON / JSONL / GeoJSON (collapsible tree + raw toggle), CSV / TSV (RFC-4180 table), XLSX, DOCX, PDF (with page/OCR summary), images (fit / actual-size toggle, dimensions, alpha checkerboard), video, audio (with duration), LaTeX, and PPTX. Each format gets its own width profile, a header subtitle showing format-specific metadata, and a unified action bar with Copy path, Show in Finder, Open externally, and Close.
• Smart PDF attachments — uploaded PDFs are copied into the workspace, summarized into a compact prompt excerpt with page/extraction metadata, and read on demand with the document parser when the user asks for summaries, Q&A, extraction, comparison, or transformation. Scanned/image-heavy PDFs keep OCR/scan status visible, PDF excerpts are marked as untrusted document data, and read_pdf_visual remains reserved for layout, formatting, and page-appearance questions.
• Spreadsheet artifacts — task-created spreadsheet files render as compact artifact cards. Excel workbooks and CSV/TSV files open in the editable right-sidebar viewer; native Numbers, Google Sheets shortcut, ODS, XLSB, and other recognized spreadsheet outputs still get the same card and external-app/folder actions. Fullscreen mode expands editable sheets across the app with cell/range/row/column selection, copy, zoom, add row/column, save, model picker, voice input, attachments, and follow-up task context. Learn more
• Presentation artifacts — generated .pptx decks render as compact artifact cards and open by default in the resizable right-sidebar presentation viewer. The viewer shows thumbnails, slide navigation, zoom, a white slide canvas, speaker notes, text-first fast loading, cached rendered slide images, fullscreen follow-up context, and background refresh after requested deck edits. Legacy PowerPoint formats are recognized with external-app/folder actions. Learn more
• Web page artifacts — generated .html / .htm pages and built React output such as dist/index.html, build/index.html, or out/index.html render as compact artifact cards and open by default in a resizable right-sidebar sandboxed iframe preview. Fullscreen mode keeps the functional follow-up composer and refreshes after the relevant file or build output changes. React-style source projects without build output show a clear build-output-needed state instead of auto-starting a dev server. Learn more
• Browser Workbench / Browser V2 — live website testing opens a visible in-app browser in the right sidebar by default. Browser-use tools target that shared webview through Browser V2, show cursor movement during actions, prefer accessibility snapshot refs over selectors, expose console/network/download/storage diagnostics, support screenshots and annotation, and can expand to fullscreen with the normal follow-up composer. Learn more
• Image generation — configurable provider ordering across Gemini, OpenAI, Azure OpenAI, and OpenRouter.
• Video generation — text-to-video and image-to-video routing with polling tools and inline preview.
• Programmatic technical video — bundled manim-video skill for Manim CE explainers, equation walkthroughs, algorithm visualizations, and animated architecture/data stories. Learn more
• React/Next.js implementation guidance — bundled react-best-practices skill for React workspace changes, Next.js feature work, reviews, refactors, data-fetching improvements, bundle-size checks, and rendering-performance fixes. Learn more
• High-agency frontend design — bundled taste-skill for stricter anti-slop frontend work with stronger layout variance, typography, motion, and implementation rules.

See Workflow Intelligence, Core Automation, I Gave CoWork OS Workflow Intelligence, And Now It Learns From Reviewable Work | Full Guide, Features, Heartbeat v3, Providers, and Plugin Packs for current runtime details.

Latest Release

0.5.44 expands Browser V2 and gateway/channel documentation, improves Browser Use approvals and gateway routing, adds DeepSeek and NanoGPT provider coverage, fixes active-runtime reliability issues found during log review, and hardens the release pipeline for macOS, Windows, Linux server, npm, and GitHub artifacts. Start with Release Notes 0.5.44, then Browser V2 Architecture, Browser Workbench, Channel User Guides, Gateway User Guide, Features, Getting Started, and the Changelog.

Quick Start

Download the App

Download the latest release from GitHub Releases:

macOS first launch: CoWork OS macOS DMGs are currently unsigned. After dragging the app to Applications, open it once. If macOS blocks it, go to System Settings > Privacy & Security, scroll to Security, click Open Anyway next to "CoWork OS" was blocked to protect your Mac, then click Open Anyway in the confirmation dialog. Maintainers can create this unsigned DMG/ZIP with npm run package:mac:unsigned.

