Muse Code

Memory for Muse Code, Meta’s terminal coding agent. Muse Code’s hooks cannot inject context back to the model, so recall is agent-invoked through MCP or skills, while hooks handle automatic capture and doc sync.

Extension point	Role
MCP (`mcp_servers`)	recall (`hydradb_query`) + ingest
Skills (`.claude/skills/`)	`/hydradb …` slash commands (manual recall / remember / sync)
Hooks (`.muse/hooks.json`)	automatic capture (`Stop`) + doc sync (`SessionStart`, `PostToolUse`)

Add the MCP server

{
  "mcp_servers": {
    "hydradb": {
      "transport": "streamable_http",
      "url": "https://mcp.hydradb.com",
      "headers": {
        "Authorization": "Bearer ${HYDRADB_API_KEY}",
        "X-HydraDB-Database": "your-database"
      },
      "enabled": true,
      "mode": "optional"
    }
  }
}

API-key requests to the MCP server require the X-HydraDB-Database header (your database ID). See the MCP reference for all supported headers.

Set credentials & run with the required flags

export HYDRADB_API_KEY="your-api-key"
export HYDRADB_DATABASE="your-database"

# --trust-workspace loads project skills + hooks; --yolo (or an approval
# profile) lets the auto-recall skill run node; both are required or recall
# silently no-ops in headless mode.
muse exec --trust-workspace --yolo "what did we decide about X?"

Only use --yolo in a workspace you trust. It bypasses command approval, so any project hooks loaded by --trust-workspace run against your local environment. In untrusted checkouts, prefer an approval profile that permits only the recall skill’s node call instead of --yolo.

It runs on the same engine as the Claude Code plugin, so the modes and config variables there apply here too.

Feedback and contributions

This plugin is open source. Feel free to raise an issue or contribute on GitHub.

Get Started

Usage

Organize

Understand

Integrations

Plugins

For Agents

Feedback and contributions

​Feedback and contributions

Feedback and contributions