DESKTOP PREVIEW

Your machine. Your models. Your agents. Any screen.

The local-first command surface for steering coding agents across your workspaces.

Desktop Preview now — not a finished remote-control or cloud IDE release.

Request Desktop Preview View certification scope Compare

Cloud/model hosting is a separate future lane. This release stays desktop-first.
Hosted preview: rawrxd-screenpilot.garrettaosborne.chatgpt.site
Agent shell: screenpilot.tech/gui/ide_chatbot_standalone.html — live use needs local http://127.0.0.1:11435/gui on your PC.

Certification posture

Internal product snapshot · 2026-08-31. Public copy lags sealed evidence — never leads it. RC1 release authority remains held until P1PRA_002 Finalize=0.

LOCAL_ONLY_001 — PASS / CERTIFIED

Deep2 + GGUF local inference, fail-closed when local policy is violated.

STREAMER_CERT_001 — 10/10 PASS

Token-generation streamer certification retained.

STREAMER_VRAM_APERTURE_001 — PASS

MASM aperture evidence retained. Not reopened by BATCH_085.

MODEL_STREAMER_PRODUCT_PATH — WIRED

Real shared InferenceEngine. Synthetic token fallback removed.

AGENTIC_AUTONOMOUS_TEMPLATE_001 — WIRED

Plan / Build / Agent are behavioral execution modes — not prompt cosmetics.

WIN32IDE_LINK — PASS

P1PRA_SYMBOL_LINKED=1 on the product EXE.

TOKEN_PRESSURE_VALVE_001 — WIRED

Orthogonal decode nozzle (needle/mist/pulse/rinse/cutoff/repair). Does not change ctx, temp, GPU split, or model path. Not a knowledge substitute.

P1PRA_002 / RC1 — HOLD

Product inference authority awaits end-to-end Finalize=0 witness (Load → Send → stream → lease release).

EGRESS_001 — OPEN

Whole-product default-deny egress not claimed. Remote phone/web steering remains Coming next.

RawrXD ScreenPilot vs top agentic IDEs

Weighted internal comparison across Cursor, Windsurf, GitHub Copilot, Claude, and Zed — not an independent industry certification. Bounded claim: real local-model streaming + behaviorally distinct Plan/Build/Agent. Release authority held pending P1PRA_002.

~85%
BATCH_085 target (defensible)
HOLD
RC1 until Finalize=0
4
Waves to 85%
INFERENCE

.gguf → InferenceEngine stream

StreamingGGUFLoader → AgenticModelStreamerBridge → Tokenize / Eval / Generate.

STEER

Plan / Build / Agent / Stop

Checklist+approve, streaming edit lane, multi-turn tools, cancel — CommandBroker (message, mode).

BOUNDARY

Fail-closed command broker

Session bind, write leases, approval queue, event journal on disk.

ROADMAP

IDE_PRODUCT_FINISH_BATCH_085

Waves 1–2 WIRED. Waves 3–4 (Git/search/persist/smoke) after P1PRA witness.

Honest gaps vs leaders

GapReferenceClose path (BATCH_085)
Token pressure / anti-loop nozzleTOKEN_PRESSURE_VALVE_001 (orthogonal to ctx/temp/GPU/path)
P1PRA Finalize=0 E2E witnessPhysical Load + Build/Ask Send → E2E.log
Context assembly depth / RAGCursorWave 1 context + Wave 3 search wire-up
Build edit → run → repair loop polishCursor / WindsurfWave 2 edit txn + terminal observe
Cloud / phone steeringCursor AgentsREMOTE_CONTROL_001 after EGRESS_001
Hosted agent parityWindsurfscreenpilot.tech stays preview; live agent is Win32IDE / local :11435
Session restore + Git/search Command UXVS Code / CopilotWave 3 backlog

Dimension matrix

DimensionRawrXD evidenceBoundary / note
Agent loopBroker + Stop + StartAgentLoopLOCAL_ONLY_001 + approvals
Local GGUF inferenceReal InferenceEngine product pathRoots include G:\OllamaModels; no default Ollama pull
Plan / Build / AgentBehavioral modes WIREDNot cosmetic prompt prefixes
Multi-file editsBuild stage → diff → Allow/DenyWave 2 edit transaction
Egress postureFail-closed architectureEGRESS_001 open
Product authorityP1PRA stage ledger linkedRC1 HOLD until Finalize=0
Command-home vs IDECtrl+Shift+W two-shellReversible workflow
Ship maturityBATCH_085 → defensible 85%~15% polish/parity reserved after
Cloud / remote agentsNot exposedComing next after EGRESS_001
Native performanceWin32 + MASMZed is Rust-native reference

Next ship

Seal P1PRA_002_E2E_PRODUCT_AUTHORITY with Finalize=0, then Wave 3–4 (Git, search, persistence, approval center, settings, smoke, evidence bundle). Frozen cert lanes stay untouched.

Tonight’s drop

Command-home UI

Workspace and Git context, agent transcript, composer, activity strip, and Stop control.

Local Deep2 / GGUF

In-process inference lane. Scan and load weights from local roots such as G:\OllamaModels.

Model hub

Browse and load local .gguf / .bin files. No cloud weight fetch, no ollama pull in this path.

Approval boundaries

Fail-closed control channel. LocalServer auto-start suppressed in Command mode.

Coming next

Remote steering

Phone and web control via paired-device service — TLS, pairing, scoped tokens. Not a public model port on your PC.

ScreenPilot Sovereign Pro

Sovereign Pro — perpetual license

$599 one-time

Full local IDE product license. Perpetual for the current major version. No subscription.

  • Ad-free screenpilot.tech documentation access
  • Local Deep2 / GGUF execution on your hardware
  • MASM-native acceleration and workspace / Git integration
  • Agentic plan / build / test workflow (Command-home + Work Mode)
  • Offline, local-first operation — one user, multiple owned devices

Native Win32 application remains ad-free for all users. Cloud inference and remote steering are not included; cloud usage stays separately metered. Remote control remains gated until EGRESS_001 closes.

Enterprise deployment

Team license

$2,500–$10,000

Per-deployment team licensing for private workspaces and controlled rollout.

Private / sovereign deployment

$50k–$2M+

Hardened private deployment, sovereign/defense builds, and enterprise support contracts.

screenpilot — desktop preview shell
[ workspace ] /projects/rawrxd   branch: main
LOCAL_ONLY_001   STREAMER_CERT_001   STREAMER_VRAM_APERTURE_001
MODEL_STREAMER_PRODUCT_PATH   AGENTIC_AUTONOMOUS_TEMPLATE_001
[!] P1PRA_002 / RC1 HOLD — Finalize=0 witness required
[!] EGRESS_001 open — egress claims withheld
[ steering ] Remote phone / web — Coming next
> Local Deep2 / GGUF → InferenceEngine stream…