Your Self-Sovereign
AI Agent

Pathfinder is a native desktop application and personal AI agent built on a single architectural guarantee: by default, nothing leaves your machine. It is not a cloud service with a privacy policy. It is a locally-run system whose architecture is structurally incapable of transmitting your data without your explicit choice.

This white paper presents the technical architecture, design philosophy, capability set, and roadmap for Pathfinder — and makes the broader case for why personal AI must be sovereign by design, not sovereign by promise.

Pathfinder is open-source under AGPL v3. It runs on Windows, macOS, and Linux. It supports five LLM providers — local and cloud — switchable per persona. It includes an encrypted vault, a Telegram gateway, a cron scheduler, an agent loop with tool calling, a knowledge base, a credentials manager, and a skills pipeline. All of it runs on your hardware. All of it answers to you.

Every major AI assistant in use today operates on the same implicit contract: hand over your data, your conversations, your habits, and your identity — and in return, receive intelligence. The provider promises to be responsible with what you share. They publish privacy policies. They hire ethics teams. They mean well.

But "trust us" is a policy. And policies change.

Policies are rewritten when business models shift. They are overridden by government requests. They are rendered meaningless by acquisitions, bankruptcies, and data breaches. A policy is a promise made by people who may not be there tomorrow, about data that already left your hands yesterday.

We are at an inflection point. AI is moving from novelty to infrastructure — from something you try to something you depend on. The decisions made in the next few years about how personal AI is architected will shape who owns the most intimate layer of human cognition for decades.

Your AI assistant learns your schedule, your relationships, your health concerns, your finances, your fears. That is an extraordinary amount of trust to place in a terms-of-service agreement.

Pathfinder offers an alternative: personal AI that is genuinely personal. Intelligence that serves its owner because it is built to — not because it promised to.

Pathfinder is built on a single foundational principle: a system that is architecturally incapable of acting against its owner cannot betray them — regardless of who builds it, funds it, or acquires it.

This is not a privacy policy. It is the structure of the system itself.

Pathfinder is a native desktop application built with Tauri 2 and a Rust backend. It is not a web app. It is not an Electron app. It is a single compiled binary with a bundled SQLite database and an AES-256-GCM encrypted vault — all running on your hardware, with no external server required.

The Rust backend enforces all security constraints at the command layer before any data is processed. The frontend is a pure consumer of Tauri commands — it never accesses the vault or database directly. This separation means tier enforcement cannot be bypassed by modifying frontend code.

The SQLite database uses WAL (Write-Ahead Logging) mode with rusqlite 0.37 bundled — no system dependency required. The vault is a key-value store encrypted with AES-256-GCM using a key derived from the user's passphrase via PBKDF2/SHA-256 at 100,000 iterations. The key is cached in RAM as a VaultSession and zeroized from memory on drop — it is never written to disk.

Pathfinder uses a four-tier security model. Each tier adds capability — none remove sovereignty. Tiers are per-persona, not per-installation. A single Pathfinder installation can simultaneously run a Scout persona (local LLM only) and a Pioneer persona (cloud LLM enabled), each with its own model, memory, and security constraints.

The Scout tier block is the clearest example of sovereignty by architecture. When a persona is set to Settler or Scout (tiers 0 or 1), the agent_chat Rust command checks security_tier <= 1 before routing to any cloud provider and returns an error — not a warning, an architectural rejection. This check runs in compiled Rust — it is not a frontend setting that can be toggled or bypassed. A Scout persona will never send inference to Anthropic, OpenAI, or Google, regardless of what API keys are stored in the vault.

A single Pathfinder installation can run a Scout persona (local-only) and a Pioneer persona (cloud-enabled) simultaneously — each with its own model, system prompt, and capability set. Tiers are per-persona, not per-installation.

The first-run wizard performs an environment scan — pinging Ollama, listing installed models, checking RAM and disk — and auto-recommends the appropriate starting tier based on what is already installed on the machine.

The sovereignty claim is architectural, not absolute. Here is exactly what stays local and what travels, when you choose to let it:

Pioneer tier personas never appear in the "Travels When" column for conversation context — cloud LLMs are blocked by the Rust command layer before the call is ever made.

Pathfinder is designed for community extensibility from the start. Five defined extension points allow contributors to add capabilities without touching the sovereignty core — the vault encryption, tier enforcement, and core DB schema remain founder-controlled.

Community contributions are gated on Block 12 — which requires a cryptographic contribution ledger (blockchain-based) to be production-ready. This is intentional: the contribution process itself must be sovereign. Every merge will require a CLA and an on-chain record. No contribution can reach users without explicit founder approval until the DAO governance model is operational.

The community layer sits above individual Pathfinder instances and can never reach inside them. It pushes capabilities downward. It never pulls data upward.

Cloud AI tools — Claude.ai, ChatGPT, Gemini — are excellent at inference. They are not designed for sovereignty. The comparison below is not a criticism of their models; it is an observation about their architecture.

Alongside Pathfinder as a product, OpenNextFrontier is developing an AI Literacy Initiative — a locally-rooted educational program that teaches practical AI skills to businesses, nonprofits, and individuals, with a built-in access model that ensures those who can't afford tuition still get a seat.

The goal is not to produce AI power users. It is to produce people who understand what AI is, what it is not, where it is heading, and what questions to ask. People who can evaluate claims made by vendors, employers, and politicians. People who know when to embrace a technology and when to push back.

What students take home: A working AI assistant that runs locally on their own computer — no subscription, no ongoing cost, no vendor dependency. Something real, not just a certificate.

The model: Businesses and professionals who can afford quality AI education fund seats for nonprofit staff, community organizations, students, and individuals who couldn't otherwise participate. Paying students aren't just buying education — they're enabling access for others.

Governance vision: A foundation arm (nonprofit, eventual 501(c)(3)) holds the educational mission and access fund. A commercial arm sustains operations and generates the cross-subsidy. Community governance expands over time to include educators, builders, and participants.

Personal AI is becoming infrastructure. The architecture decisions being made today will determine who owns the most intimate layer of human cognition for the next generation.

Pathfinder is a bet on a different outcome: that it is possible to build personal AI that is genuinely personal. That sovereignty by design is not a niche concern but an architectural necessity. That the most trustworthy AI is not the one that promises best — it is the one that cannot betray you by structure.

The project is open-source. The architecture is visible. The vault is yours. The data is yours. The personas are yours. The memories are yours.

That is not a claim. That is a design.

Get started: Download Pathfinder from opennextfrontier.com. Single .exe installer. No account. No subscription. Your data never leaves.

Read the source: github.com/epictetusmemory-design/pathfinder-tauri — AGPL v3.

Explore the docs: opennextfrontier.com — architecture, wire diagram, layer map, capabilities, this white paper.