HydraDB for Muse Code

Persistent, cross-session memory for Muse Code (Meta's terminal coding agent), powered by HydraDB.

Muse Code exposes three extension points - MCP, skills, and hooks - and this integration uses all three over the same shared engine (scripts/plugin.mjs) as the Claude Code / Codex / Cursor / OpenCode plugins.

How recall works on Muse Code

Muse Code hooks cannot inject context back to the model (per the docs: hooks "enforce checks, format code, or block actions"). So there is no per-prompt auto-recall via hooks here - recall is agent-invoked, through MCP or skills:

Extension point	Role	HydraDB use
MCP (mcp_servers)	tools the agent may call	recall (hydradb_query) + ingest
Skills (.claude/skills/)	/-invoked slash commands	manual recall / remember / sync
Hooks (.muse/hooks.json)	run commands on lifecycle events	automatic capture + doc sync

Setup

1. MCP (recall)

Add HydraDB to your Muse settings mcp_servers block (see settings.snippet.json):

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

2. Skills (slash commands)

Muse Code auto-discovers skills from .claude/skills/ (and .codex/skills, .agents/skills). This repo ships the HydraDB skills there, so /hydradb … commands are available once the workspace is trusted. Manage with muse skills list / muse skills enable <id> --scope project.

3. Hooks (automatic capture + sync)

.muse/hooks.json wires lifecycle events to the engine:

Event	Engine action
SessionStart	sync workspace docs into HydraDB
PostToolUse (file writes)	incremental sync
Stop	capture the completed turn

Then set credentials (resolved by the shared engine):

export HYDRADB_API_KEY="your-api-key"
export HYDRADB_DATABASE="your-tenant-id"

Running (two flags are required)

Muse gates project skills/hooks behind workspace trust, and the recall skill needs permission to run node. Both are required or recall silently no-ops:

# --trust-workspace  loads the project skills + hooks (else they are skipped)
# --yolo             allows the auto-recall skill to run Bash(node *) (or use an
#                    approval profile that permits it) so recall can execute
muse exec --trust-workspace --yolo "what did we decide about X?"

Verified end to end: the agent invokes the model-invocable auto-recall skill, which queries HydraDB and answers from recalled context. Without --yolo (or an equivalent approval profile) the skill's tool call is blocked in headless mode.

To verify against your Muse build

• The exact .muse/hooks.json schema and whether hook commands run with the repo root as cwd (the commands use ./scripts/run-plugin.sh, which self-resolves the plugin root from its own location).
• The auth handshake for mcp.hydradb.com (bearer header vs. an MCP auth step).

License

Apache-2.0 - Copyright (c) 2026 HydraDB