PARANOID

PARANOID MCP

Ask AI models in private from Claude Code, Cursor or any MCP client. Every question carries a fresh zero-knowledge proof made on your own computer, so the PARANOID gateway can check that it was paid for without learning which member asked or tying it to any other question. To keep your network address apart as well, see What stays visible.

PARANOID runs open models such as DeepSeek, Kimi, Qwen and GLM, including uncensored ones. Questions are bought once with USDC on Solana, and agents can buy them on their own with x402. The Solana program and the circuit behind every proof are open at paranoiddotred/paranoid, and that source builds to the exact bytes deployed on mainnet.

It is listed in the official MCP Registry as io.github.paranoiddotred/paranoid-mcp, published straight from each release by CI, and on Glama.

Install

The quick way

Claude Desktop. Download paranoid.mcpb from the latest release and double-click it, or drag it into the Claude window. Paste the connector code from the Keys tab on paranoid.red into its settings, or choose a Solana wallet file holding USDC so the buy tool can get questions. Claude Desktop brings its own Node.js, so there is nothing else to install.

Cursor.

Claude Code.

claude mcp add paranoid -- npx -y github:paranoiddotred/paranoid-mcp#v1.4.1 mcp

Cursor and Claude Code run it straight from this repository with npx, pinned to the release tag so you get the version whose hashes are published, and that needs Node.js and git. To use questions you already bought, run it once in a terminal and paste the connector code when it asks, so the code never lands in your shell history. It then serves the local endpoint until you stop it with Ctrl+C.

npx -y github:paranoiddotred/paranoid-mcp#v1.4.1

You can also put the code in the client's settings as PARANOID_CODE, the way the Claude Desktop extension does.

By hand

It is one file with no packages to install and needs Node.js 20.18, 22.8 or newer. Download it and check it against the current release in one go:

mkdir -p ~/paranoid && curl -o ~/paranoid/connect.mjs https://paranoid.red/connect.mjs \

&& echo "8f90ee28b74883a46dcf928efaecfcd452591753851283210edf2788cbdd73c1 $HOME/paranoid/connect.mjs" | shasum -a 256 -c

The file paranoid.red serves is byte for byte the connect.mjs in this repository, and every release lists its SHA-256, so a copy that was changed anywhere on the way fails that check before it ever runs.

You can also build it yourself from src/ and the prover and circuit files where they are published (snarkjs 0.7.6 on npm, the circuit in paranoiddotred/paranoid). CI does the same on every push and fails if a single byte differs.

bash scripts/fetch-zk.sh && node build.mjs && git diff --exit-code connect.mjs

Claude Code

claude mcp add paranoid -e PARANOID_PAYER=$HOME/agent-wallet.json -- node $HOME/paranoid/connect.mjs mcp

Cursor (~/.cursor/mcp.json)

{

"mcpServers": { "paranoid": { "command": "node", "args": ["/Users/you/paranoid/connect.mjs", "mcp"], "env": { "PARANOID_PAYER": "/Users/you/agent-wallet.json" } } } }

PARANOID_PAYER points at a Solana keypair file (the JSON the Solana CLI writes) whose wallet holds USDC. It is needed only to buy questions. If you already have questions on paranoid.red, run node connect.mjs once and paste the connector code from the Keys tab instead.

Tools

Questions cost 0.002 USDC each, 250 at least, plus a 0.30 USDC setup fee when an agent buys with x402. The gateway pays the Solana network fees.

How it stays private

When you buy questions, your computer makes a secret and sends only a commitment to it, which goes into a member list on Solana. For each question, the connector proves in zero knowledge that it holds the secret of some member with questions left, and marks one numbered question as spent with a nullifier that names no member. The gateway verifies the proof and forwards the question to the model, but it never learns which member asked, two questions cannot be tied together through their proofs, and each request travels on a connection of its own.

The gateway runs inside an Intel TDX chip. Before the first question, and every ten minutes after, the connector checks Intel's signature on that chip, a nonce it made itself, and the chip's build against the list of approved builds that PARANOID's 2-of-3 multisig keeps on Solana (p2UnUp...Sfw8), read from a public Solana RPC that PARANOID does not run. Then it seals every request to that chip with Oblivious HTTP and sends it through an independent relay. Nothing on the way, paranoid.red included, can read the question or the answer, and the gateway sees the relay's address instead of yours. If a check fails, nothing is sent. node connect.mjs check runs the checks and one sealed request without spending a question.

Spending the same numbered question twice reveals the secret. Anyone can then slash the membership and recompute every question it ever asked, so keep each membership in one place. The connector writes each number to disk before it is used and reads the public list of spent questions before every key, so this cannot happen by accident.

The design follows zkAPI by Vitalik Buterin and Davide Crapis, from the Ethereum Foundation and the Open Anonymity Project. PARANOID brings it to Solana.

Use a PARANOID key in any OpenAI app

The Keys tab on paranoid.red gives a key that starts with sk-paranoid-. It works like an OpenAI API key, so any app or SDK that lets you set the OpenAI base URL can run on PARANOID with two settings.

Base URL  https://paranoid.red/api/v1

API key sk-paranoid-...

With the OpenAI SDK for Node.js:

import OpenAI from "openai";

const client = new OpenAI({ baseURL: "https://paranoid.red/api/v1", apiKey: process.env.PARANOID_KEY }); const res = await client.chat.completions.create({ model: "default", messages: [{ role: "user", content: "Explain zero-knowledge proofs in two lines." }], }); console.log(res.choices[0].message.content);

With the OpenAI SDK for Python:

import os

from openai import OpenAI

client = OpenAI(base_url="https://paranoid.red/api/v1", api_key=os.environ["PARANOID_KEY"]) res = client.chat.completions.create(model="default", messages=[{"role": "user", "content": "Hello"}]) print(res.choices[0].message.content)

With curl:

curl https://paranoid.red/api/v1/chat/completions \

-H "Authorization: Bearer $PARANOID_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"default","messages":[{"role":"user","content":"Hello"}]}'

GET https://paranoid.red/api/v1/models lists the models and what one message to each costs. A model name the gateway does not know goes to the default model, so apps that insist on a name of their own still work.

The key is a bundle of proofs made on your device and stops working when the day ends. For an app or agent that stays on, run the connector instead: it makes a fresh key for every request, so nothing ever expires.

Other ways to run it

The same file serves the OpenAI API on your computer for any app that speaks it.

node connect.mjs                 # serves http://127.0.0.1:4141/v1, any API key works

PARANOID_PAYER=wallet.json node connect.mjs buy 500

Full documentation is at paranoid.red/docs/agents and paranoid.red/docs/connector.

From the terminal

node connect.mjs ask "What does RFC 9458 change for a relay?"

git diff | node connect.mjs ask "Review this change" node connect.mjs chat -m

The answer goes to stdout and everything else to stderr, so it can be piped on.

Files it downloads

On its first question the connector downloads the prover and the circuit from paranoid.red and keeps them in ~/.paranoid/zk. With them comes Intel's quote verifier, @phala/dcap-qvl-web 0.3.3 exactly as published on npm. The prover is snarkjs 0.7.6 exactly as published on npm, and the circuit files are the ones published in paranoiddotred/paranoid, so anyone can compare them. Their SHA-256 hashes are written inside connect.mjs, it refuses to run anything that does not match, and the prover runs in a separate JavaScript realm with no network, files or process to reach, so the proof is the only thing that leaves it. SECURITY.md has the details.

License

MIT