Can not start transcription of any recording

[Jibaru]

When loading a new Heed instance, I got the following error when it started processing audio:

GET http://localhost:5170/api/models

{"error":"Unable to connect. Is the computer able to access the url?"}

Logs:

npx create-heed                          

  heed — local-first meeting transcription
  every voice, even when they speak at once.
  Local. Open. Yours.

> Detected: macOS (Apple M4 Pro)


[1/7] Bun runtime
✓ Bun 1.3.5 already installed

[2/7] Python 3.10+
✓ Python 3.14.2 found

[3/7] AI models (faster-whisper + pyannote)
✓ faster-whisper already installed

[4/7] ffmpeg (audio capture)
✓ ffmpeg already installed

[5/7] Ollama (local AI engine)
✓ Ollama ollama version is 0.31.1 already installed

[6/7] Download heed
✓ heed already exists at /Users/ignaciorueda/Documents/Jibaru/gamemaker/heed
> Pulling latest changes...
> Running: cd "/Users/ignaciorueda/Documents/Jibaru/gamemaker/heed" && git pull
Already up to date.
✓ Updated
> Installing JavaScript dependencies...
> Running: cd "/Users/ignaciorueda/Documents/Jibaru/gamemaker/heed" && bun install
bun install v1.3.5 (1e86cebd)

Checked 401 installs across 448 packages (no changes) [11.00ms]
✓ Dependencies installed

[7/7] Launch heed

  All set!

? Open as floating desktop panel? (recommended) (Y/n) Y
  Starting services + floating panel...

> Starting services...
[heed] Opening http://localhost:5170
Opening in existing browser session.

[isjunrod]

Hey Jibaru — thank you, this is a genuinely useful report. You were the first to run heed on a clean M4 and you surfaced a real gap.

Root cause: the npx create-heed you hit was a stale build — it set up the old Whisper + pyannote path without building the Parakeet engine, and it didn't guard against Python 3.14 (which has no wheels yet for the AI deps). So the transcription server never actually started, which is why /api/models couldn't connect.

Fixed in create-heed@0.2.1 (just published). It now:

• pins the Python venv to a supported version (3.10–3.12, installing 3.12 if needed) and rebuilds a stale venv instead of reusing it;
• always leaves you with a working engine — Parakeet on the Apple Neural Engine, or an automatic fallback;
• self-heals an out-of-date local copy, and surfaces a clear, actionable error instead of failing silently.

Just re-run:

npx create-heed@latest

That's it — it realigns your local copy, rebuilds the venv on the right Python, and builds the Parakeet engine. (If you'd rather start 100% clean: rm -rf the heed folder first, then the same command.)

If anything still trips, paste the final health-check output from the Launch step and I'm on it. Really appreciate you kicking the tires — made heed better on day one.

[Jibaru]

Hey Jibaru — thank you, this is a genuinely useful report. You were the first to run heed on a clean M4 and you surfaced a real gap.

Root cause: the npx create-heed you hit was a stale build — it set up the old Whisper + pyannote path without building the Parakeet engine, and it didn't guard against Python 3.14 (which has no wheels yet for the AI deps). So the transcription server never actually started, which is why /api/models couldn't connect.

Fixed in create-heed@0.2.1 (just published). It now:

• pins the Python venv to a supported version (3.10–3.12, installing 3.12 if needed) and rebuilds a stale venv instead of reusing it;
• always leaves you with a working engine — Parakeet on the Apple Neural Engine, or an automatic fallback;
• self-heals an out-of-date local copy, and surfaces a clear, actionable error instead of failing silently.

Just re-run:

npx create-heed@latest
That's it — it realigns your local copy, rebuilds the venv on the right Python, and builds the Parakeet engine. (If you'd rather start 100% clean: rm -rf the heed folder first, then the same command.)

If anything still trips, paste the final health-check output from the Launch step and I'm on it. Really appreciate you kicking the tires — made heed better on day one.

So now i have:

npx create-heed@latest

  heed — local-first meeting transcription
  every voice, even when they speak at once.
  Local. Open. Yours.

> Detected: macOS · 12 cores · 48GB RAM
>   Apple M4 Pro
✓ Engine: Parakeet on the Apple Neural Engine — fastest, 100% local

  Fast     record + transcribe + diarize + copy (lightweight)
  Complete + AI notes (Ollama + LLM, ~2-5GB — you can enable it later)
? Add AI notes now (Complete)?  No = Fast (Y/n) n
✓ Preset: Fast

[1/8] Bun runtime
✓ Bun 1.3.5 already installed

[2/8] Python 3.10+
✓ Python 3.14.2 found

[3/8] ffmpeg (audio capture)
✓ ffmpeg already installed

[4/8] Download heed
✓ heed already exists at /Users/ignaciorueda/Documents/Jibaru/goals/heed
> Pulling latest changes...
> Running: cd "/Users/ignaciorueda/Documents/Jibaru/goals/heed" && git merge --ff-only origin/main
Already up to date.
✓ Updated
> Installing JavaScript dependencies...
> Running: cd "/Users/ignaciorueda/Documents/Jibaru/goals/heed" && bun install
bun install v1.3.5 (1e86cebd)

Checked 401 installs across 448 packages (no changes) [17.00ms]
✓ Dependencies installed

[5/8] Python runtime (lightweight core)
✓ Python core already installed in .venv

[6/8] Parakeet speed engine (Apple Neural Engine)
✓ Parakeet sidecar already built

[7/8] System audio permission (one-time)
> heed needs permission to capture your meetings' system audio.
> A macOS dialog will appear — approve it. (Mic-only recording works without this.)
✓ System audio permission granted — press record and you're done

[8/8] Launch heed
> Health check (confirming transcription + diarization work here)...
> Running: "/Users/ignaciorueda/Documents/Jibaru/goals/heed/.venv/bin/python3" "/Users/ignaciorueda/Documents/Jibaru/goals/heed/packages/transcription/doctor.py"

heed doctor — checking this machine end-to-end...

  [OK  ] ffmpeg (audio I/O): found
[heed] Capabilities (cached): apple_gpu parakeet rtf(small)=0.0x
  [FAIL] hardware probe: apple_gpu/parakeet, reference small @ 0.0x real-time
  [OK  ] model policy: final=base, live=base
  [FAIL] transcription: [Errno 32] Broken pipe
  [OK  ] diarization (optional): FluidAudio (CoreML, no token) — 1 speaker(s) on sample
  [OK  ] notes engine / Ollama (optional): notes engine reachable on :11434

heed has a problem in a CORE stage above — transcription may not work yet.
✗ Failed: "/Users/ignaciorueda/Documents/Jibaru/goals/heed/.venv/bin/python3" "/Users/ignaciorueda/Documents/Jibaru/goals/heed/packages/transcription/doctor.py"
! A check failed above. If transcription failed, install the fallback engine:
  npx create-heed fallback

  All set!

? Open as floating desktop panel? (recommended) (Y/n) y
  Starting services + floating panel...

> Starting services (loading models)...
[heed] Opening http://localhost:5170
Opening in existing browser session.