A local desktop app for monitoring Codex, Claude Code, and Cursor conversations across projects. Name chats, add notes, attach Jira references, and see their last reported activity. Opt in per source to answer Claude Code questions and Claude Code/Codex permission requests. The app does not submit follow-up prompts, launch chats, or control processes.
Requires Node.js 24+ and npm. Python is not needed for the desktop app.
npm ci
npm run dev
For a production build or a separate demo workspace:
npm run build
npm start
# Or:
npm run demo
The development demo command opens sample projects and chats in a separate SQLite database. Restart normally to return to your real workspace.
The blue Operations workspace keeps projects on the left, compact chat records in the center, and the selected chat’s notes, ticket, and activity in the right sidebar. Chat name, project assignment, project renaming, and Archive/Restore are visible near the top of chat details. Every project name is editable, including repository labels carried over from older versions. New chats default to their repository project; accessible local Git worktrees share the parent repository's project. Moving a chat with the Project selector replaces its membership, including choosing No project. Names, assignments, and archives survive provider updates and restarts.
Use the flag beside a project in the left sidebar to mark it as priority; the pencil renames any project. Active chats in priority projects get a single red line along the left edge when they need input, report an error, finish a turn, or end a session. Hover over a marked chat to see the attention reason. The project's red lamp remains visible across chat filters, and its tooltip shows the number of chats needing attention. Unfinished or uncertain subagent work does not count as completed work. Priority follows the chat's current project, survives restarts, and clears from a chat when it resumes, is archived, or moves out of that project. Remove the flag to stop marking that project's chats.
An orange lamp appears on the Windows taskbar icon or macOS Dock icon while any open, unarchived chat reports Needs input. It clears when those chats resume, end, or are archived, regardless of list filters or window focus. The indicator reflects the last reported event. The app-rendered, draggable title bar integrates navigation. macOS keeps its traffic lights and system menu; Windows uses native window buttons without a separate title or menu bar.
Open Settings → Sources → Add source. Choose the provider and native or WSL environment. The app generates an authenticated HTTP hook configuration under its local data directory. Merge it into the provider's user settings, preserving unrelated settings/hooks and replacing the deck's old entries. Provider configuration is never edited automatically.
Native Claude Code uses HTTP hooks directly. WSL sources configured by the Windows deck use command hooks that invoke Windows curl.exe, keeping HTTP on the Windows side. Cursor and Codex also use command hooks that POST stdin with curl. No PowerShell or Bash collector scripts, separate Node installation, or Python are required.
To answer in the deck, select a Claude Code or Codex source and enable Answer questions and permissions in the deck (Codex: Answer permissions in the deck). Merge the regenerated hook fragment and review/trust the new Codex hooks. Existing sources remain observation-only until you opt in. Requests wait up to five minutes and then return to the agent's own interface; disabling responses also returns pending requests. Keep waiting dismisses Spotlight while the alert stays available. Codex questions remain in Codex, and Claude network sandbox permissions may still require its native interface. See live responses for coverage and delivery limits.
Subagent start/stop hooks keep background work visible after the main turn finishes: Main turn finished · 2 subagents working. These chats remain in Running until the observed workers finish; missing identities or observations older than five minutes show an uncertain status. Merge the latest configuration from Sources for existing installations, and review/trust new Codex hooks. See subagent observations for provider differences and event coverage.
Claude's existing Stop/SubagentStop hooks also expose background monitors and scheduled checks. The deck shows Monitoring · waiting for next check between turns, with a separate Monitoring filter/count and observation details. No hook configuration changes are needed when these events already forward their complete payloads. Detection requires a Claude version that reports these fields; Codex/Cursor monitoring is not inferred. See background monitoring for freshness and snapshot limits.
For Claude Code in WSL, run the Windows app, create a WSL source per distro, and configure Claude inside that distro. The receiver defaults to http://127.0.0.1:43821. WSL hooks launch "/mnt/c/Windows/System32/curl.exe" through Windows executable interoperability and POST to Windows localhost, without requiring Windows directories on the Linux PATH. This route needs no WSL host-IP, mirrored-networking configuration, or inbound firewall exception. Windows interoperability must be enabled. Adjust the executable path in your hook commands if the Windows drive or WSL mount root differs. See WSL setup and diagnostics.
Keep the deck running while using your agents. New and existing/resumed chats appear on their next submitted prompt; merely opening or closing an unused chat does not add it. Historical conversations are not scanned. Events sent while the deck is closed are not captured. Accepted observation events are queued on disk before ingestion; pause/resume and storage retries retain that queue. Live permission and question requests remain bound to their current hook connection and are never replayed after restart.
Command hooks finish quietly and successfully when the deck is unavailable, without interrupting the agent. The Sources diagnostic command still reports delivery failures so you can test the connection. Native Claude HTTP hooks allow execution to continue on connection failure, but Claude may display a hook error notice.