Windows first launch: Windows SmartScreen may show a warning for unrecognized apps. Click More info > Run anyway to proceed.

Works out of the box — defaults to OpenRouter's free model router, no API key needed.

Or Install via npm

npm install -g cowork-os
cowork-os

Windows npm install notes:

• Run npm install -g cowork-os / npm uninstall -g cowork-os from %USERPROFILE% (or another neutral directory), not from %APPDATA%\npm\node_modules\cowork-os, to avoid EBUSY lock errors.
• On Windows ARM64, first launch may take longer while native modules are rebuilt; this can run multiple fallback steps before the app opens.
• If native rebuild fails, install Visual Studio Build Tools 2022 (C++) and Python 3, then retry.
• If startup logs show ERR_FILE_NOT_FOUND ... dist/renderer/index.html, reinstall the latest package and check Troubleshooting.

Or Build from Source

git clone https://github.com/CoWork-OS/CoWork-OS.git
cd CoWork-OS
npm install && npm run setup
npm run build && npm run package

Windows prerequisites: Native module setup may require Visual Studio Build Tools 2022 (C++) and Python 3. On Windows ARM64, setup automatically falls back to x64 Electron emulation when ARM64 native prebuilds are unavailable.

npm run setup also installs local git hooks (.githooks/) including a pre-commit secret scan. If needed, reinstall hooks with npm run hooks:install.

See the Development Guide for prerequisites and details.

How It Works

1. Create a task or start from Ideas — Describe what you want ("organize my Downloads by file type", "create a quarterly report spreadsheet", "draft a DOCX memo", "build a small landing page") or begin from a curated Ideas prompt. No workspace needed — a temp folder is used automatically if you don't select one.
2. Choose a mode — Pick Chat, Execute, Plan, Analyze, or Verified for the runtime behavior, then optionally toggle Autonomous (auto-approve actions), Collaborative (multi-agent perspectives), or Multi-LLM (compare providers with a judge) per task.
3. Monitor execution — Watch the real-time task timeline as the agent plans, executes, and produces artifacts. Parallel tool bursts are grouped into lane summaries, and shell commands run in a live terminal view where you can see output in real-time, stop execution, or provide input (e.g. y/n) directly.
4. Respond when needed — Destructive operations require explicit approval (unless Autonomous mode is on), and plan-mode tasks can pause for structured multiple-choice input before continuing.

Features

Agent Runtime

Task-based execution with dynamic re-planning, five runtime modes (Chat, Execute, Plan, Analyze, Verified) plus orchestration toggles (Autonomous, Collaborative, Multi-LLM, Think With Me), a shared turn kernel, metadata-driven tool scheduling, graph-backed delegation, typed worker roles, optional workflow-pipeline execution with per-phase model routing, agent teams with persistence, agent comparison, git worktree isolation, AI playbook, and performance reviews. Learn more

Skills now follow an additive runtime model: CoWork can proactively shortlist or apply a relevant skill, but the original task remains canonical. Skills add context and scoped execution modifiers instead of replacing the task prompt. Learn more

Chronicle (Desktop Research Preview)

Chronicle is an opt-in desktop-only recent-screen context lane for vague on-screen references such as this, that, the failing one, latest draft, or why is this failing. Configure it from Settings > Memory Hub > Chronicle: passive capture is consent-gated, can be paused from Settings or the tray, resolves through screen_context_resolve, and promotes only task-used observations into existing recall, evidence, and optional linked screen_context memory entries instead of creating a second memory system. Learn more

Research Vaults (llm-wiki)

CoWork OS includes llm-wiki as a bundled, first-class research-vault workflow inspired by Andrej Karpathy's LLM Wiki idea: keep a /raw corpus, build durable linked notes on top, and make the result easy for agents to traverse later.

