The trend of wrapping LLMs in heavy server-side abstraction is getting a hard reset on mobile. Pathfinder, a newly released native Android research agent, cuts out the middleman entirely. Instead of routing traffic through a proprietary backend, the app talks directly to the model providers you configure. It is a chat app that lets a language model search the web and read pages to answer, all while keeping the architecture lean and the data on the device.

Direct-to-Provider Architecture

Pathfinder supports sign-in via API keys or existing subscriptions for major players including Anthropic (Claude Pro/Max), OpenAI (ChatGPT Plus/Pro), GitHub Copilot, Kimi, xAI (Grok/X Premium), and OpenRouter. The architecture allows users to switch models mid-conversation, a feature often locked behind complex server-side session management in other apps. There is no Pathfinder server in between; the client handles the heavy lifting of orchestration.

Local Session Management and Tools

Sessions are stored on-device as plain JSONL, giving users total ownership of their conversation history. The app supports branching any turn into a new line of conversation, along with searching across sessions and exporting them. For research capabilities, it utilizes web_search via the Brave Search API and web_fetch, which renders pages and returns readable content as markdown. This setup requires Android 14 or newer and is distributed as signed APKs via GitHub Releases.

Under the Hood: A Kotlin Port

The technical foundation is a selective Kotlin port of piโ€™s agent runtime, mirroring its package layout for messages, streaming, tool calls, and OAuth. The Android layer is built with Jetpack Compose and Material 3. Notably, the project is GPL-3.0 licensed, signaling a commitment to open-source transparency. Privacy is strict: no analytics, crash reporting, or telemetry. Credentials are stored encrypted behind the Android Keystore, ensuring they never leave the device except to authenticate with their specific provider.

Key Takeaways

  • Pathfinder eliminates server-side middlemen, connecting directly to LLM providers.
  • Session data is stored locally as JSONL, enabling branching and easy export.
  • The app requires Android 14+ and is GPL-3.0 licensed.
  • Supports major providers like Anthropic, OpenAI, and xAI via API keys or subscriptions.

The Bottom Line

This is exactly the kind of infrastructure we need: raw, local-first, and free from the bloat of serverless wrappers that just want to harvest your data. Pathfinder proves that you don't need a cloud backend to run a sophisticated agent; you just need a good runtime and respect for the user's hardware.