yoelbassin/gnuradioMCP

LLM-driven RF for Claude Code. Describe what you want on the air and I survey the
spectrum, build and run the receiver, look at the signal, and leave a reproducible
RF project behind — SigMF captures, YAML pipelines, and .grc flowgraphs you own.
> Early development — v1.0 is simulation-only (no hardware yet). See ROADMAP.md.
> Proviously known as GNURadio/GR-MCP.
Requirements
- uv and Python ≥ 3.13
- GNU Radio 3.10+ installed system-wide (conda-forge,
your distro's package manager, or Homebrew) — needed to run pipelines and
simulate. Analyzing existing IQ files doesn't require it.
Install
In Claude Code:
That's it — uv starts the MCP server on first use and loads the skills. Nothing
else to configure.
Use it
Ask in plain language:
> "Simulate an FM station at 100.3 MHz and build me a receiver."
> "Here's a capture, mystery.wav — survey what's on the air, then decode the strongest carrier."
I pick the right skill and drive the workflow end to end:
- **survey-spectrum** — what's present (PSD, signal detection, spectrogram)
- **simulate-scene** — register a simulated device from a description
- **build-receiver** — channelize, demodulate, validate, run, verify, save
- **debug-no-signal** — locate the fault when a receiver outputs nothing
- **tx-experiment** — closed-loop transmit-then-receive in simulation
- **escape-hatch** — drop to the Python library when no tool fits
Artifacts land under artifacts/ in the server's working directory; set
MARCONI_WORKSPACE to put them elsewhere.
License
GPL-3.0-or-later.