Alfred Desktop

Alfred Desktop (clients/desktop) is a native Mac/Linux control surface for a local install. It is the optional client tier of the layered install: the core fleet and CLI run fully standalone without it.

Slack stays Alfred’s collaboration surface. The desktop app is for local trust and repair: what needs attention, which plans are waiting, why a run failed, which memory candidates are ready, and which local actions are safe to run next. It is a thin control surface, not a second runtime.

Design note and run commands: docs/DESKTOP_CLIENT.md.

The local control surface

flowchart TB
    subgraph client["Alfred Desktop (clients/desktop, Tauri)"]
        ui["React UI tabs:<br/>Inbox / Ask / Work / Agents / Setup"]
        native["native command allowlist:<br/>start runtime, status, agents,<br/>auth, memory, redis,<br/>safe dry-run"]
    end

    subgraph core["core fleet (headless)"]
        serve["alfred serve<br/>localhost JSON API"]
        state[("$ALFRED_HOME/state<br/>firings / plans / memory")]
        cli["Alfred CLI: bin/alfred"]
        fleet["lib/agent_runner + bin/*.py"]
    end

    gh["GitHub issue / PR links"]
    slackthread["Slack plan threads"]

    ui -->|local JSON over 127.0.0.1| serve
    native -->|narrow allowlist, no arbitrary shell| cli
    serve --> state
    cli --> fleet
    fleet --> state
    ui -.->|open outside the app| gh
    ui -.->|open outside the app| slackthread

Tabs

Tab	Job
Inbox	The decision queue: repeated failures, blocked plans, follow-ups, memory candidates, recent runs, shipped work, and the Claude/Codex capacity rail.
Ask	Plain-language planning intake backed by the same readiness engine as Slack.
Work	Kanban board, saved plans, Slack follow-ups, local draft actions, and issue queue controls.
Agents	Roster, activity feed, latest-run inspector, memory learning queue, and safe per-agent controls.
Setup	Start or reconnect to the local runtime and run fleet/auth/agent/memory/Slack checks in-app.

Boundary

The client reads and writes the same local surfaces Alfred already uses: $ALFRED_HOME, alfred serve, the Alfred CLI, GitHub issue and PR links, Slack plan threads, and the local fleet brain. It opens no public port and does not add a separate scheduler.

The boundary is enforced in the Tauri layer. The fetch command only allows Alfred JSON API paths on http://localhost, http://127.0.0.1, or http://[::1]. Local plan/run details stay in native inspector panes; explicit Slack and GitHub links open outside the app. State-changing controls use a narrow native allowlist (start runtime, run checks, safe dry-runs, pause, resume, run once, and local follow-up planning) and surface command audit detail with the result. There is no arbitrary shell execution. In a plain browser preview, native actions are unavailable.

Run it locally

alfred serve --port 7010 --no-browser
cd clients/desktop
npm install
npm run tauri dev

The app uses 7010 because macOS can reserve 7000 for Control Center. Legacy saved 7000 URLs are treated as stale local configuration and rewritten to 7010. Setup lets you enter a custom localhost URL when needed.

Native installers

tauri.conf.json builds the native installer for the host platform:

cd clients/desktop
npm run tauri -- build

Host	Artifacts
macOS 11+ on Apple silicon	.app and .dmg
Linux	.AppImage and .deb

Continuous integration builds with --no-bundle to prove the binary compiles without code signing or packaging. The release pipeline publishes a signed and notarized macOS DMG and app zip, plus Linux AppImage and Debian artifacts.