siglent-sdg-mcp

A Model Context Protocol (MCP) server that lets AI assistants control Siglent SDG waveform generators over your local network. Connect Claude to your bench and generate, modulate, and configure signals through natural language.

Overview

This MCP server communicates with Siglent SDG waveform generators via SCPI commands over TCP sockets (port 5025). No VISA drivers or NI-MAX installation required — just a network connection to your generator.

Key features:

• 20 tools covering output control, waveform generation, modulation, sweep, burst, arbitrary waveforms, and more (21 with the opt-in raw SCPI tool)

• Dual-channel support (C1 and C2)

• 8 waveform types: sine, square, ramp, pulse, noise, arbitrary, DC, PRBS

• 8 modulation types: AM, DSB-AM, FM, PM, PWM, ASK, FSK, PSK

• Auto-connect on startup via environment variable

• Query queue serializes commands automatically — tools can safely run in parallel

• Read-only SCPI query escape hatch; raw SCPI commands are opt-in

• Security guardrails: command-injection protection, query-only scpi_query, connect-target allowlist (see Security)

Related MCP server: jds6600-mcp

Compatibility

The server uses standard SCPI commands from the SDG series programming guide. Other Siglent models that support the same command set over port 5025 should work with little or no modification.

Quick Start

You need a Siglent waveform generator accessible on your network (TCP port 5025). Pick one of the three options below and add the config to your .mcp.json (in your project directory, or ~/.claude/.mcp.json for global access).

Replace 192.168.1.126 with your generator's IP address.

Option A: Docker (recommended)

No Node.js installation required. Works on Linux, macOS, and Windows (via WSL2 or Docker Desktop).

{
  "mcpServers": {
    "siglent-sdg": {
      "type": "stdio",
      "command": "docker",
      "args": [
        "run", "--rm", "-i",
        "-e", "SIGLENT_SDG_IP=192.168.1.126",
        "ghcr.io/magnusjohansson/siglent-sdg-mcp:latest"
      ]
    }
  }
}

Option B: npx

Requires Node.js 20+. Downloads and runs the package automatically.

{
  "mcpServers": {
    "siglent-sdg": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "siglent-sdg-mcp"],
      "env": {
        "SIGLENT_SDG_IP": "192.168.1.126"
      }
    }
  }
}

Option C: Clone and build

git clone https://github.com/magnusjohansson/siglent-sdg-mcp.git
cd siglent-sdg-mcp
npm install
npm run build

{
  "mcpServers": {
    "siglent-sdg": {
      "type": "stdio",
      "command": "node",
      "args": ["/path/to/siglent-sdg-mcp/build/index.js"],
      "env": {
        "SIGLENT_SDG_IP": "192.168.1.126"
      }
    }
  }
}

Replace /path/to/siglent-sdg-mcp with the actual path to your clone.

Environment Variables

Auto-Connect Behavior

If SIGLENT_SDG_IP (or SIGLENT_IP) is set, the server attempts to connect to the generator immediately after starting. This runs in the background and does not block the MCP server — Claude can start using other tools right away. If the generator is offline or unreachable, the server logs a warning and you can connect manually later using the connect tool.

If neither variable is set, the server starts without a connection. Use the connect tool to connect when ready. In that case connect only accepts private-LAN IP addresses; see Security.

Using with Other AI Clients

The Quick Start examples above use Claude Code's .mcp.json format, which includes a "type": "stdio" field. Other AI clients use the same JSON structure but without the "type" field and with different config file locations.

Claude Desktop

Edit claude_desktop_config.json:

• Windows: %APPDATA%\Claude\claude_desktop_config.json

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

Docker

{
  "mcpServers": {
    "siglent-sdg": {
      "command": "docker",
      "args": [
        "run", "--rm", "-i",
        "-e", "SIGLENT_SDG_IP=192.168.1.126",
        "ghcr.io/magnusjohansson/siglent-sdg-mcp:latest"
      ]
    }
  }
}

npx

{
  "mcpServers": {
    "siglent-sdg": {
      "command": "npx",
      "args": ["-y", "siglent-sdg-mcp"],
      "env": {
        "SIGLENT_SDG_IP": "192.168.1.126"
      }
    }
  }
}

Note: You must fully restart Claude Desktop after changing the config file.

Cursor

Edit one of:

• User-level: ~/.cursor/mcp.json (available across all projects)

• Project-level: .cursor/mcp.json (shared with your team via version control)

You can also add servers through the UI: Settings > Cursor Settings > MCP > Add new global MCP server.

Docker

{
  "mcpServers": {
    "siglent-sdg": {
      "command": "docker",
      "args": [
        "run", "--rm", "-i",
        "-e", "SIGLENT_SDG_IP=192.168.1.126",
        "ghcr.io/magnusjohansson/siglent-sdg-mcp:latest"
      ]
    }
  }
}

npx

{
  "mcpServers": {
    "siglent-sdg": {
      "command": "npx",
      "args": ["-y", "siglent-sdg-mcp"],
      "env": {
        "SIGLENT_SDG_IP": "192.168.1.126"
      }
    }
  }
}

Windsurf

Edit mcp_config.json:

• Windows: %USERPROFILE%\.codeium\windsurf\mcp_config.json

• macOS/Linux: ~/.codeium/windsurf/mcp_config.json

You can also configure servers through the UI: Cascade panel > MCP icon > Manage MCP Servers > View raw config.