You can launch it from the GUI with a normal prompt such as Build a persistent Obsidian-friendly research vault for GRPO papers, from the welcome/onboarding starter cards, or with /llm-wiki when you want explicit slash syntax.

llm-wiki creates and maintains a workspace-local markdown wiki with:

• immutable raw/ source captures
• deterministic ingest helpers for articles, papers, repos, datasets, and images
• durable Obsidian-friendly notes and maps
• deterministic vault-first search across notes, raw captures, and filed slide decks
• filed-back Marp slide decks and SVG charts under outputs/
• SCHEMA.md, index.md, log.md, and inbox.md
• a GUI vault browser on the welcome screen for core files, recent notes, recent queries, outputs, and raw captures
• deterministic vault analysis for orphans, broken links, bridge pages, surprising cross-section links, and suggested follow-up questions

It works in desktop and gateway channels, supports inline chaining, and writes inspectable run artifacts alongside the persistent vault. GUI-first prompts can start the flow even before you supply a topic, in which case CoWork asks one short scoping question first. Learn more

Operator Runtime Visibility makes the runtime's learning and routing visible: task detail surfaces now show the learning progression, unified recall spans tasks/messages/files/workspace notes/memory/KG, shell sessions preserve operator state, and live routing/fallback events are surfaced in Mission Control and the task UI. Learn more

Workflow Intelligence reflections now use the same runtime with stricter safeguards: they start only after memory services are initialized, write durable target-scoped artifacts under .cowork/subconscious/ for compatibility, default to reviewable suggestions, learn from act/edit/snooze/dismiss/ignore feedback, require isolated git worktrees for trusted code-change auto-create paths, and skip non-git workspaces when isolation is required. See Workflow Intelligence and Troubleshooting.

Output Completion UX

Completion state and file availability are now explicit:

• Everything Workbench: generated knowledge-work outputs share one artifact model across docs, sheets, decks, pages, PDFs, and previews, reducing the need to jump into separate office apps for review, light editing, and follow-up changes
• High-signal completion toast: finished tasks with outputs show Task complete with filename/count and actions for Open file, Show in Finder, and View in Files
• Right-sidebar focus: if you are viewing the completed task and the panel is collapsed, the Files panel auto-opens and highlights the primary output
• Unseen-output badge: if completion happens in another task/view, the collapsed right-panel toggle shows a numeric badge until you open Files
• Filename-first rows with clear location context: Files rows stay filename-only, with output folder context shown separately (or Workspace root)
• Artifact parity: artifact-only outputs are treated the same as normal file outputs in completion toasts, timeline details, and Files panel
• Spreadsheet artifact workbench: spreadsheet outputs render as compact cards. .xlsx, .xls, .xlsm, .csv, and .tsv open into a resizable editable sidebar by default; .numbers, .gsheet, .ods, and .xlsb are recognized as spreadsheet artifacts with external-app/folder actions. Editable sheets can expand into fullscreen spreadsheet mode with copy/save controls, zoom, attachments, voice input, and model selection for follow-up prompts
• Document artifact workbench: Word-style outputs render as compact cards. .docx opens into a resizable editable sidebar by default and can expand into fullscreen document mode with a Google Docs-style toolbar, save, copy, attachments, voice input, model selection, follow-up context, and automatic preview refresh after follow-up edits. .doc, .rtf, .odt, .ott, .pages, and related formats are recognized with preview/external actions according to local parser support
• LaTeX/PDF artifact workbench: explicit LaTeX/TikZ paper tasks can write .tex, compile it with a system TeX engine, and show the source plus rendered PDF together with Summary, source, and PDF tabs
• Smart PDF attachment reading: user-uploaded PDFs keep the initial prompt compact while preserving the workspace-relative path, page count, extraction mode, OCR/scan status, and a safe excerpt. For deeper PDF content tasks, CoWork calls parse_document on demand instead of stuffing the whole PDF into every turn; visual PDF inspection stays on read_pdf_visual
• Presentation artifact workbench: .pptx outputs show inline deck cards in task events and assistant summaries, then open into a resizable sidebar or fullscreen viewer with thumbnails, previous/next navigation, zoom, speaker notes, fast text-first loading, cached rendered slide images, external actions, and follow-up prompt context. .ppt, .pptm, .potx, .potm, .ppsx, and .ppsm are recognized with external-app/folder actions
• Web page artifact workbench: generated .html / .htm files and built React output entrypoints show compact web page cards, then open into a resizable sidebar or fullscreen sandboxed iframe preview with browser/folder/copy actions, persisted sidebar width, and follow-up prompt context. React-style projects without dist, build, or out HTML output show a build-output-needed preview state rather than auto-running a dev server
• Semantic completion labels: completed tool batches and verifier verdicts now feed the richer completion text shown in timelines, feed relays, and export surfaces

