Not-Happy-Jan

Multi-sensory feedback for AI coding agents. When your agent finishes a task, Jan tells you. When something goes wrong, Karren lets you know — loudly.

"Not happy Jan, the issue is…"

☎️ The Call-Centre Experience

Ever feel like you're on hold, waiting for your agent to finish processing?

Not-Happy-Jan makes Claude Code feel like a genuine call to a government department — multiple tiers of staff, hold music, and constant feedback delivered with a customisable level of competence and boganism — all while you wait for your token quota to fill back up.

• 🎷 Hold music plays while the agent's busy, pauses the instant someone picks up

• 👩 Jan — reception. Handles the routine stuff. Cheerful, occasionally scatty.

• 🧑‍💼 Bazza — middle management. Escalates when things get warm. Apathetic to anxious, your call.

• 📣 Karren — the manager. Takes over when it all goes wrong. She is NOT. HAPPY. JAN.

• 🎚️ Dial in each character's competence, boganism and chaos to taste — by just asking.

Related MCP server: Notifications MCP Server

What it does

By default NHJ generates its voices with a local TTS model, running right on your own machine — building on earlier agent-feedback work, but going much further. NHJ is extensible to a whole range of feedback tools: from the haptic Logitech MX Master mouse to $60 Ulanzi pixel displays, LaMetric Time, the Divoom Times Gate / Frame — or anything you can POST to, like a Vestaboard. When Karren's not happy, Jan and you are going to know about it.

When Claude Code (or any MCP-compatible agent) completes a task, NHJ fires across every output channel you have configured:

The cast & modes

Three tiers answer the phone depending on severity — Jan (reception), Bazza (middle management), Karren (the manager). Each has its own zero-shot-cloned voice and live dials (boganism, competence, chaos) you tune just by asking Claude. One-word modes (rave, call-centre, quiet, special-forces, went-full-bogan…) reconfigure the whole vibe — music, voice, displays, haptics — at once.

→ Characters & dials · Modes

Install

curl -fsSL https://raw.githubusercontent.com/guruswami-ai/not-happy-jan/main/install.sh | bash

Or install from a checkout:

git clone https://github.com/guruswami-ai/not-happy-jan
cd not-happy-jan
bash install.sh

One command, non-interactive. It downloads the source and bootstraps uv plus a managed Python runtime when required. The default profile installs the full experience — venv, Qwen3-TTS voice model, the ocker-bogan-nano LLM, bundled media, Claude Code hooks, and MCP server (stdio, localhost-only) — using on-demand model loading (no always-on daemon, so no resident RAM cost at idle).

👉 Just run bash install.sh — and go for the full experience. The default is the full experience, and it's what Not-Happy-Jan is for: Jan, Bazza and Karren in their own live, cloned voices, the dynamic ocker-bogan-nano brain writing a fresh in-character line every time, and 4+ hours of hold music ducking under it all — 100% local, loaded on demand (nothing resident at idle). It's a one-off ~5 GB download on an Apple Silicon Mac, then it just works.

--minimal (pre-recorded bank, zero downloads) is there for CI and RAM-tight machines — the cast still speaks, but you miss the live voices, the dynamic lines, and the music. Don't reach for it unless you have to.

The MCP server always binds to 127.0.0.1. LAN access requires an explicit opt-in: bash install.sh --full --listen 0.0.0.0.

Uninstall — one command stops the TTS/LLM/MCP services, then removes everything (runtime, config, downloaded models/media, the nhj launcher, and NHJ's Claude hooks/MCP/markers/skill); your Claude install is left intact. Like the install, it runs entirely in your user account — no sudo (user-domain LaunchAgents, files under $HOME):

nhj uninstall          # add --yes to skip the confirmation

The dynamic brain — included in the default install. bash install.sh (and --full) download ocker-bogan-nano, NHJ's own fine-tuned LLM, so every line is freshly generated in character instead of replayed from the bundled bank. It needs llama.cpp (brew install llama.cpp); if that's missing the installer notes it as a follow-up and the bundled bank still works. To (re)install or point at your own model:

nhj install-model    # ~940 MB Q4_K_M GGUF, runs as a local llama.cpp server — 100% local, sub-second

Why it's worth it, the BYO-model path, and the swearing/bleep behaviour are all in Dynamic voices.

Requirements: macOS Apple Silicon (M1+), Python 3.10+, ~8 GB RAM (16 recommended), ~5 GB disk. The initial alpha supports Apple Silicon macOS only. TTS_ENGINE=none runs text-only; low-spec machines can run from pre-recorded samples. The three ways to run it (silent · pre-recorded · full live experience) → Minimum specs; install profiles → Install profiles; settings → Configuration & requirements.

Quick start

# Test all configured adapters
nhj test ok
nhj test err --message "something broke"

# List available voices
nhj voices

# List configured devices
nhj devices

# Pre-render audio bank for faster playback
nhj build-bank --voice jan

# Live settings — no restart needed (or just ask Claude in plain language)
nhj muzak on                # inference muzak while the agent works
nhj set jan.ockerism 11     # tune any character dial (jan.competence, bazza.stress, …)
nhj mute                    # silence everything;  nhj unmute to restore
nhj status                  # show current settings

Documentation

Your new friends

NHJ is fully customisable: 72 HR-approved hold-music tracks (over 4 hours of on-hold listening pleasure), replaceable personalities, and zero-shot voice cloning — drop in any voice you like. But we reckon that once you've got to know Jan, tuned her competence and bogan settings to your liking, and met anxious Bazza and cranky Karren, you won't want to code with Claude without your new friends.

Credits

Evolved from an earlier personal agent-feedback project by Paul Nevin. TTS powered by Qwen3-TTS via mlx-audio.

License

Code is available under the MIT License. Downloadable media and voice assets are licensed separately from the code and have mixed provenance — mostly AI-generated (released CC0) or procedurally synthesised, with hold music generated under a licensed Suno plan. See Media provenance for per-asset terms. Do not assume the code license grants rights to media, and use only voice references you own or have documented consent to use.