Ghost

Ghost is the deterministic trust layer for desktop automation.

Ghost lets users automate sensitive desktop file workflows by recording once, reviewing exactly what happens, approving within policy boundaries, auditing every operation, and undoing changes when needed.

The Problem

Bookkeepers, practice admins, and operations teams spend hours every week organizing invoices, receipts, statements, and exports across local folders and desktop apps. This work is repetitive, high-frequency, and painful — but it is also too sensitive for black-box AI agents, too desktop-specific for Zapier or Make, and too small-business-oriented for enterprise RPA.

Existing tools don't fit:

• Macro recorders replay blind coordinates; they break when windows move.
• Zapier/Make excel at cloud APIs, not local file workflows.
• UiPath/Enterprise RPA are powerful but expensive and overbuilt for SMBs.
• AI agents are unpredictable and hard to audit—terrifying with financial files.
• Manual scripts require technical skill.
• Keyboard Maestro/BetterTouchTool don't offer audit trails, policy enforcement, or undo.

Ghost's Answer

Ghost is not an autonomous agent. Ghost is a deterministic, auditable automation engine built around a simple pipeline:

Intent → Plan → Policy Check → Human Approval → Execution → Audit → Undo

Deterministic: Same input always produces the same output. No hallucination. No surprise behavior.

Auditable: Every operation is logged with full context. Workflows are reviewable before execution.

Policy-enforced: Users define Zones (folders Ghost may touch) and Capabilities (what Ghost may do). Everything else is denied by default.

Undoable: Reversible operations write undo data before execution. Users can recover from mistakes.

Get Ghost

Platform	Download
macOS (Apple Silicon + Intel)	Ghost.dmg
Windows 10 / 11 (64-bit)	Ghost_Setup.exe

Note: current builds are developer-preview quality. macOS builds may be ad-hoc signed; if macOS blocks the app, approve it under System Settings → Privacy & Security. Notarized releases are in progress.

How It Works: Ghost Organizer

The starting workflow is Ghost Organizer: safe, reviewable file organization with audit and undo.

Step-by-Step Flow

1. Select Zone — Choose a folder Ghost may access (e.g., Downloads, Desktop).
2. Scan Files — Ghost analyzes filenames, extensions, and patterns.
3. Classify — Deterministic rules categorize files (invoices, receipts, statements, etc.).
4. Propose Plan — Ghost suggests moves and renames, with confidence scores and matched rules.
5. Review & Approve — User sees before/after preview. Can edit destinations, deselect actions, or cancel.
6. Policy Check — Ghost Guard flags conflicts, overwrites, and out-of-Zone operations.
7. Write Undo Journal — Before touching anything, Ghost writes a local undo record.
8. Execute — Only approved actions run. Progress is shown step-by-step.
9. Audit — Every operation logged with timestamps, decisions, and hashes.
10. Undo — One-click recovery of the entire run.

Demo: Five Minutes

1. Open Ghost Organizer.
2. Select ~/Downloads.
3. Click "Scan."
4. Ghost proposes: "Move invoice_acme.pdf to /Clients/Acme/Invoices/2026/" with matched rule shown.
5. Review the before/after tree. Click "Approve."
6. Ghost executes and audits. Shows: "43 moved, 31 renamed, 5 folders created, 7 skipped, 0 blocked."
7. Undo is available.

That's the product. Boring. Trustworthy.

Who Ghost Is For

• Bookkeepers and practice admins organizing client invoices, receipts, statements, and exports.
• Office managers and small business operations teams drowning in downloads and attachments.
• Students and nonprofits managing financial or operational workflows that require audit trails.

Ghost is deliberately not for:

• Autonomous agents you launch and forget about.
• Unattended, scheduled automation.
• Workflows that don't need review before execution.

Every meaningful action requires your approval.

The Trust Pipeline

Ghost's architecture enforces one principle:

Ghost may suggest anything, but it only does what you have explicitly approved, inside boundaries you control.

Every operation flows through the same pipeline — no exceptions:

Intent → Plan → Policy Check → Human Approval → Execution → Audit Log → Undo Path

Core Guarantees

• Deny by default — Operations are blocked unless explicitly allowed by your Zones and Capabilities.
• Deterministic execution — Suggestions (including any AI-assisted ones) never execute directly; only reviewed, approved plans do.
• Suppressed secrets — Secure-field input is never retained—not in logs, not in recordings, not in previews.
• Local-first — Workflows, logs, and settings live on your machine. No account. No cloud dependency. No telemetry unless you opt in.
• Auditable — Every run leaves a tamper-evident audit log and undo journal.