Guided Input & Runtime Recovery

Long-running tasks now have clearer operator handoffs and stronger recovery defaults:

• Structured input cards: plan-mode tasks can pause with 1-3 short multiple-choice prompts, with answers captured inline in the desktop UI or via the Control Plane web dashboard
• Adaptive turn recovery: main execute-mode tasks are uncapped at the turn-window level by default, still reserve room for a final answer, and use lifetime/emergency safety stops plus bounded follow-up recovery instead of strategy-assigned 30/30 style windows; explicit maxTurns / windowTurnCap still opt tasks into capped behavior
• Context overflow retry: context-capacity failures trigger compaction and retry instead of immediate hard failure when the model context window is exceeded
• Path repair: /workspace/... aliases and drifted relative paths can be normalized back into the active workspace or pinned task root, with strict-fail policies available when you want hard enforcement
• Parallel timeline lanes: read-only tool batches render as grouped timeline rows so the UI stays readable even when searches/fetches run concurrently

Mission Control

Centralized orchestration and monitoring cockpit with a Kanban board, real-time activity feed, core automation profile visibility, and a Core Harness view for traces, failure clusters, evals, experiments, and learnings. Learn more | Core Automation

Devices

The Devices tab turns CoWork OS into a multi-machine control surface. Save and reconnect remote CoWork nodes, inspect device summaries (activity, apps, storage, alerts, resource signals), launch tasks against a selected machine, browse that machine's remote workspaces, and attach files directly from the remote filesystem before dispatching a task. Learn more

Automations

Automations are now organized around a hard boundary: Workflow Intelligence is the always-on cognitive loop, while Routines are the main saved-automation product layered on top of lower-level execution surfaces. Scheduled Tasks, Webhooks, and Event Triggers still exist, but they now also serve as advanced or compiled backends for routines rather than competing first-class automation concepts. Task view can also turn the current task into a real scheduled task from the three-dot menu with Add automation..., preserving the source task title, ID, and cowork://tasks/<taskId> deeplink. The home dashboard, welcome-screen Next actions, and routines panel surface recent automation suggestions/runs so you can monitor background systems without hunting through tabs. Learn more | Task Automations

Zero-Human Company Ops

CoWork OS can be configured as a founder-operated autonomous company shell: venture workspace kit context, a dedicated Settings > Companies control surface, company-linked operator agents, automation profiles, strategic planner issue generation, and Mission Control ops monitoring. Create the company in Companies, activate operator personas such as Company Planner and Founder Office Operator, then attach automation where needed and monitor the company loop from Mission Control. Learn more | Core Automation

Digital Twin Personas

Role-specific AI twins that handle cognitive overhead as optional persona presets. Pick a template (Software Engineer, Engineering Manager, Product Manager, VP, Founder Office Operator, Company Planner, and more), customize it, and activate it as a role preset with recommended skills and prompt/personality defaults. Twins can be linked to a company for company-aware operations, but they no longer own heartbeat or Workflow Intelligence policy directly. Learn more

Live Canvas & Build Mode

Agent-driven visual workspace for interactive HTML/CSS/JS content, data visualization, and iterative image annotation. Build Mode adds a phased idea-to-prototype workflow with named checkpoints and revert support. Learn more

Multichannel Gateway

Unified AI gateway across 17 channels with security modes, rate limiting, ambient mode, scheduled tasks, and a shared message lifecycle for commands, active-task follow-ups, cancellations, progress delivery, skill slashes, and scheduled outputs. WhatsApp supports /new, /new temp, /stop, editable progress updates, and hidden temporary scratch workspaces; Slack supports multiple workspaces, Telegram supports group-routing policies and allowlists, Discord can be limited to specific guilds, and Feishu/Lark plus WeCom are first-class channels. Learn more | Per-channel guides | User guide | Message lifecycle

Inbox Agent

Local-first inbox workspace that turns email into an action queue while preserving normal email-client controls. It keeps cached mail visible on restart, syncs in the background, supports manual reply/reply-all/forward, and surfaces the right next step for each thread: triage, draft, cleanup, commitment tracking, and scheduling. Learn more

