LOCAL_ONLY_001
PASS / CERTIFIED
Deep2 + GGUF local inference, fail-closed when local policy is violated.
Local-capable coding agents. Honest certification boundary.
Your machine. Your models. Your agents. Any screen.
ScreenPilot is the professional Command surface for steering RawrXD local coding agents: workspace-aware, Git-aware, approval-aware, and built around a certified local inference lane.
Cloud/model hosting is a separate future lane. Tonight’s release stays desktop-first.
Command · This PC · local workspace
01
Plan
02
Build
03
Verify
Approval waiting
Network-facing remote control remains fail-closed until EGRESS_001.
Deep2 localGit visibleAgent active
Plan, build, verify, or steer an agent...
Tonight’s drop
Drop-ready means scoped, working, and defensible.
Command-home desktop surface
Local Deep2/GGUF agent lane
Workspace and Git context binding
Live activity, stop, and approval states
Phone/browser control designed, not exposed until egress is certified
Certification posture
Local-first is certified. No-dependency product is not claimed yet.
STREAMER_CERT_001
10/10 PASS
TinyLlama Q4_K_M live lane at ~11–13 e2e tok/s.
EGRESS_001
OPEN
Whole-product default-deny egress and residual provider quarantine are not claimed yet.
Internal benchmark snapshot · 2026-08-31
RawrXD ScreenPilot vs top agentic IDEs
Fourteen weighted dimensions across Cursor, Windsurf, GitHub Copilot, Claude, and Zed. The score is a date-stamped internal comparison, with N/A dimensions excluded per product—not an independent industry certification.
Command Home is confirmed live with journal resume, Plan / Build / Agent steering, and Ctrl+Shift+W work mode. Cloud and phone steering remain outside the certified boundary.
RawrXD weighted score
72%
Rank vs five references
#3
Lead dimensions
4
Weighted comparison72 / 100
CursorWindsurfCopilotClaudeZed
What the score means
RawrXD leads where local sovereignty, native execution, and bounded action matter most. The gap is concentrated in cloud agents, codebase retrieval wired into the Command home, marketplace depth, and final UI polish.
Long-tail tools such as Continue, Cline, Cody, and Aider are outside this five-reference scorecard. Product capabilities and plan limits change; re-score before using this as a purchasing or investment claim.
Where RawrXD leads
Sovereignty is the differentiator.
LOCAL_ONLY_001
In-process GGUF / Deep2
Local weights load without an Ollama pull or cloud fetch.
BOUNDARY
Fail-closed command broker
Session bind, write leases, approval queue, and event journal on disk.
WORKFLOW
Two-shell UX
Command-home landing plus Ctrl+Shift+W full Win32 IDE, reversible.
PERFORMANCE
Native Win32 lane
No Electron; optional cloud compile paths remain gated off by default.
Honest gaps vs leaders
The close path is explicit.
The next gains are integration work, not a claim that the open gates are already closed.
| Gap | Reference | Close path |
|---|---|---|
| Codebase @ indexing / RAG | Cursor | Wire semantic index into the Command-home steer handler |
| Cloud / phone steering | Cursor Agents | REMOTE_CONTROL_001 after EGRESS_001 |
| UI polish / menu E2E | Cursor / VS Code | P1_UI_MENU_E2E_001; shell phases 5–6 |
| Hosted agent parity | Windsurf | screenpilot.tech remains preview; live use needs local :11435 |
| Marketplace DX | Copilot | Work Mode VSIX; hosted HTML remains honest offline preview mode |
Dimension matrix
Evidence, boundary, next move.
The matrix records RawrXD’s current evidence and the comparison note used in the internal scorecard.
| Dimension | RawrXD evidence | Boundary / comparison note |
|---|---|---|
| Agent loop (multi-step autonomy) | Broker + stop + journal | Bounded by LOCAL_ONLY_001 and approvals |
| Codebase @ indexing / RAG | Semantic index in Work Mode | Not wired to Command-home steer |
| Multi-file Composer edits | Agent diff in Work Mode | Command-home is chat-first |
| Local GGUF / in-process inference | Deep2 + LOCAL_ONLY_001 | No default Ollama or cloud pull |
| Egress / air-gap posture | Fail-closed architecture | EGRESS_001 open; quarantine remains explicit |
| Approval gates + write leases | Session bind + write lease | Git commit/push approval and SharedAgentWorkRegistry |
| Command-home vs IDE split | Plan / Build / Agent + Ctrl+Shift+W | Reversible two-shell workflow |
| Native performance (no Electron) | Win32 + MASM | Zed is the Rust-native reference |
| Cloud / remote agents | Not exposed yet | REMOTE_CONTROL_001 waits for EGRESS_001 |
| Extension marketplace | VSIX in Work Mode | Hosted HTML is preview-only |
| Terminal / tool execution | PowerShell + bounded agent | CapExecute gated in Command mode |
| Git integration in UI | Context bar + rail | Branch and M/A/D state visible |
| Pricing / self-host | Local hardware and self-host | No cloud IDE subscription |
| Ship maturity / polish | P1 UI ~72% | Desktop Preview label stays honest |
Next ship
Rebuild with Plan / Build / Agent prefixes and the loopback LocalServer; load GGUF in the hub; smoke on the new SHA; record STATIC_DEP only when SSL is active.
Cloud lane
Want hosted models later?
Use real GPU compute for model serving, not shared website hosting. Keep the marketing site, app shell, and model API as separate layers so local certification stays clean.