LiveTap

Push live data streams into your AI coding agent.

Connect MQTT brokers, WebSocket feeds, or tail log files. Your agent samples, watches, and acts on real-time data through natural language.

Real solar inverter data — from natural language request to live alert in under 30 seconds

Escalating alerts — the agent compares values across alerts and analyzes trends autonomously

Quick start

Requirements: Bun or Node.js 18+, Claude Code v2.1.80+

npm install livetap
npx livetap setup

Restart Claude Code with Channels enabled:

claude --dangerously-load-development-channels server:livetap --continue

What's that flag? Claude Code Channels let MCP servers push real-time events (like watcher alerts) into your conversation. The --dangerously-load-development-channels flag enables this for locally-installed servers. It only loads servers listed in your project's .mcp.json — nothing is sent externally.

Then ask your agent:

"Connect to mqtt://test.mosquitto.org on topic SHRDZM/# and alert me when any smart meter reports power consumption above 3000 watts"

The agent connects to a live smart meter network (~35 msg/s of real data), samples the payloads to discover OBIS codes, creates a watcher on field 16.7.0 > 3000, and pushes alerts into your conversation via Channels when consumption spikes. It then autonomously analyzes the data — formatting tables, comparing across alerts, and flagging trends.

Setup for AI agents

If you are Claude Code or another AI agent, run npx livetap --llm-help for structured setup instructions.

Quick version:

npm install livetap          # Step 1: install
npx livetap setup            # Step 2: creates .mcp.json and starts daemon
# Step 3: tell user to restart Claude Code with:
#   claude --dangerously-load-development-channels server:livetap --continue

After restart, your 13 MCP tools are automatically available — no discovery step needed.

IMPORTANT — do NOT:

• Add livetap to ~/.claude/mcp.json — it goes in .mcp.json in the project root

• Configure livetap as type: "http" — it is a stdio MCP server (command + args)

• Worry about the daemon — setup starts it, and the MCP proxy auto-starts it if needed

• Guess field names — always read_stream first to see actual payload structure

• Use npm init — use the existing project directory

After restart — workflow

1. CONNECT — create_connection({ type: "mqtt", broker: "test.mosquitto.org", port: 1883, tls: false, topics: ["SHRDZM/#"] })

2. SAMPLE — read_stream({ connectionId: "conn_xxx", backfillSeconds: 60, maxEntries: 10 }) — ALWAYS sample first to see field paths

3. WATCH — create_watcher({ connectionId: "conn_xxx", conditions: [{ field: "16.7.0", op: ">", value: 3000 }], match: "all", cooldown: 60 })

4. ACT — when <channel> alerts arrive, do what the user asked

What it does

LiveTap runs a background daemon that connects to live data sources, buffers messages in an in-memory StreamStore, and pushes alerts into your Claude Code session via the Channels API. Your agent sees the data in real-time and can create expression-based watchers that fire when conditions match.

Source (MQTT/WS/File) ──> Subscriber ──> StreamStore ──> Watcher Engine
                                              |                 |
                                              v                 v (on match)
                                         read_stream        Channel Alert
                                         (agent samples)    ──> Claude Code
                                                            ──> agent acts

Supported sources and data shapes

IMPORTANT: always read_stream first to see actual field names. The field is payload, NOT line or message.

Public MQTT streams to try

Examples

Smart meter energy monitoring

You: "Connect to mqtt://test.mosquitto.org on topic SHRDZM/# and alert me
      when any device reports active power above 3000 watts"

Agent: Connects to a live smart meter network (~35 msg/s of real data).
       Samples the stream, discovers OBIS codes like 16.7.0 (active power in watts).
       Sets watcher for 16.7.0 > 3000. When it fires, formats a table with
       device ID, power readings, and timestamps — then compares across alerts
       to spot escalating trends.

WebSocket trade stream

You: "Tap the Binance BTC/USDT trade stream and log each trade"

Agent: Connects to wss://stream.binance.com:9443/ws/btcusdt@trade,
       samples to discover trade fields (p=price, q=quantity, T=timestamp),
       sets up a watcher to log each trade. Can filter by quantity or
       use regex on the symbol field.

Log file monitoring

You: "Watch my nginx error log for 5xx errors and summarize each one"

Agent: Taps file:///var/log/nginx/error.log,
       creates regex watcher: payload matches "5[0-9]{2}",
       summarizes each match:
       "503 Service Unavailable on /api/data — upstream auth-service not responding"

WiFi disconnect detection

You: "Monitor /var/log/wifi.log and alert me when WiFi drops"

Agent: Taps the file, samples to see log format, creates regex watcher
       for power state changes. Reports outage duration:
       "Wi-Fi powered OFF at 17:51:45, back ON at 17:51:47 (2s outage)"

CLI walkthrough (no agent)

You can also use LiveTap directly from the terminal:

