Maple AI

Maple AI is a private chat assistant built by OpenSecret on AWS Nitro Enclaves, with end-to-end encryption from the client device to the inference enclave. Conversations are encrypted client-side, sent through an encrypted channel into a hardware-isolated enclave (CPU-isolated, no persistent storage, no admin access, no external network), and the response is encrypted in the enclave before return. The architecture combines Nitro Enclaves with reproducible builds so that anyone can build the open-source codebase from GitHub and compare the resulting hash against the live cryptographic attestation from the enclave, verifying that the deployed binary matches the audited source.

The product currently serves DeepSeek R1 and V3.1, OpenAI's GPT-OSS-120B, Qwen3 Coder 480B, Llama 3.3 70B, and Gemma 3 / Qwen3-VL 30B (multimodal). All inference runs inside the enclave; OpenSecret does not host an unencrypted log of prompts because the operator cannot see plaintext at any stage. The free tier covers core features; paid Starter, Pro, and Team plans add higher usage limits and pooled organizational access. Native applications ship for iPhone, iPad, and Android, with web access at trymaple.ai.

The OpenSecret platform itself is open source and provides the underlying confidential-compute primitives: the developer platform is in Rust, an OpenSecret SDK is published in TypeScript, and a nitro-toolkit in Python collects utilities for Nitro Enclave development. Source is at github.com/OpenSecretCloud, with the Maple application code under github.com/OpenSecretCloud/Maple.

HRF awarded Maple AI in its November 24, 2025 AI for Individual Rights round, the foundation's first dedicated AI grant cohort. The target user is a journalist, organizer, or human rights defender in an authoritarian environment who needs frontier-quality LLM help but cannot give an external provider an unencrypted log of every prompt. Maple AI's enclave architecture is the structural defense against subpoena or coercion of the provider, because the provider physically does not have access to plaintext to surrender. The competing alternatives are running local models (capability-limited at the laptop tier) or trusting a commercial provider's privacy policy (no cryptographic guarantee).