Skip to main content
The @sovseal/mcp-server package runs a local Model Context Protocol server over stdio, giving AI clients like Claude Desktop, Claude Code, Cursor, Windsurf, and Antigravity persistent zero-knowledge memory with 0 RTT local latency on every read or write. The embedding model (Xenova/all-MiniLM-L6-v2, SHA-256-pinned) and the LanceDB vector index live entirely on your device — no plaintext ever reaches a remote server.

​
Stdio transport configuration

Register the server with your AI client by adding a single JSON block to its configuration file. The server launches on demand via npx and communicates over stdin/stdout.
claude_desktop_config.json
Place this block in the appropriate config file for your client:

​
Key custody

sovseal uses a layered key custody model so that your encryption master key is never stored in a plaintext configuration file:
  • OS keychain: The master key is held in the operating system keychain (@napi-rs/keyring: macOS Keychain, GNOME Secret Service, or Windows Credential Manager).
  • HKDF subkeys: Derived per purpose — k_rest for local LanceDB at-rest encryption and k_sync for replication.
  • Headless hosts: Set SOVSEAL_KEY_FALLBACK=file to use 0600 file permissions fallback when an OS keychain is unavailable.

​
Tool reference

​
store_memory

Masks high-risk PII (SSN, Luhn-validated credit cards, API keys) → embeds (384-dim) on cleaned text → encrypts text at rest (k_rest) → writes to local LanceDB index. Exact duplicate facts increment reinforce_count rather than inserting a duplicate row. Parameters
​
string
required
A self-contained, third-person factual statement to store (e.g., "User prefers TypeScript and Vite for dashboard apps").
Example MCP response

​
recall_memory

Embeds query → over-fetches 8×K candidates by vector distance → re-ranks candidates by composite score: score=similarity×decay×reinforcement\text{score} = \text{similarity} \times \text{decay} \times \text{reinforcement} Returns top-K {id, text, score} with 0 network round-trips. Parameters
​
string
required
Natural-language search string (e.g., "frontend stack preferences").
​
integer
Number of results to return (1–20, default 5).

​
sovseal CLI

The package also installs a sovseal binary, so the same build that serves MCP tools gives you terminal access to the identical vault — memory commands, health diagnostics, and device linking. See the CLI reference for the full surface.
The most-used subcommand is mind, a one-shot token-budgeted digest (~1,200 tokens) of your local memory store that never starts the stdio server:
Prints two sections:
  1. Core Mind State — procedural & semantic memories ranked by composite score.
  2. Recurring Patterns — memories reinforced more than once (reinforce_count > 1).
For scripts and agents, every command accepts --json and emits a structured envelope; sovseal help --json publishes the entire command tree as machine-readable data.

​
MCP Resources