Connect the local bridge to run your own model. The interface will never pretend an offline model produced an answer.
SOVEREIGNAGENT
A cinematic command surface for a custom local agent core: reasoning, research, code, memory, browser automation, voice and image workflows — without making a paid cloud model the foundation.
ENTERSYSTEM↘
*Bound by your hardware, storage and electricity — not token billing.
02 / COMMAND CENTER
Ask once.
Make the system work.
Fast mode keeps latency low. Deep mode adds planning and critique. Swarm mode fans the task across research, code, verification and a final judge.
03 / SPECIALIST SWARM
One model.
More discipline.
COREcustom python orchestrator
Planner
Decomposes goals into dependency-aware steps before execution.
PLANResearcher
Searches self-hosted SearXNG, reads pages, reconciles sources.
WEBCoder
Inspects repositories, edits files, runs tests and interprets failures.
CODECritic
Finds weak assumptions, omissions and contradictions before delivery.
REVIEWVerifier
Checks evidence, commands, tests, citations and task completion.
VERIFYJudge
Chooses the strongest candidate and requests another pass when needed.
DECIDEVision
Optional multimodal model path for screenshots, images and video frames.
SEEVoice
Optional whisper.cpp speech input and local text-to-speech workflow.
HEAR04 / LOCAL TOOLCHAIN
Own the runtime.
Own the workflow.
Qwen3 source
Reasoning / tool-use model family source reference.
MODELTransformers
Direct Python inference path for local checkpoints.
INFERENCEllama.cpp
GGUF inference through local CLI — server mode remains optional.
GGUFFAISS + SentenceTransformers
Local vector memory, semantic retrieval and RAG.
MEMORYPlaywright + Trafilatura
Browser automation and clean web-document extraction.
BROWSERTree-sitter + ripgrep
Repository structure, symbols and high-speed code search.
CODEPodman
Container isolation for risky or reproducible tool execution.
SANDBOXwhisper.cpp
Local speech recognition without sending audio to a cloud model.
VOICEDiffusers + ComfyUI
Optional local image generation, editing and node workflows.
IMAGE05 / MEMORY SYSTEM
Remember locally.
Retrieve deliberately.
SQLite stores runs and notes. FAISS and SentenceTransformers can add semantic memory when installed. The core still works without cloud memory.
06 / SOURCE VAULT
Every supplied repository.
Preserved, not flattened.
Large archives are split into Cloudflare-safe chunks and can be reconstructed byte-for-byte with the included restore script.
07 / LAUNCH
Cloud surface.
Local intelligence.
Local mode
Run the bundled bridge, point it to llama-cli and your GGUF model, then open the local interface.
powershell -ExecutionPolicy Bypass -File .\START_LOCAL_AGENT_WINDOWS.ps1
Search
Optional: run your own SearXNG with Podman for private metasearch grounding.
.\START_SEARXNG_PODMAN.ps1
Cloudflare UI
Deploy this ZIP to Pages. The browser can connect back to the localhost bridge after you enter your local token.