American PhD in economics Khaled Eltohi has launched a 24/7 AI radio station, WRIT-FM, where Claude is responsible for all the creative. The station operates continuously, generating all content in real time.
American PhD in economics Khaled Eltohi has launched a 24/7 AI radio station, WRIT-FM, where Claude is responsible for all the creative. The station operates continuously, generating all content in real time.
As Khaled Eltohi explained on Reddit, all the code was developed using Claude Code: the CLI writ command interface, the streaming pipeline, the host persona system, the content generators, and the schedule parser.
Technology stack: Python, ffmpeg, Icecast, Claude CLI for scripting, Kokoro TTS for voiceover, ACE-Step for creating music intros using AI. Runs on Mac Mini.
Every word that is spoken on the air is created through Claude CLI. The station has 5 unique AI hosts: The Liminal Operator (night philosophy), Dr. Resonance (music history), Nyx (night reflections), Signal (news analysis), and Ember (soul/funk). Each has a defined identity, voice style, philosophy, and distinct “anti-patterns” (phrases or topics they would never say).
The system generates scripts of 1,500–3,000 words for 7 types of segments: in-depth analytical analysis, simulated interviews, panel discussions (debates between two AI hosts), news analysis (based on real RSS feeds), stories, music essays, and columns with listener letters.
Kokoro TTS synthesizes audio from scripts by breaking up long fragments beyond sentences and combining them using ffmpeg.
The weekly schedule features 8 different shows, and Claude writes scripts for each of them, adapting the tone, focus on topics, and style of speaking to each individual host. The news show uses real headlines from RSS feeds, which Claude interprets through the prism of the nightly broadcast, rather than simply retelling them.
The speech module (stream_gapless.py) works solely on heuristics — without AI during the broadcast. It determines the active show by searching in the file, plays speech blocks from the corresponding show's queue, inserts AI-generated musical intros (ACE-Step) between them, and deletes segments after playback.
Daemon scripts monitor the number of segments and trigger generation when the stock falls below a threshold. Replay history in SQLite prevents replays within a 4-hour window.
“The key architectural decision was to avoid using AI directly in the runtime. Pre-generating content into file system queues, which are then consumed by the streamer, ensures that the broadcast never stops waiting for a response from the API. It also turned out that character “anti-patterns” (clear lists of what the host should NEVER do) are more important for image stability than positive instructions,” the developer shared his conclusions.
The radio station is available at WRIT-FM at the link . You can also check out the project on GitHub .