Docker

{
  "mcpServers": {
    "siglent-sdg": {
      "command": "docker",
      "args": [
        "run", "--rm", "-i",
        "-e", "SIGLENT_SDG_IP=192.168.1.126",
        "ghcr.io/magnusjohansson/siglent-sdg-mcp:latest"
      ]
    }
  }
}

npx

{
  "mcpServers": {
    "siglent-sdg": {
      "command": "npx",
      "args": ["-y", "siglent-sdg-mcp"],
      "env": {
        "SIGLENT_SDG_IP": "192.168.1.126"
      }
    }
  }
}

Google Antigravity

Configuration is managed through the IDE's UI:

1. Open the Agent pane on the right side of the workspace

2. Click the ... button at the top

3. Select MCP Servers

4. Click Manage MCP Servers

5. Click View raw config

6. Add the configuration below and save

Docker

{
  "mcpServers": {
    "siglent-sdg": {
      "command": "docker",
      "args": [
        "run", "--rm", "-i",
        "-e", "SIGLENT_SDG_IP=192.168.1.126",
        "ghcr.io/magnusjohansson/siglent-sdg-mcp:latest"
      ]
    }
  }
}

npx

{
  "mcpServers": {
    "siglent-sdg": {
      "command": "npx",
      "args": ["-y", "siglent-sdg-mcp"],
      "env": {
        "SIGLENT_SDG_IP": "192.168.1.126"
      }
    }
  }
}

Replace 192.168.1.126 with your generator's IP address in all examples above.

ChatGPT Desktop

ChatGPT Desktop only supports remote HTTPS MCP servers (called "connectors"), not local stdio servers. Since this MCP server uses stdio transport, it is not directly compatible with ChatGPT Desktop.

Tools

20 tools across 9 categories by default; scpi_command is only registered when SIGLENT_ALLOW_RAW_SCPI=true.

Example Conversations

Generate a basic waveform

You: Output a 1 kHz sine wave at 2 Vpp on channel 1.

Claude calls configure_basic_wave with channel: "C1", waveform_type: "SINE", frequency: 1000, amplitude: 2, then configure_output with channel: "C1", state: "ON".

Set up modulation

You: Add AM modulation to channel 1 with 80% depth at 100 Hz.

Claude calls configure_modulation with channel: "C1", state: "ON", type: "AM", depth: 80, frequency: 100.

Configure a frequency sweep

You: Sweep channel 2 from 100 Hz to 10 kHz over 5 seconds.

Claude calls configure_sweep with channel: "C2", state: "ON", start: 100, stop: 10000, time: 5, then configure_output with channel: "C2", state: "ON".

Check current settings

You: What's the current waveform setup on both channels?

Claude calls get_basic_wave on both C1 and C2 in parallel and reports the waveform type, frequency, amplitude, and other settings for each channel.

Copy channel configuration

You: Make channel 2 match channel 1's settings.

Claude calls copy_channel with source: "C1", destination: "C2".

Set up burst mode

You: Configure channel 1 for 5-cycle bursts triggered externally.

Claude calls configure_burst with channel: "C1", state: "ON", burst_mode: "NCYC", cycles: "5", trigger_source: "EXT".

Security

An MCP server turns a language model's output into actions. For this server those actions are physical: a wrong command can put 20 Vpp into a 3.3 V circuit. Tool arguments are therefore treated as untrusted input, whether they came from your request, from a document the model read, or from a prompt injection.

What the server enforces, regardless of what the model was told:

• Output voltage. This server does not limit amplitude or offset. The model can set anything the generator allows, up to 20 Vpp. To protect a circuit, set the generator's own limit on the front panel or with C1:BSWV MAX_OUTPUT_AMP,<Vpp> (on models that support it). This backstop only holds while raw SCPI is off: with SIGLENT_ALLOW_RAW_SCPI=true, the model can send MAX_OUTPUT_AMP itself and raise the limit. Whether reset (*RST) restores the generator's default limit isn't documented; check it on your model before relying on the backstop. Server-side limits are being designed separately: a first version was reviewed and found to be bypassable, so it isn't shipped (see the threat model).

• Other clients on port 5025. Anyone with network access to the instrument can control it directly, without going through this server. The instrument has no authentication. Network segmentation is the control for that.

The full threat model, including residual risks, is in docs/threat-model.md.

Architecture

Claude Code <-- stdio/JSON-RPC --> siglent-sdg-mcp <-- TCP/SCPI --> Generator:5025

• Transport: MCP over stdio (JSON-RPC 2.0)

• Protocol: SCPI commands over raw TCP sockets, newline-terminated

• Query Queue: All SCPI queries are serialized through an internal queue. The generator processes one command at a time, so even when tools issue parallel requests, the queue ensures they're sent sequentially.

