About ncrypted
ncrypted is an MCP server and toolkit for sharing passwords, API keys, and notes as one-time encrypted secrets. It encrypts locally so the service stores ciphertext, and can deliver a requested secret to a Vercel environment variable or `.env` file without exposing the value to the model.
Key features
- 01Locally encrypted, one-time secret links
- 02AES-256-GCM encryption with optional passphrase
- 03Secret requests and environment-file delivery
- 04Optional email delivery of links
- 05CLI, SDK, MCP server, and Agent Skill
- 06Per-secret x402 USDC payments on Base
Use cases
- 01Share an API key with an agent once
- 02Request a secret from a user without model exposure
- 03Write credentials to a Vercel env var or `.env` file
- 04Send a one-time secret link by email
MCP support
The stdio MCP server provides tools for creating, reading, checking, and burning secrets, along with wallet functions. It encrypts locally by default; an optional hosted MCP endpoint is labeled as using server-side encryption.
Configuration
Remote servers need no local install. Pick your client, then copy the config or install in one click.
https://ncrypted.ai/mcpclaude mcp add --transport http seidl-family-ncrypted https://ncrypted.ai/mcp{"mcpServers":{"seidl-family-ncrypted":{"type":"http","url":"https://ncrypted.ai/mcp"}}}Open in Cursor{"servers":{"seidl-family-ncrypted":{"type":"http","url":"https://ncrypted.ai/mcp"}}}Open in VS Code{"mcpServers":{"seidl-family-ncrypted":{"type":"http","url":"https://ncrypted.ai/mcp"}}}{"mcpServers":{"seidl-family-ncrypted":{"serverUrl":"https://ncrypted.ai/mcp"}}}Run commands in a terminal; paste JSON into the client’s MCP config and restart. Servers that need sign-in will prompt on first connect.
Tools
4| Parameter | Type | Required | Description |
|---|---|---|---|
| secret | string | Yes | The sensitive value to share. Max 64 KB. |
| passphrase | string | No | Optional second factor the recipient also needs. Share it through a different channel. |
| ttl_seconds | integer | No | Expiry if never read. Default 86400 (24h), max 7 days. |
Details
From the project READMEncrypted
One-time, end-to-end encrypted secrets for AI agents.
An agent encrypts a password, API key or note locally and gets a link that opens exactly once. Share it with a person, email it, or hand it to another agent; reading destroys it. Agents can also request a secret from a user and write it straight into a Vercel env var or .env file without the model ever seeing the value.
The first 3 secrets per day are free; after that each costs 0.005 USDC over x402 on Base. Reads, requests and deliveries are free.
Works with Claude Code, Codex, Cursor, Gemini CLI, any MCP client, the Vercel AI SDK, LangChain and CrewAI (Coinbase AgentKit support is in review). Guides: https://ncrypted.ai/guides
https://ncrypted.ai/s/{id}#{key}
└─ decryption key, never sent to the server
Install
You need a dedicated, low-balance Base wallet with some USDC (no ETH needed). Export its key as NCRYPTED_WALLET_PRIVATE_KEY.
Claude Code
/plugin marketplace add Seidl-Family/ncrypted
/plugin install ncrypted@ncrypted
Any MCP client (Claude Desktop, Cursor, Codex, …)
{ "mcpServers": { "ncrypted": { "command": "npx", "args": ["-y", "@seidl-family/ncrypted-mcp"], "env": { "NCRYPTED_WALLET_PRIVATE_KEY": "0x..." } } } }
Agent Skill: npx skills add Seidl-Family/ncrypted
CLI / SDK
echo "hunter2" | npx ncrypted create --ttl 1h [--email [email protected]] [--passphrase X]
npx ncrypted read "https://ncrypted.ai/s/abc#key"
import { Ncrypted } from "ncrypted";
const nc = new Ncrypted(); // reads NCRYPTED_WALLET_PRIVATE_KEY
const { link, receiptToken, id } = await nc.create("hunter2", { ttl: 3600 });
const value = await nc.read(link); // free, destroys it
How it works
- Zero-knowledge: AES-256-GCM with a random 256-bit key, derived via HKDF (plus PBKDF2 if you set a passphrase). The server stores ciphertext only.
