Local Servers — Piper, Whisper, Studio, crawl

Settings → Local Servers is in the Assistants group. The tab subtitle is exactly “Piper, Whisper, HyperFrames, Skill Seekers, and web crawl.” Voice devices, Piper voice/speed/volume, and the Whisper model stay on Settings → Voice. MCP connections (Google, GitHub, Discord, Telegram, Linear, custom stdio) live on Settings → Extensions — not here.

Assistants localhost only Start on first use

Five cards, as they appear

The tab is a vertical stack: Piper TTS, Whisper STT, HyperFrames, Skill Seekers, Web crawler. Toggles persist immediately. Status pills poll every few seconds (ready / connecting / error / disabled).

These processes bind localhost. Do not expose 4123, 8935, 4567, 8765, 11235, or 51452 to the internet.

Launch no longer auto-starts sidecars

VibeFuse no longer starts Whisper, Crawl4AI, HyperFrames Studio, or Skill Seekers on every boot. Auto-starting them saturated the machine and made the canvas wait for a free core. Each engine starts on first use or when you press the card button.

EngineHow it starts
PiperPer speak request — piper.exe is not a daemon. No Start button.
WhisperOne-shot whisper-cli on the next listen, if Use local Whisper is on.
HyperFramesStart Studio, Open widget, or the first Jarvis HyperFrames tool.
Skill SeekersConnect on the card (or the first Jarvis lookup that needs it).
Crawl4AI + SearXNGStart web tools, or the first web_search / scrape call.

Piper TTS

Card description: “Local text-to-speech server for Jarvis replies. Voice, speed, and volume are in Settings → Voice.” Voice / speed / volume are not on this tab.

Controls on this card

ControlWhat it does
Use local PiperOn: spawn piper.exe per request (and the local :4123 API if the exe is missing). Off: never spawn local Piper; speech uses the remote Piper endpoint.
Piper endpointOpenAI-compatible /v1/audio/speech URL. Placeholder http://127.0.0.1:4123.
Test PiperProbes STT + TTS and prints Speech ready/down · Piper ready/down.

Local Piper is a per-request piper.exe spawn — not a long-running daemon, and there is no Start / Restart / Stop. Full voice list, ack clips, and fallbacks: Piper TTS.

Whisper STT

Card description: “Local speech-to-text server. The Whisper model is in Settings → Voice.”

ControlWhat it does
Use local WhisperOn: run local whisper-cli. Off: never run the CLI; transcription uses the remote Whisper endpoint.
Whisper endpointFallback transcribe API when local CLI is down. Placeholder http://127.0.0.1:8935.

Local STT is a one-shot whisper-cli process per take — not a Whisper daemon. Pick the ggml model on Settings → Voice. Engine, ffmpeg, and models: Whisper STT.

HyperFrames

Card description: “Local HyperFrames Studio for HTML-to-video compositions.” Needs Node.js 20+ on PATH.

  • Enable HyperFrames Studio — Off: Studio is stopped and HyperFrames AI tools are hidden. Rendered videos already on the canvas keep playing.
  • Workspace C:\vibefuse\hyperframes · Studio port 4567 (http://127.0.0.1:4567).

Chips

Studio UI · Init · Catalog · Preview · Check · Render · Capture · Media

Buttons

Start Studio when it is down; Restart / Stop when it is ready; Open widget always available (places Studio on the canvas).

Jarvis can init, compose, check, preview, and render through this server. Walkthrough: HyperFrames Studio.

Skill Seekers

Card description: “Find existing docs-to-skill presets and package new skills locally.” There is no on/off toggle — only Connect / Disconnect.

  • Local Python MCP from yusufkaraaslan/Skill_Seekers
  • Path C:\vibefuse\Skill_Seekers
  • Requires Python 3.11+ (3.12 preferred) and Git. No API key.
  • HTTP port 8765 on localhost
  • First Connect clones the repo and builds a venv; later Connects reuse it

Chips

Find skills · Docs → skill · GitHub

Jarvis uses these tools to search presets (React, Godot, Blender, Claude Code, Unity, and more) before calling create_skill. Launch does not auto-start Skill Seekers — press Connect, or let the first lookup start it. How to write a skill after you find one: Write a skill.

If tools are missing, Disconnect then Connect so the venv rebuilds. VibeFuse sets PYTHONUTF8=1 so Windows cp1252 logs do not crash the MCP.

Web crawler

Card description: “SearXNG for AI search. Crawl4AI (open-source Python) for scrape, map, and crawl — no Firecrawl Cloud, no Docker required.”

ControlWhat it does
Enable SearXNG searchOff: SearXNG stops and web_search / firecrawl_search are hidden.
Enable Crawl4AI scraperOff: the Python sidecar stops and firecrawl_scrape / firecrawl_map are hidden.
Card titleCrawl4AI + SearXNG
SearXNG endpointDefault http://127.0.0.1:51452
Crawl4AI endpointDefault http://127.0.0.1:11235
Start web tools / RefreshStarts the sidecar; Refresh when it is already ready.

Chips

Search · Scrape · Map · Crawl

  • No Firecrawl Cloud and no Docker. Crawl4AI is a local Python process on port 11235.
  • The Firecrawl MCP server is forced disabled — you cannot turn that MCP card on. Search and scrape are built-in tools, not an MCP package.
  • Jarvis tools: web_search, firecrawl_search, firecrawl_scrape, firecrawl_map.
  • If the sidecar is down, scrape still uses a built-in fetch fallback. Search needs SearXNG running.

Connect other MCP servers (Google, GitHub, Discord, Telegram, Linear) from Settings → Extensions.

Troubleshooting

SymptomFix
Speech / Piper chips say Down after launchExpected until first use. Hold the orb or click Test Piper. Confirm Use local Piper / Use local Whisper are on.
Test Piper failsRe-run the installer so %LOCALAPPDATA%\VibeFuse\bin\piper.exe and jarvis.onnx exist. Voice/speed/volume are on Settings → Voice, not this tab.
Whisper never startsFirst listen starts whisper-cli. Need ffmpeg on PATH and a model in Settings → Voice. See Whisper STT.
HyperFrames Studio will not startInstall Node.js 20+. Enable HyperFrames Studio, then Start Studio. Workspace must stay C:\vibefuse\hyperframes.
Skill Seekers failed to startInstall Python 3.11+ and Git. Press Connect again so the venv rebuilds. Confirm C:\vibefuse\Skill_Seekers exists. There is no enable toggle — only Connect / Disconnect.
Crawl4AI not readyEnable Crawl4AI scraper, click Start web tools. Search can still work via SearXNG on 127.0.0.1:51452. Scrape falls back to built-in fetch.
Cannot enable Firecrawl MCPBy design. The Firecrawl MCP preset is forced off. Use the built-in web_search / firecrawl_* tools.
Engines start then the canvas stuttersThat is why launch no longer auto-starts Whisper, Crawl4AI, HyperFrames, or Skill Seekers. Start only the card you need.
Port already in useAnother process is bound to 4123 / 8935 / 4567 / 8765 / 11235 / 51452. Stop the extra copy, then Refresh or Start again.