Current Status

Early-stage but functional. Working now:

• ✅ Ghost Organizer — Safe file organization with preview, policy checks, audit, and undo
• ✅ Deterministic compression — Semantic steps from raw events
• ✅ Ghost Guard policy engine — Deny-by-default rules engine
• ✅ Zones and Capabilities — Local boundary enforcement
• ✅ Audit logs and undo journals — Full operation history with recovery
• ✅ macOS and Windows — Desktop builds for both platforms

Not yet production-ready:

• ⚠️ Release signing — macOS notarization and Windows code signing in progress
• ⚠️ Cross-app reliability — Some workflows still depend on screen coordinates (fallback for robustness)
• ⚠️ Richer target resolution — Window-title awareness and window-relative positioning in development

The Wedge: Ghost Organizer

Why start with file organization?

1. High frequency — Bookkeepers and admins do this every week.
2. Sensitive — Files contain financial and client data. Black-box automation is unacceptable.
3. Audit-heavy — These users need compliance trails anyway.
4. Measurable ROI — 2–3 hours/week saved per user is easy to quantify.
5. Clear UX — Folder preview and file list are familiar patterns.

Ghost Organizer proves the trust pipeline works for a real, painful workflow. Then we expand.

Roadmap

Phase 1: Scope Reduction (Weeks 1–2)

• Remove AI language and experimental features from public UI
• Publish formal trust pipeline spec
• Define canonical workflows and success metrics

Phase 2: Reliability Foundation (Weeks 3–5)

• Build benchmark suite with ≥99.5% success target for canonical workflows
• Add comprehensive unit and integration tests
• Publish reliability metrics

Phase 3: Production Hardening (Weeks 6–8)

• Apple Developer ID signing and notarization
• Windows code signing
• Security documentation and threat model

Phase 4: Organizer UX Polish (Weeks 9–10)

• First-run onboarding flow
• Improved preview and policy UI
• Dark mode and keyboard shortcuts

Phase 5–8: Traction and YC (Months 4–6)

• Private beta with 10 users
• 3 paid pilots from SMB firms
• Public beta launch
• YC application with metrics and testimonials

Architecture

Stack: Rust backend (Tauri 2), vanilla HTML/CSS/JS frontend, local SQLite storage.

Core Modules:

• organizer/ — File scanner, classifier, planner, executor, undo
• policy/ — Deny-by-default capability engine
• audit/ — Append-only audit logs and undo journals
• core/compression/ — Event compression to semantic steps
• platform/ — macOS and Windows capture/replay backends

See IMPLEMENTATION.md for detailed status and docs/ for architecture specs.

Security & Privacy

• Local-first by default — No cloud, no account, no telemetry unless opted in.
• No camera, no microphone — Input capture only during explicit recording/approved replay.
• Suppress secrets — Secure fields are never stored, logged, or transmitted.
• Signed releases — Code signing and notarization for safe installation.

See SECURITY.md for threat model and responsible disclosure.

Development

# Install Rust and Tauri CLI
cargo install tauri-cli --version "^2.0" --locked

# Run checks
make ci         # fmt + test + clippy
make check      # cargo check
make test       # cargo test
make dev        # cargo tauri dev

# Build
make build      # cargo tauri build --no-bundle

See AGENTS.md for contribution guidelines, architecture details, and the binding contract for humans and AI agents working on Ghost.

Contributing

Contributions are focused on:

• Replay reliability and semantic target resolution
• Ghost Organizer UX and correctness
• Safety and policy hardening
• Docs and examples
• Bug fixes

Not accepting:

• Feature bloat or experimental half-finished work
• Cloud-dependent features (local-first is non-negotiable)
• Autonomous execution modes (user approval is the product)

FAQ

Does my data leave my machine? No. Workflows, logs, and settings are local. No account. No server. Telemetry is off by default.

Can Ghost delete files? Not in Ghost Organizer—delete is blocked by default. All mutations are previewed, approved, audited, and undoable.

What about passwords? Secure-field input is suppressed at capture time and never retained. Ghost Guard blocks replay of workflows containing secrets.

Can it run while I'm away? No. Replay runs when you start it and stays interruptible. Unattended execution is deliberately not a current feature.

Why not just use a macro recorder? Raw macros replay coordinates and break when windows move—and they'll happily type your password into the wrong field. Ghost reviews, guards, re-resolves targets semantically, and leaves an audit trail.

License

MIT — see LICENSE.

---

Ghost is boring. That is the point. Trust means predictability, auditability, and control. Ghost is built to earn your trust by showing exactly what it does—before, during, and after.