LoopPrism

See what your agents actually do.

---

What is LoopPrism?

LoopPrism is the strace for AI agents. It sits between your agent and MCP servers, recording every tool call so you can see exactly what happened, how long it took, and where your token budget is going.

No code changes. Just point your MCP clients at loopprism.

pip install "git+https://github.com/AMark-CS/LoopPrism.git"
loopprism proxy --port 4317

Then open http://localhost:8080 and watch every MCP call appear in real time.

Features

• Waterfall traces — See every MCP tool call with timing, result size, and error state
• MCP-native — Understands tools/call, tools/list, prompts/get, and more
• Cost breakdown — Know which tool generates the most data
• Zero config — One command, SQLite on disk, no Docker, no cloud
• Local-first — All data stays on your machine

Quick Start

1. Install

pip install "git+https://github.com/AMark-CS/LoopPrism.git"
# or: uv tool install git+https://github.com/AMark-CS/LoopPrism.git

2. Start the proxy

loopprism proxy --port 4317

╭─────────────────────────────────────────────────────╮
│              loopprism                              │
│      "See what your agents actually do"            │
├─────────────────────────────────────────────────────┤
│  Proxy:     http://localhost:4317                   │
│  Dashboard: http://localhost:8080                   │
│  Database:  ~/.loopprism/traces.db                  │
│                                                     │
│  Ready. Point your MCP clients to localhost:4317   │
╰─────────────────────────────────────────────────────╯

3. Point your MCP clients to loopprism

// Claude Desktop config (~/.claude/settings.json)
{
  "mcpServers": {
    "github": {
      "url": "http://localhost:4317/github?backend=https://api.github.com/mcp"
    },
    "filesystem": {
      "url": "http://localhost:4317/fs?backend=http://localhost:3001"
    }
  }
}

Or via environment variable:

export GITHUB_MCP_URL="http://localhost:4317/github?backend=https://api.github.com/mcp"

4. Open the dashboard

Visit http://localhost:8080 — every MCP tool call appears in real time with:

• Waterfall chart showing timing and latency
• Cost breakdown by tool
• MCP server health stats

Architecture

Agent ──▶ loopprism Proxy ──▶ MCP Server
              │
              ├── Intercepts JSON-RPC
              ├── Extracts tool name, args, latency
              ├── Writes to SQLite (~/.loopprism/traces.db)
              └── Serves Dashboard (http://localhost:8080)

Commands

# Start proxy + built-in dashboard (default)
loopprism proxy

# Custom ports
loopprism proxy --port 4317 --dashboard-port 8080

# Proxy only (no dashboard)
loopprism proxy --no-dashboard

# Standalone dashboard (connect to a remote DB)
loopprism dashboard --db ~/.loopprism/remote.db

# Analyze a specific trace
loopprism analyze <trace-id>

# Debug verbose mode
loopprism proxy --verbose

API

The dashboard is backed by a REST API:

Endpoint	Description
GET /api/traces	List traces
GET /api/traces/:id	Get trace with all spans
GET /api/mcp-servers	Stats per MCP server
GET /api/mcp-tools	Stats per tool
GET /api/cost/summary	Cost overview
GET /api/cost/trace/:id	Cost breakdown per trace
GET /api/health	Health check

Development

git clone git@github.com:AMark-CS/LoopPrism.git
cd LoopPrism
pip install -e ".[dev]"
pytest

# Dashboard dev server
cd dashboard
npm install
npm run dev

Rename compatibility

LoopPrism was previously named toolglass. The Python package and primary CLI are now loopprism. The toolglass CLI alias, TOOLGLASS_* environment variables, and incoming X-Toolglass-* headers remain supported. New environment variables use LOOPPRISM_* and take precedence. Existing ~/.toolglass/traces.db and config.toml are used when their new equivalents are absent; no data is moved or deleted.

Install from GitHub for now; this repository does not yet provide a published loopprism PyPI release.

License

MIT