Your personal AI agent. Give me an objective and I’ll plan, use tools, and work toward the result.
ULTRON is the operating identity. The imported model is its reasoning brain.
Intent recognition, task classification, complexity routing, dependency-aware plans, continuation and completion checks.
Conversation continuity, semantic knowledge retrieval, relevance ranking, confidence, source tracking and contradiction handling.
Tool relevance checks, explicit-intent detection, approval gates, scoped execution and untrusted-output isolation.
Draft validation, evidence checks, failure reporting, uncertainty handling and verification of claimed actions.
Model-aware context budgets, minimal inference passes, bounded retries and device-capability awareness.
Stable ULTRON voice, language matching, proportional detail, accessibility and sensitive-topic handling.
ULTRON runs a programmed orchestration layer around the selected reasoning model. Custom preferences cannot override its core policies, permissions, or approval gates.
One secure place for compatible model-provider keys.
Sign in to connect provider keys to the private ULTRON cloud runtime.
Automatic routing uses only free-tier model IDs where the provider publishes them. Open OpenRouter API keys ↗Choose a provider .txt file. The provider is detected from its filename.
The file is read locally, never uploaded. Only individual keys are sent securely to ULTRON and stored in Supabase Vault. Use filenames such as openrouter.txt, groq.txt or google-ai-studio.txt.No API key saved
ULTRON detects supported key formats automatically. Android stores the key encrypted with the device Keystore. Only compatible chat-completion APIs are enabled; unsupported key types are identified instead of being silently accepted.One connected knowledge graph from every conversation.
Choose a node to inspect its linked conversations and source messages.
Inspect what ULTRON noticed, what it proposes to improve, and what has actually been verified.
ULTRON records diagnostic metadata only. It cannot rewrite its own code, change permissions, or mark an improvement verified without your explicit review and evidence.
A proposal is not a code change. “Record external implementation” only records that you changed something elsewhere. Verification requires test or runtime evidence that you provide.
Tools ULTRON builds, changes he proposes, and the actual history of what ran.
Checking owner session and Supabase connection…
Cloud-backed tools and skills run through ULTRON’s private Supabase runtime. Phone-only capabilities remain on the device. This panel shows real versions, proposals, execution records and failures.
Give ULTRON controlled access to your repositories. Start read-only, then approve each change when it matters.
Review the exact action and arguments before an agent can change or send anything.
Choose 2–5 agents. Each receives the objective independently and runs in parallel.
Checking Android access…
Voice backend: device speech fallback. ULTRON's dedicated local neural voice model is not installed yet.
Not connected. Local-only mode is active.
Authentication uses Supabase. On Android, the refreshable session is encrypted with AES-GCM and a key held by the Android Keystore; web preview uses browser storage.Checking native model runtime…
Checking device memory and storage…
Recommended: Qwen 2.5 1.5B for a balanced starter, Gemma 3 1B for lower storage, or Qwen 3 1.7B for more demanding reasoning. Keep extra free storage for context and temporary files.
Search the Hugging Face catalog. Downloaded weights stay in Ultron's private app storage.
Models are community-published. Review each model's license and source before use. Downloads can be several gigabytes and require Wi-Fi and free storage. On-device llama.cpp runs GGUF weights. Import a .tgz or .tar.gz archive containing GGUF files to extract them automatically. Safetensors and PyTorch checkpoints need a separate runtime and are not executable in this build.