We live in a fundamental security dilemma. Modern AI models (LLMs) have become critical infrastructure for knowledge work. But the price for this intelligence is steep: you have to send your most intimate thoughts, trade secrets, and ideas to centralized "black boxes" like OpenAI, Google, or Anthropic.
For cypherpunks, lawyers, and freedom-loving aliens, that's unacceptable. The industry's answer so far has been: "Just trust us." Maple AI (trymaple.ai) aims to break this paradigm. Not with promises, but with cryptographic proofs. This report analyzes the technology, the risks, and why this project is more than just another AI wrapper.
1. Origins: From the Ashes of Mutiny Wallet
Maple AI is not a classic Silicon Valley startup jumping on the hype train. It is the direct successor of the team behind Mutiny Wallet — an innovative, self-custody Bitcoin Lightning wallet.
When the team had to shut down the wallet in 2024/2025 due to regulatory hurdles, it made a strategic pivot. The ideological foundation remained identical: data is toxic waste. Whoever collects data creates liability risk and surveillance infrastructure. The Human Rights Foundation (HRF) already listed the project as one of the "Top 15 Freedom Tech Projects of 2025".
2. The Architecture: OpenSecret & AWS Nitro Enclaves
How can Maple claim not to see your data when it's being processed on their servers? The answer lies in Confidential Computing. At its core, Maple is an implementation of the backend platform "OpenSecret".
The Vault: AWS Nitro Enclaves
Normally a server administrator (root) has access to everything happening in memory. Maple uses AWS Nitro Enclaves instead — isolated compute environments:
- No SSH access: Even Maple's own developers cannot log into the enclave.
- No disk storage: The enclave is diskless. Once power is cut or the process ends, the data is gone (zero data retention).
- No open internet: The enclave communicates only through a strictly defined local channel.
The Proof: Remote Attestation
This is the crucial point ("don't trust, verify"): before your device sends a single byte of data, it requests an "Attestation Report" — a cryptographic signature directly from the AWS hardware.
This report confirms: "I am a real Nitro Enclave and I am running exactly the program code whose hash you expect." Since the code is open source (reproducible builds via NixOS), anyone can verify that no backdoors have been built in.
How Live Web Search (Brave) Works
A common problem with secure enclaves is that they cannot have internet access. Maple solves this with a "privacy proxy":
- Your encrypted prompt lands in the enclave.
- The enclave decrypts it, detects a search query, and sends it via a proxy to Brave Search.
- Brave sees only the Maple server's IP, not your IP.
- The search results flow back into the enclave, are processed there, and sent to you encrypted.
This gives you live data without Google or Bing being able to build a profile on you.
3. The Models: The China Paradox & Vision
Maple doesn't train its own models — it hosts the best open-weight models in the world.
- DeepSeek R1 & Kimi K2 (China): Maple uses Chinese top-tier models that often beat GPT-4 in benchmarks. Normally a privacy nightmare. But: since these models run in the isolated US enclave without internet access, they can't beam data back to China. Using the enemy's weapon against itself.
- GPT-OSS-120B (USA): A rare open model from OpenAI with strong reasoning capabilities.
- Gemma 3 (Vision): Image recognition is also available. You can analyze screenshots or photos — also fully processed inside the enclave.
4. Maple Proxy: AI for Developers Without Surveillance
For developers, the "Maple Proxy" feature is a game changer. It's a small program that runs locally on your machine and pretends to be the OpenAI server.
You can use tools like Cursor, VS Code, or Windsurf and simply redirect them to your local port. The proxy encrypts everything and sends it into the enclave. That way you can use proprietary coding tools without your code landing at Microsoft or OpenAI.
5. Pricing & True Anonymity
The business model is transparent: Maple funds itself through direct user payments. The pricing structure is similar to competitors, but offers real "de-banking" resistance through Bitcoin.
| Plan | Price | Target & Limit |
|---|---|---|
| Free | $0 | For testing. Limited to ~25 messages/week. |
| Starter | $5.99 | Basic models, moderate usage. Entry point. |
| Pro | $20 | Standard. Access to DeepSeek R1, Kimi K2 & GPT-OSS. |
| Max | $100 | Power users & developers (20x limits). Ideal for Proxy. |
| Team | $30 / user | For law firms/practices. Centralized billing. |
Pay with Bitcoin & Lightning
As heir to Mutiny Wallet, Maple integrates native Bitcoin payments.
- Anonymity: You can pay without a bank account or real name (no PII required).
- Lightning Network: Instant, low-fee payments.
- Discount: Paying with Bitcoin sometimes gets you a discount.
6. Critical Analysis: Where Are the Catches?
Where there's light, there's shadow. Security costs convenience.
- Latency (Ping): The cryptographic handshake takes time. Maple feels less "snappy" than ChatGPT.
- Loss risk: Lose your recovery phrase and the account is gone. There is no support reset.
- Traffic analysis (metadata): This is a risk for high-value targets (dissidents). Even if the content is encrypted, an ISP could infer from packet size and timing that you're actively communicating with an AI. Maple doesn't (yet) offer obfuscation for this.
- Sync limits: Synchronization between devices happens via encrypted "blobs." The server doesn't see the content, but it does see when and how much is being synced (metadata leak).
7. Alien Verdict: Niche or Future?
For the average user, Maple AI is overkill. But for lawyers, doctors, developers, and anyone handling sensitive data, it's potentially without alternative. It's proof that "Confidential Computing" has left the lab.