Send diagnostic event checks local inbox ingestion only. The Settings → Sources section also provides an HTTP test command to run in the provider environment. Submit a real prompt in an existing chat to verify the complete hook installation.
Today is the main workspace: plan a dated card with Coding, Agentic, and Other sections, create or continue tasks, and keep their notes, tickets, links, and observed chats together. Open linked chats inline to see the latest question, approval request, prompt, and response. Import a pasted YAML/Markdown transcription or upload a file, review it beside the previous card, and save an optional evening wrap-up. Repositories provides observed chat history grouped by detected repository. See daily planning for the workflow and card format.
SQLite databases live under Electron’s OS application-data directory, with separate workspace/ and demo/ subdirectories. Jira settings are saved separately in each database. The exact active path appears in the app footer. Hook files stay local. There is no cloud sync, telemetry, or Jira API access.
Optional Settings → AI chat matching uses OpenRouter or direct Amazon Bedrock to link today's chats to tasks. Saving a key does not enable matching; the checkbox controls sending bounded chat/task context to the selected model. Keys use OS-backed encryption and never appear in snapshots or the database. See automatic chat links.
Local CPU matching keeps startup and inference diagnostics in local-matching/worker.log beneath the active data directory. The log is capped at 4 MiB and replaced when a worker starts; exception details can include chat input, so review it before sharing. UI and exported errors stay generic and point to the log.
The app backs up an existing database before schema migrations. For a manual backup, quit the app and copy the displayed data directory. Keep backups private: they contain your notes and previews.
On machines used for work, install the agent-based commit confidentiality guard
with npm run guard:install (Claude Desktop review by default, with no standalone
Claude CLI; choose Codex with npm run guard:install -- --reviewer codex). It blocks commits containing suspected work data
or enterprise configuration, and blocks when review is uncertain or unavailable.
See commit confidentiality review for setup and limits.
For the standalone, synthetic chat-to-task AI experiment, run npm run test:matching
and npm run eval:matching. See matching evaluation
for adding an OpenRouter key and explicitly enabling live requests.
npx playwright install chromium --only-shell # once per machine, for headless UI tests
npm run typecheck
npm test # storage, provider, and HTTP tests; Electron's Node runtime only, no windows
npm run test:ui # builds and tests the renderer in headless Chromium, including database restarts
npm run check # routine desktop checks; no visible windows
npm run test:electron # opt-in Electron startup, IPC isolation, native lamps and window controls
npm run check:all # routine checks plus the windowed Electron suite
npm run package # unpacked app in release/
npm run test:packaged # run the small Electron suite against the built application bundle
npm run dist # platform ZIP in release/
Routine UI tests use the real SQLite store, HTTP hooks, durable inboxes and workspace commands behind a test-only adapter for window.deck; they do not validate Electron preload isolation or OS window behavior. Clipboard writes are validated and external-browser requests are captured in the test adapter. The explicit Electron suite checks startup, preload/IPC isolation, restart persistence, native input lamps and window controls. Routine CI runs the headless suite on macOS and Windows; the packaging workflow runs the small packaged Electron suite.
Build on the target OS. CI is configured for macOS and Windows; real workplace provider/WSL validation still needs to be done there. Distribution builds are unsigned. Managed computers may require IT approval to run the app or allow user-defined hooks.
CI runs Desktop checks (npm run check) on pull request updates and pushes to main. Feature-branch pushes do not also start a duplicate run, and newer updates cancel superseded runs for the same PR or branch.
Packaged app checks runs on macOS and Windows when dependency manifests, packaging/build configuration, the packaged-test runner or its Electron tests change, and on v* release tags. Before a release, use Actions > Packaged app checks > Run workflow to validate the current build; enable upload_packaged_builds there when downloads are needed. Downloads are retained for three days. Bridge reference checks runs only for changes under bridge/ or schema/, changes to Makefile or its workflow, or a manual run. Keep the path-filtered packaging and bridge workflows optional in branch protection; the always-running desktop checks can remain required.
The packaging tool's @electron/get dependency is overridden to v5 to avoid its older vulnerable HTTP cache dependency; Node 24 supports that downloader. The lockfile pins the tested dependency graph.
src/main/: SQLite, hook adapters/inboxes, HTTP reception/configuration, and Electron lifecycle.src/preload/: narrow validated IPC interface; the renderer has no Node access.src/renderer/: React workspace and source setup UI.src/shared/: types and input validation.tests/: persistence/integration tests and desktop UI automation.The reference bridge still runs with make bridge-install && make bridge-serve. make check checks both desktop and bridge; npm run check checks only the desktop.
Hardware source is preserved at tag hardware-final-2026-10-04. Firmware, simulator, and BLE dependencies are retired. The separate LilyGo checkout is untouched.