• Auto-Connect: If SIGLENT_SDG_IP is set, connects in the background on startup without blocking the MCP server.

• Security choke point: every command and query passes through connection.ts, which applies the checks in src/security/ before anything is written to the socket.

Development

npm run build       # Compile TypeScript
npm run watch       # Watch mode — recompile on changes
npm run dev         # Build and run
npm run inspector   # Launch with MCP Inspector for debugging
npm test            # Unit tests + end-to-end tests against the SCPI simulator

The end-to-end tests run the real server and connection code over a real TCP socket against a small SCPI simulator (tests/simulator/), so no hardware is needed. Security tests assert on what actually reached the simulator, not only on the error returned.

To check against real hardware (never turns an output on, aborts unless both outputs are off, because channel coupling can carry C1 changes to C2, uses 1 Vpp on C1, and changes C1's waveform settings):

npm run hardware-check -- 192.168.1.126

Releasing

Publishing is triggered by publishing a GitHub Release, not by pushing a tag:

1. Bump version in package.json and add a section to CHANGELOG.md, then merge to main.

2. Create a release (draft first) with tag vX.Y.Z on main, using the changelog section as the release notes.

3. Publish the release. This runs the npm publish (tests first, with provenance) and the GHCR image build.

The npm workflow fails if the tag doesn't match the package.json version. A pre-release publishes to the npm next dist-tag, and doesn't move the Docker latest tag. A draft release publishes nothing.

Docker (local build)

Build the image locally:

docker build -t siglent-sdg-mcp .

Then use the local image in your .mcp.json:

{
  "mcpServers": {
    "siglent-sdg": {
      "type": "stdio",
      "command": "docker",
      "args": [
        "run", "--rm", "-i",
        "-e", "SIGLENT_SDG_IP=192.168.1.126",
        "siglent-sdg-mcp"
      ]
    }
  }
}

Project Structure

src/
  index.ts              # Entry point: stdio transport, auto-connect
  server.ts             # MCP server factory, tool registration
  config.ts             # Environment parsing, effective security settings
  connection.ts         # TCP socket manager with query queue (security choke point)
  security/
    scpi-guard.ts       # Command-boundary and query-only checks
    host-policy.ts      # Connect-target allowlist
  tools/
    connection.ts       # connect, disconnect, identify
    output.ts           # get_output, configure_output
    basic-wave.ts       # get_basic_wave, configure_basic_wave
    modulation.ts       # get_modulation, configure_modulation
    sweep.ts            # get_sweep, configure_sweep
    burst.ts            # get_burst, configure_burst
    arbitrary.ts        # get_arbitrary_wave, set_arbitrary_wave
    utility.ts          # reset, copy_channel, configure_sync, equal_phase
    scpi.ts             # scpi_query, scpi_command (opt-in)
tests/
  unit/                 # Guard, host policy and config tests
  integration/          # End-to-end security tests via MCP client
  simulator/            # SCPI simulator (TCP)

Troubleshooting

"Not connected to waveform generator"

The generator isn't connected yet. Either set SIGLENT_SDG_IP in your .mcp.json env for auto-connect, or use the connect tool manually.

Connection timeout

• Verify the generator's IP address (check the generator's Utility > Interface menu)

• Ensure port 5025 is accessible (try telnet <generator-ip> 5025 from your machine)

• Check that no firewall is blocking the connection

• Close other SCPI clients (another MCP server instance, a second Claude session, NI-MAX, scripts). The SDG1032X accepts more than one TCP connection, so a second client does not fail to connect. It interferes instead (see below)

Query timeout

Some SCPI queries can take a few seconds. The default timeout is 5 seconds. For scpi_query, you can increase the timeout with the timeout_ms parameter.

Timeouts on queries that normally answer instantly (identify, get_output) usually mean another client is talking to the generator at the same time. A common cause is a second Claude Code session whose MCP config also starts this server with SIGLENT_SDG_IP set, so both auto-connect. The server serializes its own queries, but it can't coordinate with other processes, and their commands and replies interleave on the instrument. Close the other client and reconnect.

Docker: can't reach the waveform generator

By default, Docker containers can reach LAN devices via the bridge network (NAT). If the container can't connect to your generator:

• Verify the generator is reachable from your host: telnet 192.168.1.126 5025

• On Linux, try adding --network host to the Docker args:

  "args": ["run", "--rm", "-i", "--network", "host", "-e", "SIGLENT_SDG_IP=192.168.1.126", "ghcr.io/magnusjohansson/siglent-sdg-mcp:latest"]

  Note: --network host does not work on macOS or Windows Docker Desktop.

Docker: wrong architecture / exec format error

The published image supports linux/amd64 and linux/arm64. Docker should pull the correct one automatically. If you see an exec format error, pull explicitly:

docker pull --platform linux/amd64 ghcr.io/magnusjohansson/siglent-sdg-mcp:latest

"CHDR" appears in responses

This shouldn't happen — the server sets CHDR OFF on connect. If you see command headers in responses, try disconnecting and reconnecting.

License

MIT — see LICENSE for details.