• Classic + Today modes: use the familiar three-pane inbox or the Today lanes for Needs action, Happening today, Good to know, and More to browse
• Action cards: Unread, Needs reply, Suggested Actions, Open Commitments
• Mailbox views: Inbox, Sent, All, plus Recent/Priority sorting, saved views, account filters, and practical domain filters
• Normal email actions: manual Reply, Reply all, Forward, To/Cc/Bcc, editable subject/body, and provider-backed send
• AI draft handling: generated replies are editable before send, and sent drafts disappear after successful send
• Ask Inbox: ask mailbox questions in the right sidebar, see live retrieval/reading/answer steps, and search local FTS, semantic mailbox embeddings, provider-native mail search, and indexed attachment text with LLM summaries when configured
• @Inbox composer routing: type @Inbox or @inbox ... in the main composer to open Inbox Agent, switch to Ask Inbox, and run the query there
• Workflow buttons: Cleanup, Follow-up, Reply, Forward, Mark done, Prep thread, Extract todos, Schedule, Refresh intel
• Gmail auto-forwarding: create forwarding automations from a Gmail thread with dry-run support, attachment filters, per-message dedupe, and thread-scoped execution
• Commitment tracking: accept commitments into real follow-up tasks, mark already-handled threads done, or dismiss items
• Background sync: load from the local database immediately and refresh in the background without blanking the inbox on restart

Infrastructure

Built-in cloud infrastructure tools — no external processes or MCP servers needed. The agent can spin up sandboxes, register domains, and make payments natively.

• Cloud Sandboxes (E2B): Create, manage, and execute commands in isolated Linux VMs. Expose ports, read/write files — all from natural language.
• Domain Registration (Namecheap): Search available domains, register, and manage DNS records (A, AAAA, CNAME, MX, TXT).
• Crypto Wallet: Built-in USDC wallet on Base network. Auto-generated, encrypted in OS keychain. Balance displayed in sidebar.
• x402 Payments: Machine-to-machine HTTP payment protocol. Agent can pay for API access automatically with EIP-712 signed USDC transactions (requires approval).

All infrastructure operations that involve spending (domain registration, x402 payments) require explicit user approval. Configure in Settings > Infrastructure. Learn more

Web Scraping

Advanced web scraping powered by Scrapling with anti-bot bypass, stealth browsing, and structured data extraction. Three fetcher modes — fast HTTP with TLS fingerprinting, stealth with Cloudflare bypass, and full Playwright browser. Includes batch scraping, persistent sessions, proxy support, and five built-in skills (web scraper, price tracker, site mapper, lead scraper, content monitor). Configure in Settings > Web Scraping. Learn more

Integrations

• Cloud Storage: 6 integrations, including Notion, Box, OneDrive, Google Workspace, Dropbox, and SharePoint
• 44 MCP Connectors: pre-built enterprise integrations for CRM, support, productivity, analytics, and payments (Salesforce, Jira, HubSpot, Zendesk, Stripe, Tavily, Grafana, Metabase, Socket, and more), with connector notifications available as trigger inputs for automations
• Composer mentions: type @ in the message box to choose configured integrations. Google Workspace is shown as Gmail, Google Drive, and Google Calendar; gateway channels and MCP connectors appear only when locally usable. Mentions render as icon+name chips and are passed as soft routing hints, not permission grants. Learn more
• Developer Tools: glob/grep/edit_file, Playwright browser automation, MCP client/host/registry

Learn more

Active Context Sidebar

Real-time overview of your active integrations, always visible in the right panel. Shows connected MCP connectors (44 available) and native integrations with branded Lucide icons (HubSpot, Salesforce, Google Workspace, Discord, GitHub, Postgres, and more) and green status dots, plus enabled skills from active packs. Each section shows 4 items with internal scrolling for more. Auto-refreshes every 30 seconds. Learn more

Usage Insights

Dashboard with task metrics, cost/token tracking by model, prompt-cache read telemetry (cachedTokens and cache-read rate where available), activity heatmaps (day-of-week and hourly), top skills usage, per-pack analytics, persona-level success/retry/cost breakdowns, and task-result thumbs up/down quality signals with 7/14/30-day period selection. Access from Settings > Usage Insights. Learn more

LLM Providers

34 provider options, with 13 built-in providers and 21 compatible/gateway providers. Use cloud APIs or run fully offline with Ollama, configure an ordered fallback chain for runtime failover, and get default-on prompt caching on supported Claude and GPT-style routes. Claude supports both direct API keys and Claude subscription tokens from claude setup-token, with live Refresh Models support in Settings. Learn more

Plugin Platform & Customize

Unified plugin platform with 19 bundled packs (Engineering, DevOps, Product, Sales, QA, Finance, CoWork Shortcuts, and more), each bundling skills, agent roles, connectors, slash command aliases, and "Try asking" prompts. Packs can link to Digital Twin personas as optional role presets.

