Install both tools in 5 minutes

1. Install the vault as a Claude Code skill

One cp from the cloned repo into your skills directory.

git clone https://github.com/froggychips/mcp-skills-vault.git
mkdir -p ~/.claude/skills
cp -r mcp-skills-vault/mcp-ecosystem-intelligence ~/.claude/skills/

2. Scan your current project

The orchestrator detects your stack from manifests and prints what to install. No network needed for the offline gate.

node mcp-skills-vault/mcp-ecosystem-intelligence/scripts/orchestrate.cjs \
  --cwd .

3. Install mcp-trace

Same shape: clone, verify version. Zero npm deps.

git clone https://github.com/froggychips/mcp-trace.git
cd mcp-trace
node bin/mcp-trace.cjs --version

4. Wrap an MCP server with trace

Edit your project’s .mcp.json (or ~/.claude.json). Point command at mcp-trace, with the real server after --.

{
  "mcpServers": {
    "github": {
      "command": "node",
      "args": [
        "/abs/path/to/mcp-trace/bin/mcp-trace.cjs",
        "proxy", "github", "--",
        "docker", "run", "-i", "--rm",
        "-e", "GITHUB_PERSONAL_ACCESS_TOKEN",
        "ghcr.io/github/github-mcp-server@sha256:..."
      ],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "${GITHUB_TOKEN}" }
    }
  }
}

Restart Claude Code so it picks up the new command.

5. After a session, look at what happened

The SQLite file at ~/.mcp-trace/traces.db has every JSON-RPC frame’s metadata. Three subcommands query it.

node bin/mcp-trace.cjs stats
node bin/mcp-trace.cjs tools github
node bin/mcp-trace.cjs tokens github

You now know which tools your servers actually use, their p50/p95 latency, error rates, and a rough token-cost estimate.

6. Optional: audit your installed setup

Diff what’s actually installed against the vault DB. Flags drift, untrusted entries, unbounded heavy servers, and scope issues.

node ~/.claude/skills/mcp-ecosystem-intelligence/scripts/audit_setup.cjs