- Exactly once: reads use Redis
GETDEL, so concurrent readers can't both get it. Unread secrets expire after their TTL (5 min to 7 days). - Preview-safe: opening the web page doesn't consume the secret; the recipient has to click Reveal.
- Pay per secret: x402
exactscheme, USDC on Base, settled only after the secret is stored. The SDK refuses to pay more than $0.01 per request. - Email: the creator can email the link once. It passes through the server for that request and isn't stored. Use a passphrase for very sensitive values.
Payments and safety
ncrypted charges a fixed 0.005 USDC per secret created, paid over x402 on Base. Everything else is free: reading, status checks, revoking.
- Narrow scope: the only payment the plugin can make is ncrypted's own fee for
ncrypted_create. It has no tool to send arbitrary amounts, trade, swap, or pay anyone else. - Hard cap: the SDK refuses to pay more than 0.01 USDC for a single request, whatever the server asks.
- Dedicated wallet: payments come from a low-balance wallet the user creates on purpose (
npx ncrypted wallet new) and funds themselves. The plugin never asks for, reads or stores the user's main wallet, and the skill tells the agent never to ask for private keys in chat. - Visible and explicit: every charge is an explicit
ncrypted_createtool call. There are no background or recurring charges. - No charge on failure: payment settles through Coinbase's x402 facilitator only after the secret is stored.
- Zero-knowledge by default: the local MCP server encrypts on the user's machine, so the service stores ciphertext only. The optional hosted endpoint (
https://ncrypted.ai/mcp) is clearly labeled as server-side encryption.
Data and services contacted
The plugin's MCP server, the CLI and the skill contact only these services:
| Service | What is sent | When |
|---|---|---|
ncrypted API (https://ncrypted.ai, Vercel + Upstash Redis) | Ciphertext, IV, salt, passphrase flag and TTL. Never the key, passphrase or plaintext. | Creating, reading, checking status of or burning a secret |
| Coinbase x402 facilitator / Base network | A signed USDC payment authorization (payer address, amount, payee) | Creating a secret (0.005 USDC) |
Base public RPC (mainnet.base.org) | The payer wallet address | ncrypted_wallet / ncrypted wallet balance checks |
| Resend, via the ncrypted API | Recipient email address, the share link (which includes the key) and an optional note. Relayed once and not stored by ncrypted. | Only when the user asks to email a link |
Retention: ciphertext is deleted on first read, when burned, or at expiry (at most 7 days). Receipt metadata is deleted 7 days after expiry. No analytics, no cookies. Details: privacy policy.
Repo layout
| Path | What |
|---|---|
apps/web | Next.js app on Vercel: API, viewer page, llms.txt, openapi.json |
packages/sdk | ncrypted on npm: crypto, client, CLI |
packages/mcp | @seidl-family/ncrypted-mcp on npm: stdio MCP server |
skills/ncrypted | Agent Skill (SKILL.md) |
plugins/ncrypted | Claude Code plugin; .claude-plugin/marketplace.json lists it |
server.json | MCP Registry manifest (published by .github/workflows/mcp-registry.yml) |
Development
npm install
npm test # crypto unit tests
cp apps/web/.env.example apps/web/.env.local # set NCRYPTED_PAY_TO
npm run dev # in-memory store when Redis isn't configured
npm run build && npm run e2e # needs .env.e2e with a funded Base Sepolia wallet
After editing skills/ncrypted/SKILL.md, run npm run sync:skill.
Tool-selection evals: ./scripts/run-evals.sh runs evals/ (14 cases, mocked MCP tools, ~$2.50) against a temporary copy of the plugin. Re-run after changing tool descriptions or the skill, and refresh evals/mocks/ncrypted/_tools.json from tools/list when tools change.
On each MCP release, bump the pinned version in Dockerfile, plugins/ncrypted/.mcp.json, plugins/ncrypted/package.json (then npm install --package-lock-only there) and gemini-extension.json (the Claude directory rejects @latest), then run the "Publish to MCP Registry" workflow.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.