Skip to content
Back to Wyrm

Wyrm Docs

v7.0.232-verb typed surface · 150 tools · run-attributed fleet memory · fleet negative learning · Live Memory · hybrid recall · OODA loop · wyrm-loop daemon

Everything you need to install, configure, and use Wyrm — persistent memory for AI agents.

// INSTALLATION

Install Wyrm

npm (global)
$ npm install -g wyrm-mcp
npx (no install)
$ npx wyrm-mcp
Homebrew (macOS / Linux)
$ brew tap ghosts-lk/wyrm
$ brew install wyrm-mcp
Docker
$ docker run -d -p 3000:3000 \
    -v wyrm-data:/data \
    ghcr.io/ghosts-lk/wyrm-mcp:latest
// CLIENT CONFIGURATION

Configure Your AI Client

The MCP config is identical for every client — only the config file path differs.

mcp config (all clients)
{
  "mcpServers": {
    "wyrm": {
      "command": "wyrm-mcp"
    }
  }
}
Claude Desktop
~/Library/Application Support/Claude/claude_desktop_config.json(macOS)
Claude Desktop
%APPDATA%\Claude\claude_desktop_config.json(Windows)
GitHub Copilot
.vscode/settings.json(key: "github.copilot.chat.mcpServers")
Cursor
.cursor/mcp.json(project root)
Windsurf
~/.codeium/windsurf/mcp_config.json(global)
// TOOLS REFERENCE

The Tool Surface

A frozen, typed 32-verb surface advertised to clients, with all 150 tools still callable. Organized by category — every tool is available in the Free tier.

Projects— 6 tools
1 / 24
wyrm_scan_projectsScan a directory for git projects and register them. Use this to auto-discover all projects.
wyrm_list_projectsList all registered projects with their status
wyrm_project_contextGet full context for a specific project including recent sessions, quests, and stored context
wyrm_global_contextGet overview of all projects, pending quests across projects, and global context
wyrm_set_globalSet global context that applies across all projects
wyrm_constellationCross-project memory query — search every registered Wyrm project at once for matching truths, artifacts, quests, or design references. Returns candidates grouped by project; the calling AI does semantic ranking on its side. Use this when the operator asks 'have I solved this before?', 'where else have I used X?', or any portfolio-spanning question. Honours per-row cross_project_visibility ('within' = invisible across projects, 'org' = shared, 'public' = anywhere).
// CONFIGURATION

Environment Variables

VariableDescription
WYRM_DB_PATHPath to the SQLite database file
WYRM_ENCRYPTION_KEYAES-256 encryption key for data at rest. Set to enable encryption.
WYRM_SYNC_PASSPHRASEPassphrase for encrypted snapshot export/import. Set to skip interactive prompt.
WYRM_VECTOR_PROVIDERVector embedding provider for semantic search. 'auto' detects Ollama + nomic-embed-text automatically. Supports 'auto', 'ollama', 'openai', 'none'.
WYRM_LOG_LEVELLogging verbosity: debug, info, warn, error
WYRM_CACHE_SIZEMaximum entries in the in-memory read cache
WYRM_PORTPort for the HTTP API server (wyrm serve). The visual dashboard is accessible at http://localhost:WYRM_PORT/ui when --ui is passed.
WYRM_ALLOWED_PATHSColon-separated list of additional project root directories Wyrm is allowed to scan. Extends the default allowlist (~/, ~/Projects, ~/dev, ~/code, ~/repos).
WYRM_DEVEnable development mode — disables auth enforcement for localhost connections. Never set in production.
Example with encryption
WYRM_DB_PATH=~/.wyrm/wyrm.db \
WYRM_ENCRYPTION_KEY=your-32-byte-hex-key \
WYRM_LOG_LEVEL=debug \
wyrm-mcp
// ARCHITECTURE

How Wyrm Works

Wyrm is built as a Model Context Protocol (MCP) server that any compatible AI client can connect to. It uses a local SQLite database as its storage engine, enhanced with FTS5 for full-text search.

SQLite with WAL mode for concurrent read/write access
FTS5 full-text search engine for instant queries
In-memory cache layer with auto-invalidation on writes
Optional AES-256-GCM encryption with per-install random salt
Hybrid search: Ollama auto-detected for semantic vectors + FTS5 keyword (RRF fusion)
MCP protocol with cache_control hints for prompt caching
TypeScript with full type safety, zero external runtime deps
Security hardened: rate limiting, input validation, path traversal guards, no error leakage
Architecture

AI Client

Claude · Copilot · Cursor

MCP Protocol

Wyrm MCP Server

MemCache Layer

sub-ms reads

Encryption Layer

AES-256-GCM

SQLite + FTS5

WAL mode

// FAQ

Frequently Asked Questions

Ready to Get Started?

Install Wyrm in under 60 seconds and give your AI persistent memory.

Get WyrmBack to Wyrm