• Search & filter: Real-time sidebar search across pack names, descriptions, categories, and skill names
• Per-skill control: Enable or disable individual skills within a pack without toggling the whole pack
• Persistent toggles: Pack and skill states survive app restarts
• Update detection: Background version checks against the registry with visual indicators
• "Try asking" in chat: Empty chat shows randomized prompt suggestions from active packs
• Plugin Store: In-app marketplace for browsing, installing (Git/URL), and scaffolding custom packs, now with install-time security scanning and quarantine handling for imported packs
• Skill Store & external skills: Desktop GUI support for CoWork Registry skills, direct ClawHub installs, and generic external skill imports from Git repos, raw manifests, and SKILL.md bundles, with managed scan reports and quarantine for unsafe imports
• External skill directories: Add shared read-only skill folders without importing or copying those skills into CoWork's managed directory
• Remote Registry: Community pack catalog with search and category filtering
• Admin Policies: Organization-level controls — allow/block/require packs, restrict installations, set agent limits, distribute org-managed packs from a shared directory
• Per-pack analytics: Usage Insights dashboard groups skill usage by parent pack

Access from Settings > Customize. Learn more

Best-Fit Workflows

CoWork OS ships purpose-built packs and Tier-1 connectors for three operational lanes where governed AI delivery has the clearest ROI:

These are the workflows where approval gates, local data control, and measurable outcome delivery pay off most — and where CoWork OS is a vendor-swap-friendly alternative to point solutions or BPO tooling. Learn more

Extensibility

• 147 built-in skills across developer, productivity, communication, documents, frontend, game development, mobile development, financial analysis, infrastructure-as-code, and more
• Custom skills in ~/Library/Application Support/cowork-os/skills/ (macOS) or %APPDATA%\cowork-os\skills\ (Windows)
• 19 bundled plugin packs with 100+ pack skills/shortcuts, message-box slash aliases, and Digital Twin integration where applicable
• Plugin Store — browse, install from Git/URL, scaffold custom packs, and review quarantine/report state for imported packs
• Skill Store — browse CoWork Registry skills, search ClawHub, import external skills from Git, raw JSON, or raw SKILL.md, and review quarantine/report state for imported skills
• MCP support — client, host, and registry

Voice Mode

Text-to-speech (ElevenLabs, OpenAI, Web Speech API), speech-to-text (Whisper), and outbound phone calls. Learn more

Knowledge Graph

Built-in structured entity and relationship memory backed by SQLite. The agent builds a knowledge graph of your workspace — people, projects, technologies, services, and their relationships — with 10 dedicated tools, FTS5 search, multi-hop graph traversal, temporal validity on edges (valid_from / valid_to), historical as_of queries, auto-extraction from task results, and confidence scoring with decay. Learn more

Memory & Context

Persistent memory with privacy protection, FTS5 search, LLM compression, and a contract-driven workspace kit (.cowork/) for durable human-edited context. The runtime now makes memory explicit as a four-layer wake-up model: L0 Identity and L1 Essential Story are prompt-visible by default, while L2 Topic Packs and L3 Deep Recall stay tool-driven through memory_topics_load, search_sessions, search_memories, and exact-span search_quotes. Runtime-native checkpoints capture both compact structured summaries and verbatim evidence packets before compaction, on meaningful task completion, and periodically during long runs.

The workspace kit separates workspace-wide files such as AGENTS.md, USER.md, MEMORY.md, TOOLS.md, SOUL.md, IDENTITY.md, RULES.md, VIBES.md, and LORE.md from project-scoped files such as .cowork/projects/<projectId>/CONTEXT.md and .cowork/projects/<projectId>/ACCESS.md. Special files get dedicated lifecycle handling: BOOTSTRAP.md is a one-time onboarding checklist tracked through .cowork/workspace-state.json, while HEARTBEAT.md is reserved for recurring Heartbeat v3 checklist work instead of general task context.

