Bring your own key. Get back reasoning you can audit.

DUD3P0 ("P0") is a bring-your-own-key chatbot, research tool, and cognitive amplifier for iPhone. Point it at something hard — a question to research, a system to design, a problem to think through — and it reasons out loud, labels its uncertainty, cites what it leans on, and writes real code at length. Every claim ships with an auditable receipt. When it cannot answer, it says so plainly instead of inventing one.

P0 is not another wrapper. It is built around honesty about what a model does and does not know, and it keeps your key and your work on your device.

• Bring your own keyYou supply your own model provider key. It is stored on-device in the iOS Keychain — never bundled, never hardcoded, never sent anywhere except the provider endpoint you chose. OpenAI, any OpenAI-compatible gateway, Anthropic natively (Messages API — Claude models such as claude-fable-5), or local models (Ollama, llama.cpp) on your own network.
• No phone-homeNo Everplay-Tech analytics, telemetry, or account server. Prompts, keys, and dossiers stay between your device and your provider.
• Honest receiptsEvery reasoning step is captured: its phase, the model used, token counts, and — for research claims — entailment chains labeled correct or plausible plus the citations they rest on. A multi-pass adversarial audit pressure-tests every claim, backed by machine-checked truth tables. Auditable, not a black box.
• Honest empty / error statesBefore a key is connected, the app says so plainly. If a phase cannot complete, P0 surfaces the real error rather than fabricating an answer.
• Offline reasoning floorA deterministic on-device Rust core is built in — core logic and structure with no cloud key and no download. A fuller on-device model (Phi-4-mini, ~2.5 GB GGUF) is strictly optional: a resumable background download you start yourself, or a file you import — never bundled, never required.

A "research paper" in DUD3P0 is a Mission. You build the parameters first, then reasoning begins. Setup is a seven-step wizard:

• 1 · QuestionGive the mission a title, the research question that anchors the pipeline, and any key claims you already want examined.
• 2 · Domain & RigorPick a domain (Software, Mathematics, Systems, Security, Science, History, Philosophy, Markets, Policy, Open Inquiry) and a rigor posture (Exploratory → Balanced → Skeptical → Adversarial → Formal → Forensic, and more). Rigor sets how hard P0 argues against itself.
• 3 · SourcesAdd source profiles, each tagged with a role — Primary, Corroborating, Adversarial, Expert, Dataset, Artifact, Witness — plus its central claim, known bias, and contradictions.
• 4 · ContextDescribe the background that grounds the work: what is established, the stakeholders, the terminology, the open contradictions.
• 5 · DepthChoose a target depth — Brief (~1k words), Survey (~2.5k), Deep Dive (~7.5k), Monograph (~15k) — and a section count (1–30, with a recommended default).
• 6 · ProcessPer-phase model routing. Each reasoning phase can inherit or override its model and base URL. Every phase authenticates with your single on-device Keychain key.
• 7 · Review & GenerateConfirm the parameters, then press Generate to push the mission through the pipeline.

Once you press Generate, the mission moves through six reasoning phases and produces a receipt-backed dossier instead of a one-shot answer:

• FramingScope the question, constraints, and what counts as an answer.
• DecompositionSplit the question into atomic, testable sub-questions.
• InvestigationGather evidence, receipts, and citations per sub-question; profile the sources.
• DraftingWrite the dossier sections strictly from the gathered evidence.
• VerificationAudit every claim, label uncertainty, flag missing receipts.
• ConvergenceConverge all verified material into one coherent, receipt-backed dossier.

Every mission keeps a Process Log — the audit trail. Each captured generation step is stamped with its reasoning phase, the pipeline status (Building Evidence, Profiling Sources, Investigating, Drafting, Verifying, Converging), a timestamp, the model used, and token usage.

Unlike the server-required siblings, P0 needs nothing but your key — and even without a key it has an offline reasoning floor.

1. Get DUD3. Free on the App Store.
2. Open Settings → Connection.
3. Choose your provider. OpenAI, Anthropic (native Messages API), an OpenAI-compatible gateway (set the base URL), or a local endpoint like Ollama / llama.cpp on the same network as your phone.
4. Paste your provider API key. It is written to the iOS Keychain and never leaves the device except as the Authorization header to the endpoint you chose.
5. Start a chat or a research mission. Until a key is connected, the app tells you plainly and will not fabricate answers.

DUD3 for Mac is a free, notarized installer for your Mac. It installs DUD3's program, starts it in the background, and connects it to the AI apps it finds: Claude Code, Claude Desktop, Cursor, VS Code and other MCP clients. They can then call DUD3's tools on your own Mac: Reality Lens, truth tables, logic verification, receipts, adversarial and premortem checks, repository reads, diff drafting, reviewed diff apply, approved test runs, research missions and the phone's mission offload.

DUD3P0 is a tool for research, reasoning, and code. It is not designed or marketed for generating sexual, violent, hateful, or otherwise objectionable content — that use is prohibited by the Acceptable Use Policy.

This page is the human face. The same content is published as canonical JSON — the machine face. Hand the URL to your own LLM or MCP client (HELM ships it by default) and it can onboard you, with nothing phoning home.

Discord is where P0 users compare research-mission setups, swap rigor and source-profiling patterns, and ask the maintainer about provider routing and the offline floor. The manual bot answers from this same JSON manual.

Direct contact until the invite link lands: info@llmtech.net.

• Machine manual (JSON)The canonical dump. Hand it to your agent; HELM ships it by default.