A local desktop app for monitoring Codex, Claude Code, and Cursor conversations across projects. Name chats, add notes, attach Jira references, and see their last reported activity. Opt in per source to answer Claude Code questions and Claude Code/Codex permission requests. The app does not submit follow-up prompts, launch chats, or control processes.
Requires Node.js 24+ and npm. Python is not needed for the desktop app.
npm ci
npm run dev
For a production build or a separate demo workspace:
npm run build
npm start
# Or:
npm run demo
The development demo command opens sample projects and chats in a separate SQLite database. Restart normally to return to your real workspace.
The blue Operations workspace keeps projects on the left, compact chat records in the center, and the selected chat’s notes, ticket, and activity in the right sidebar. Chat name, project assignment, project renaming, and Archive/Restore are visible near the top of chat details. Every project name is editable, including repository labels carried over from older versions. New chats default to their repository project; accessible local Git worktrees share the parent repository's project. Moving a chat with the Project selector replaces its membership, including choosing No project. Names, assignments, and archives survive provider updates and restarts.
Use the flag beside a project in the left sidebar to mark it as priority; the pencil renames any project. Active chats in priority projects get a single red line along the left edge when they need input, report an error, finish a turn, or end a session. Hover over a marked chat to see the attention reason. The project's red lamp remains visible across chat filters, and its tooltip shows the number of chats needing attention. Unfinished or uncertain subagent work does not count as completed work. Priority follows the chat's current project, survives restarts, and clears from a chat when it resumes, is archived, or moves out of that project. Remove the flag to stop marking that project's chats.
An orange lamp appears on the Windows taskbar icon or macOS Dock icon while any open, unarchived chat reports Needs input. It clears when those chats resume, end, or are archived, regardless of list filters or window focus. The indicator reflects the last reported event. The app-rendered, draggable title bar integrates navigation. macOS keeps its traffic lights and system menu; Windows uses native window buttons without a separate title or menu bar.
Open Settings → Sources → Add source. Choose the provider and native or WSL environment. The app generates an authenticated HTTP hook configuration under its local data directory. Merge it into the provider's user settings, preserving unrelated settings/hooks and replacing the deck's old entries. Provider configuration is never edited automatically.
Native Claude Code uses HTTP hooks directly. WSL sources configured by the Windows deck use command hooks that invoke Windows curl.exe, keeping HTTP on the Windows side. Cursor and Codex also use command hooks that POST stdin with curl. No PowerShell or Bash collector scripts, separate Node installation, or Python are required.
To answer in the deck, select a Claude Code or Codex source and enable Answer questions and permissions in the deck (Codex: Answer permissions in the deck). Merge the regenerated hook fragment and review/trust the new Codex hooks. Existing sources remain observation-only until you opt in. Requests wait up to five minutes and then return to the agent's own interface; disabling responses also returns pending requests. Keep waiting dismisses Spotlight while the alert stays available. Codex questions remain in Codex, and Claude network sandbox permissions may still require its native interface. See live responses for coverage and delivery limits.
Subagent start/stop hooks keep background work visible after the main turn finishes: Main turn finished · 2 subagents working. These chats remain in Running until the observed workers finish; missing identities or observations older than five minutes show an uncertain status. Merge the latest configuration from Sources for existing installations, and review/trust new Codex hooks. See subagent observations for provider differences and event coverage.
Claude's existing Stop/SubagentStop hooks also expose background monitors and scheduled checks. The deck shows Monitoring · waiting for next check between turns, with a separate Monitoring filter/count and observation details. No hook configuration changes are needed when these events already forward their complete payloads. Detection requires a Claude version that reports these fields; Codex/Cursor monitoring is not inferred. See background monitoring for freshness and snapshot limits.
For Claude Code in WSL, run the Windows app, create a WSL source per distro, and configure Claude inside that distro. The receiver defaults to http://127.0.0.1:43821. WSL hooks launch "/mnt/c/Windows/System32/curl.exe" through Windows executable interoperability and POST to Windows localhost, without requiring Windows directories on the Linux PATH. This route needs no WSL host-IP, mirrored-networking configuration, or inbound firewall exception. Windows interoperability must be enabled. Adjust the executable path in your hook commands if the Windows drive or WSL mount root differs. See WSL setup and diagnostics.
Keep the deck running while using your agents. New and existing/resumed chats appear on their next submitted prompt; merely opening or closing an unused chat does not add it. Historical conversations are not scanned. Events sent while the deck is closed are not captured. Accepted observation events are queued on disk before ingestion; pause/resume and storage retries retain that queue. Live permission and question requests remain bound to their current hook connection and are never replayed after restart.
Command hooks finish quietly and successfully when the deck is unavailable, without interrupting the agent. The Sources diagnostic command still reports delivery failures so you can test the connection. Native Claude HTTP hooks allow execution to continue on connection failure, but Claude may display a hook error notice.