Every tracked file follows a shared parser/linter model with freshness windows, secret detection, missing-file status, and revision snapshots stored under .cowork/**/.history/. Workspace kit health is surfaced in the app and can be checked locally with npm run kit:lint for human-readable output or JSON export. Import your ChatGPT history to eliminate the cold-start problem — CoWork OS knows you from day one. All imported data is stored locally and encrypted on your machine. Structured memory observations add inspectable local metadata, progressive recall tools, Memory Hub privacy controls, deterministic rebuild/backfill, and soft-delete suppression on top of archive memory. Optional Supermemory integration adds an external provider lane with supermemory_profile, supermemory_search, supermemory_remember, and supermemory_forget, plus optional prompt-time profile injection and background mirroring of non-private local memory captures. Proactive session compaction automatically generates comprehensive structured summaries when context reaches 90% capacity, and checkpoint capture preserves exact supporting spans so recall quality survives compaction. Learn more | Structured Memory | Supermemory | Context Compaction

Role-specific agent personas and intent-first task startup

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                    Security Layers                               │
│  Channel Access Control │ Guardrails & Limits │ Approval Flows   │
└─────────────────────────────────────────────────────────────────┘
                              ↕
┌─────────────────────────────────────────────────────────────────┐
│                    React UI (Renderer)                           │
│  Task List │ Timeline │ Approval Dialogs │ Live Canvas           │
└─────────────────────────────────────────────────────────────────┘
                              ↕ IPC
┌─────────────────────────────────────────────────────────────────┐
│                 Agent Daemon (Main Process)                      │
│  Task Queue │ Agent Executor │ Tool Registry │ Cron Service      │
└─────────────────────────────────────────────────────────────────┘
                              ↕
┌─────────────────────────────────────────────────────────────────┐
│                    Execution Layer                               │
│  File Ops │ Skills │ Browser │ LLM Providers (30+) │ MCP        │
│  Infrastructure (E2B Sandboxes │ Domains │ Wallet │ x402)       │
└─────────────────────────────────────────────────────────────────┘
                              ↕
┌─────────────────────────────────────────────────────────────────┐
│  SQLite DB │ Knowledge Graph │ MCP Host │ WebSocket │ Remote Access  │
└─────────────────────────────────────────────────────────────────┘

See Architecture for the full technical deep-dive.

Security

Top security score on ZeroLeaks — outperforming many commercial solutions
View Full Report

• Configurable guardrails: Token budgets, cost limits, iteration caps, dangerous command blocking
• Approval workflows: User consent required for destructive operations
• Sandbox isolation: macOS sandbox-exec (native), Docker containers, or process-level isolation on Windows
• Encrypted storage: OS keychain + AES-256 fallback
• 4,932 automated tests in the suite across 390 test files (4,854 passed, 78 skipped), including 135+ security unit tests under tests/security/ and 250+ control-plane and WebSocket protocol tests

See Security Guide and Security Architecture for details.

Deployment

Screenshots

Live Canvas for visual, iterative build workflows

Structured spreadsheet outputs generated directly in chat

Enterprise connector setup with OAuth and credential management

Usage Insights dashboard with performance, efficiency, and activity trends

Extensions manager for installed plugin packs and lifecycle controls

Pack-level customization with per-skill toggles and suggested prompts

Roadmap

Planned

• VM sandbox using macOS Virtualization.framework
• Network egress controls with proxy
• Linux desktop support

See CHANGELOG.md for the full history of completed features.

Documentation

Data Handling

• Stored locally: Task metadata, timeline events, artifacts, workspace config, memories (SQLite)
• Sent to provider: Task prompt and context you choose to include
• Not sent: Your API keys (stored via OS keychain), private memories

Compliance

Users must comply with their model provider's terms: Anthropic · AWS Bedrock

Contributing

See CONTRIBUTING.md for guidelines.

License

MIT License. See LICENSE.

_{"Cowork" is an Anthropic product name. CoWork OS is an independent open-source project and is not affiliated with, endorsed by, or sponsored by Anthropic. If requested by the rights holder, we will update naming/branding.}