@sovseal/mcp-server

SovSeal MCP Server

Unkillable, zero-knowledge memory for local AI agents (Claude Desktop, Cursor).

Your agent's context is encrypted on your machine with a key only you hold, persisted to a decentralized object store, and retrieved on demand. The server speaks the Model Context Protocol over stdio and exposes two tools to the LLM: save_context and load_context.

Install (Claude Desktop)

Paste this into ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows), then restart Claude Desktop.

{
  "mcpServers": {
    "sovseal": {
      "command": "npx",
      "args": ["-y", "@sovseal/mcp-server"]
    }
  }
}

Install (Cursor)

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "sovseal": {
      "command": "npx",
      "args": ["-y", "@sovseal/mcp-server"]
    }
  }
}

Requires Node.js 20+. The first invocation creates a local identity at ~/.sovseal/config.json (mode 0600) and bootstraps your project token. No signup, no API key.

Tools

save_context

save_context({ key: string, content: string }) -> { status, agent_id, sequence_number, arweave_tx_id, byte_size }

Canonicalizes content, encrypts client-side with AES-256-GCM, POSTs the envelope. The server only ever sees ciphertext.

load_context

load_context({ key: string }) -> string

Fetches the latest snapshot for that key, decrypts locally, returns plaintext to the LLM.

Keys are scoped to your project; the agent identifier sent over the wire is sha256(project_id || ":" || key).slice(0, 64) so the server never learns your raw key names.

How the identity works

On first run the server generates:

• a UUIDv4 project_id — used as your bearer token in the form sov_proj_<uuid>
• a 256-bit AES-GCM encryption key — never leaves your disk

Both are written to ~/.sovseal/config.json with file mode 0600. Schema:

{
  "schema_version": 1,
  "project_id": "…",
  "api_key": "sov_proj_…",
  "encryption_key_b64": "…",
  "endpoint": "https://ksrlmubaxzwufziwarps.supabase.co/functions/v1/v2-agent-state",
  "created_at": "…"
}

Back up this file. If you lose it, every snapshot you've ever made is permanently unreadable. There is no recovery flow — that's the point.

Configuration

Env var	Default	Effect
SOVSEAL_ENDPOINT	hosted SovSeal endpoint	override for self-hosting or testing

Self-hosting

The server is a thin client. The actual snapshot endpoint is an open-source Supabase Edge Function in supabase/functions/v2-agent-state. Deploy it to your own Supabase project, then point SOVSEAL_ENDPOINT at it.

Threat model — read this

• Confidentiality: AES-256-GCM with a 96-bit random IV per snapshot. The server cannot read your context. Anyone with your encryption_key_b64 can.
• Authentication: bearer token = your sov_proj_<uuid>. Anyone with the token can read and overwrite your snapshots. Treat it like an SSH key.
• Storage: ciphertext lands in a public Supabase Storage bucket. Object paths are sha256-derived and unguessable without your project_id, but enumeration is not the threat model — confidentiality is.
• Loss: lose ~/.sovseal/config.json → lose everything. There is no key escrow. There is no "forgot my password."

If your threat model includes a state-level adversary with your laptop, generate a passphrase-wrapped key system on top of this. The current MVP optimizes for friction-free developer install over key-escrow ceremony.

Build from source

git clone https://github.com/Inheribase/Inheribase
cd Inheribase/packages/sovseal-mcp-server
pnpm install
pnpm build       # → dist/index.js
node dist/index.js   # stdio MCP server

License

Apache-2.0.