Send diagnostic event checks local inbox ingestion only. The Settings → Sources section also provides an HTTP test command to run in the provider environment. Submit a real prompt in an existing chat to verify the complete hook installation.
Today is the main workspace: plan a dated card with Coding, Agentic, and Other sections, create or continue tasks, and keep their notes, tickets, links, and observed chats together. Open linked chats inline to see the latest question, approval request, prompt, and response. Import a pasted YAML/Markdown transcription or upload a file, review it beside the previous card, and save an optional evening wrap-up. Repositories provides observed chat history grouped by detected repository. See daily planning for the workflow and card format.
SQLite databases live under Electron’s OS application-data directory, with separate workspace/ and demo/ subdirectories. Jira settings are saved separately in each database. The exact active path appears in the app footer. Hook files stay local. There is no cloud sync, telemetry, or Jira API access.
Optional Settings → AI chat matching uses OpenRouter or direct Amazon Bedrock to link today's chats to tasks. Saving a key does not enable matching; the checkbox controls sending bounded chat/task context to the selected model. Keys use OS-backed encryption and never appear in snapshots or the database. See automatic chat links.
Local CPU matching keeps startup and inference diagnostics in local-matching/worker.log beneath the active data directory. The log is capped at 4 MiB and replaced when a worker starts; exception details can include chat input, so review it before sharing. UI and exported errors stay generic and point to the log.
The app backs up an existing database before schema migrations. For a manual backup, quit the app and copy the displayed data directory. Keep backups private: they contain your notes and previews.
On machines used for work, install the agent-based commit confidentiality guard
with npm run guard:install (Claude Desktop review by default, with no standalone
Claude CLI; choose Codex with npm run guard:install -- --reviewer codex). It blocks commits containing suspected work data
or enterprise configuration, and blocks when review is uncertain or unavailable.
See commit confidentiality review for setup and limits.
For the standalone, synthetic chat-to-task AI experiment, run npm run test:matching
and npm run eval:matching. See matching evaluation
for adding an OpenRouter key and explicitly enabling live requests.
npx playwright install chromium --only-shell # once per machine, for headless UI tests
npm run typecheck
npm test # storage, provider, and HTTP tests; Electron's Node runtime only, no windows
npm run test:ui # builds and tests the renderer in headless Chromium, including database restarts
npm run check # routine desktop checks; no visible windows
npm run test:electron # opt-in Electron startup, IPC isolation, native lamps and window controls
npm run check:all # routine checks plus the windowed Electron suite
npm run package # unpacked app in release/
npm run test:packaged # run the small Electron suite against the built application bundle
npm run dist # platform ZIP in release/
Routine UI tests use the real SQLite store, HTTP hooks, durable inboxes and workspace commands behind a test-only adapter for window.deck; they do not validate Electron preload isolation or OS window behavior. Clipboard writes are validated and external-browser requests are captured in the test adapter. The explicit Electron suite checks startup, preload/IPC isolation, restart persistence, native input lamps and window controls. Routine CI runs the headless suite on macOS and Windows; the packaging workflow runs the small packaged Electron suite.
Build on the target OS. CI is configured for macOS and Windows; real workplace provider/WSL validation still needs to be done there. Distribution builds are unsigned. Managed computers may require IT approval to run the app or allow user-defined hooks.
CI runs Desktop checks (npm run check) on pull request updates and pushes to main. Feature-branch pushes do not also start a duplicate run, and newer updates cancel superseded runs for the same PR or branch.
Packaged app checks runs on macOS and Windows when dependency manifests, packaging/build configuration, the packaged-test runner or its Electron tests change, and on v* release tags. Before a release, use Actions > Packaged app checks > Run workflow to validate the current build; enable upload_packaged_builds there when downloads are needed. Downloads are retained for three days. Bridge reference checks runs only for changes under bridge/ or schema/, changes to Makefile or its workflow, or a manual run. Keep the path-filtered packaging and bridge workflows optional in branch protection; the always-running desktop checks can remain required.
The packaging tool's @electron/get dependency is overridden to v5 to avoid its older vulnerable HTTP cache dependency; Node 24 supports that downloader. The lockfile pins the tested dependency graph.
src/main/: SQLite, hook adapters/inboxes, HTTP reception/configuration, and Electron lifecycle.src/preload/: narrow validated IPC interface; the renderer has no Node access.src/renderer/: React workspace and source setup UI.src/shared/: types and input validation.tests/: persistence/integration tests and desktop UI automation.The reference bridge still runs with make bridge-install && make bridge-serve. make check checks both desktop and bridge; npm run check checks only the desktop.
Hardware source is preserved at tag hardware-final-2026-10-04. Firmware, simulator, and BLE dependencies are retired. The separate LilyGo checkout is untouched.