embeddedci-mcp 0.1.0

MCP server exposing the EmbeddedCI BenchPod SDK as tools for AI agents

embeddedci-mcp

An MCP server that exposes the embeddedci BenchPod SDK as tools, so an AI agent can drive a real hardware-in-the-loop bench: power a target board, flash it over SWD, capture its UART, and emulate/decode an I2C sensor.

It's a thin consumer of the SDK — every tool maps directly to a embeddedci.benchpod.BenchPod method.

Install

pip install embeddedci-mcp        # pulls embeddedci from PyPI
# or run without installing:
uvx embeddedci-mcp --help

For local development from this repo, see the workspace README.

Run

# stdio (launched by an MCP client as a subprocess — the usual case):
embeddedci-mcp --transport stdio

# streamable HTTP (for a remote bench):
embeddedci-mcp --transport http --host 0.0.0.0 --port 8000

# preset a default connection so the `connect` tool needs no argument:
embeddedci-mcp --connection /dev/tty.usbserial-0001
embeddedci-mcp --connection 192.168.1.213        # wifi/TCP, default port 8080

The connection can also come from the BENCHPOD_CONNECTION environment variable.

Client configuration

Claude Desktop / Cursor (mcp.json / claude_desktop_config.json)

{
  "mcpServers": {
    "benchpod": {
      "command": "uvx",
      "args": ["embeddedci-mcp"],
      "env": { "BENCHPOD_CONNECTION": "192.168.1.213" }
    }
  }
}

Use "command": "embeddedci-mcp" instead if it's installed on PATH.

Claude Code

claude mcp add benchpod -- uvx embeddedci-mcp

Tools

Group	Tools
Lifecycle / status	connect, disconnect, ping, status
Power	power_on, power_off, target_power, target_status
Flash	flash
UART	capture_uart, power_cycle_and_capture
I2C sensor	enable_i2c_sensor, set_i2c_sensor, disable_i2c_sensor, i2c_sensor_status, i2c_sensor_regs, i2c_sensor_la_decoded, i2c_read_register
Pull-ups	enable_pullup, disable_pullup, pullup_status
Low-level	command, gpio_set, capture_adc, signal_generate, measure

Device/firmware failures come back as {"ok": false, "error": ..., "error_type": ...} rather than raising, so the agent can reason about them.

Resources

• benchpod://wiring — the default LA channel → DUT signal pin map and eFuse table.
• benchpod://help — the canonical HIL workflow order.

Example agent flow

1. connect("192.168.1.213")
2. flash(swclk=11, swdio=12, nreset=3, target="target/stm32f4x.cfg", file="app.elf", target_power=1)
3. enable_pullup([1, 2]) then enable_i2c_sensor(sda=2, scl=1, temperature_c=22.5, pressure_pa=101000)
4. power_cycle_and_capture(rx=5, tx=4, delay=1.5, duration=6.0, until_regex="APP_OK")
5. i2c_sensor_status() / i2c_read_register(address=0x76, register=0xD0) to confirm the DUT probed the sensor.

Publishing to the MCP Registry

server.json is starter metadata for the Official MCP Registry (currently in preview). Ship the package to PyPI first (that's what makes uvx embeddedci-mcp work), then publish the registry entry once the name is stable:

mcp-publisher publish      # GitHub-authenticated; reads server.json