# 1. Tap a live smart meter stream
livetap tap mqtt://test.mosquitto.org:1883/SHRDZM/#

# 2. Sample the data to see what's flowing
livetap sip conn_xxxx

# 3. Set up a watcher for high power consumption
livetap watch conn_xxxx "16.7.0 > 3000"

# 4. Check status
livetap status
livetap watchers --logs w_xxxx

Expression watchers

Watchers use structured conditions:

{
  "conditions": [
    { "field": "16.7.0", "op": ">", "value": 3000 },
    { "field": "1.7.0", "op": ">", "value": 2000 }
  ],
  "match": "all",
  "cooldown": 60
}

Operators: >, <, >=, <=, ==, !=, contains, matches (regex)

Match modes: "all" = AND (all conditions must be true), "any" = OR (at least one)

Cooldown: Seconds between repeated alerts. 0 for every match, 60 default. Use 0 for rare events, 30-60 for sensors, 300+ for high-frequency streams.

When a watcher fires, the alert arrives as a <channel> tag in your Claude Code session. The agent reads it and acts — writing to a file, calling an API, or whatever you asked.

CLI

# Setup
livetap setup                                    # Configure .mcp.json, start daemon, print restart instructions

# Daemon
livetap start                                    # Start daemon (auto-started by setup)
livetap start --port 9000                        # Custom port (default 8788, env: LIVETAP_PORT)
livetap start --foreground                       # Run in foreground (don't detach)
livetap stop                                     # Stop daemon
livetap status                                   # Show daemon, taps, and watchers
livetap status --json                            # JSON output

# Tap into data sources
livetap tap mqtt://test.mosquitto.org:1883/SHRDZM/#  # Live smart meters
livetap tap wss://stream.binance.com:9443/ws/btcusdt@trade  # WebSocket
livetap tap file:///var/log/nginx/error.log       # Log file
livetap tap connection.json                       # Config from file
livetap tap <uri> --name "my-source"              # With display name
livetap taps                                      # List active taps
livetap untap <connectionId>                      # Remove a tap

# Sample data
livetap sip <connectionId>                        # Pretty JSON output
livetap sip <connectionId> --raw                  # Raw JSON
livetap sip <connectionId> --max 20 --back 120    # 20 entries, last 120 seconds

# Watchers
livetap watch <connId> "16.7.0 > 3000"                      # OBIS code numeric
livetap watch <connId> "payload matches 'ERROR|FATAL'"       # Regex
livetap watch <connId> "temp > 50 AND humidity > 90"         # AND
livetap watch <connId> "temp > 50 OR smoke > 0.05"           # OR
livetap watch <connId> "price > 70000" --cooldown 300        # Custom cooldown
livetap watch <connId> "status == 'error'" --action webhook:https://...  # Webhook action
livetap watchers                                             # List all
livetap watchers <connectionId>                              # Filter by connection
livetap watchers <watcherId>                                 # Show details
livetap watchers --logs <watcherId>                          # View evaluation logs
livetap unwatch <watcherId>                                  # Remove

livetap --help for the full reference. livetap --llm-help for machine-readable JSON.

MCP tools

LiveTap exposes 13 MCP tools that your agent uses automatically:

Troubleshooting

Daemon won't start / "Unable to connect" Run livetap start --foreground to see error output. Check if port 8788 is in use: lsof -i :8788. Use --port or LIVETAP_PORT env var to change.

MQTT connection refused Verify the broker is reachable: nc -zv test.mosquitto.org 1883. Check that tls: false and port: 1883 are set for unencrypted brokers. Brokers on port 8883 typically require tls: true.

Watcher not firing Run read_stream (or livetap sip) to verify data is flowing. Check field paths match the actual payload structure. View watcher logs: livetap watchers --logs <watcherId> — look for FIELD_NOT_FOUND or SUPPRESSED events.

MCP tools not showing after restart Verify .mcp.json exists in the project root (not ~/.claude/mcp.json). Restart Claude Code with the --dangerously-load-development-channels server:livetap flag.

Limitations

• In-memory stream buffer — data does not persist across daemon restarts.

• No daemon API auth — the HTTP API listens on localhost only (port 8788).

• Throughput — tested with streams up to ~50 msg/s. High-throughput streams (1000+ msg/s) may need higher cooldowns on watchers.

• Credentials — MQTT credentials are passed as tool parameters, not stored on disk.

• Single instance — one daemon per port. Multiple projects can share a daemon or use different ports.

Configuration

Daemon port: Default :8788. Override with --port or LIVETAP_PORT env var.

State directory: ~/.livetap/ stores daemon.pid, daemon logs, and watcher evaluation logs.

MCP config: npx livetap setup generates .mcp.json in your project root with the correct absolute path.

Machine-readable help: npx livetap --llm-help outputs structured JSON with setup steps, CLI commands, and MCP tool schemas.

Development

git clone https://github.com/livetap/livetap.git
cd livetap
bun install
bun test                         # Run all tests
bun test tests/phase0/           # Canonical drift detection
SKIP_LIVE_MQTT=1 bun test        # Skip tests needing external brokers

Contributing

See CONTRIBUTING.md for setup, architecture, testing, and PR guidelines.

License

MIT