Phone installation and spoken answers

Take the interview with you.

IdeaForge can be installed as a home-screen web app, accept one dictated answer at a time, or run a continuous hands-free loop that reads each question aloud and listens for a finish word.

The IdeaForge setup screen with provider, API key, a model for the
                  questions and a model for the wrap-up, dictation, hands-free finish word,
                  connection check, and start controls.
Installation changes presentation and storage durability, not the data model. The installed app still stores ideas in that browser’s local site data and still needs a provider connection for grounded questions and synthesis. There is no account-based sync.

Install on Android

  1. Open IdeaForge in Chrome on the Android phone.
  2. If the setup screen offers Install IdeaForge, select it and accept the browser prompt.
  3. If no in-app install button appears, open Chrome’s menu and choose Install app or Add to Home screen.
  4. Launch IdeaForge from its home-screen icon.

Chrome’s browser recogniser is the preferred no-key dictation path when its behavioural probe succeeds. Android may end long recognition sessions even when continuous listening was requested, and it sends every growing guess at a sentence as a separate, already-final result. IdeaForge restarts the recogniser and keeps only the newest guess, so one answer is neither silently truncated nor repeated back as “I I want I want to…”. Because those guesses arrive marked final, hands-free mode waits a moment before trusting a trigger word or command in one.

Install on iPhone or iPad

  1. Open IdeaForge in Safari.
  2. Tap Share, then choose Add to Home Screen.
  3. Leave Open as Web App enabled if Safari offers it.
  4. Tap Add, then launch IdeaForge from the new icon.
Use hosted transcription for dictation in the installed iPhone app. The browser recogniser can exist, start without an error, and then produce no events. IdeaForge tests for proof of life and hides the microphone rather than leaving a capture hanging. Choose Groq Whisper or OpenAI transcription in Settings and supply the corresponding key.

What installation does offline

The service worker caches the application shell for a later launch. If the provider is unavailable, the interview can continue from the built-in question bank instead of failing to open. Provider requests are never placed in that cache, and a refined prompt still needs a working provider at wrap-up.

Back up the Ideas library before removing site data or moving devices. An installed icon does not turn local browser storage into cloud storage. See Ideas library for transfer and recovery.

Choose a provider a phone can reach

On a phone, localhost and 127.0.0.1 refer to that phone. They do not reach Ollama or LM Studio running on a laptop elsewhere on the network. IdeaForge also refuses arbitrary LAN addresses rather than opening a route that would weaken its provider allowlist.

A hosted provider is therefore the normal phone setup. If you need a local model, follow the browser, CORS, and local-network constraints in Local models and Docker. Safari cannot make an insecure HTTP request from a hosted HTTPS page to a loopback model server; running the app locally as well is the practical route.

Voice support matrix

Voice paths offered by the current app
Environment Browser recogniser Hosted transcription Hands-free
Android Chrome, tab or installed Used when the live probe succeeds; long answers survive recogniser restarts, and repeated partial guesses are merged into one answer. Available with a supported microphone, recorder, and transcription key. Shown when voice input and speech synthesis are both available.
iPhone or iPad in Safari Offered only if the recogniser proves it is alive in that context. Available with a transcription key and microphone permission. Conditional on the selected input path and speech synthesis.
Installed iPhone web app Do not rely on it; the known silent recogniser is treated as unavailable. The supported dictation route: choose Groq Whisper or OpenAI. Available when hosted transcription and speech synthesis initialise successfully.
Edge or Firefox The current app treats known dead or network-failing recognisers as unavailable. Fallback is available where microphone recording is supported. Conditional; the checkbox stays hidden if either input or speech output is missing.
Any supported browser, keyboard only Not used. Not used. Not available, but the full typed interview remains available.

The app probes behaviour rather than trusting the presence of a speech API. It requires the recogniser to emit evidence that it started; otherwise it selects the configured recorder-and-transcription fallback or withholds voice controls.

Starting the recogniser is also what asks for the microphone, so the first interview on a device waits for that answer rather than timing the engine. Declining, or a microphone already blocked for the site, is reported as blocked and is not treated as a broken browser: allow it in your browser’s site settings and start an interview again. Only a recogniser that fails on its own account is remembered as unavailable.

Configure dictation

  1. Open Settings.
  2. Under Dictation, choose the browser recogniser, Groq Whisper, OpenAI, or keyboard only.
  3. If the transcription provider differs from the question provider, enter its key in the separate transcription field. When the same supported provider supplies both, IdeaForge reuses that provider key.
  4. Choose an uncommon Hands-free finish word. The default is over; the app warns when a very short or common word is likely to cut an answer off.
  5. Start or resume the interview and allow microphone access when the browser asks.
Hosted transcription sends recorded audio to the service you selected. Its key is stored with the other encrypted credentials. Choose keyboard only if you do not want to send recorded audio for transcription.

Dictate one answer

  1. Select Answer out loud.
  2. Speak your answer. Live words appear when the browser recogniser supplies them.
  3. Select Stop listening when you are finished.
  4. Review and edit the transcript in the answer box.
  5. Select Send.

A manual microphone tap takes control away from hands-free mode. The transcript is not submitted automatically, so you can correct names or technical terms first. Dictated answers are marked in the exported transcript.

Run a hands-free interview

The hands-free checkbox appears only when IdeaForge has a working voice-input path and the browser can speak text. Once enabled, the loop reads the bridge and question aloud, reads the suggested answers after it as a separate sentence, listens, submits the answer, and continues without waiting for a tap. The suggestions are examples of the shape of answer the question wants, not a menu to choose from: answer in your own words. A question with no suggestions is read on its own.

  1. Enable Hands-free — read each question aloud, with suggestions, and listen for the answer.
  2. Wait until the question has finished playing before speaking.
  3. Say your answer naturally; a pause does not finish it.
  4. End the answer with the configured finish word, for example over.
  5. Listen for the next question or a spoken wrap-up offer.

Spoken commands

Commands recognised when they are the whole utterance
Say Result
repeat that Reads the current question again without recording an answer.
skip this one Records the turn as skipped and moves to the next question.
scratch that Discards the current, not-yet-submitted capture and listens again.
wrap it up Finishes once the interview has enough recorded answers; very early requests are refused.

Commands must be the whole utterance. A sentence such as I’d skip this one if I could remains an answer. The finish word is terminal-only, so we went over budget does not end a response unless the configured word is actually at the end.

Spoken wrap-up

When coverage, exhaustion, or a turn limit suggests stopping, IdeaForge asks whether it should write the idea up. Say a clear yes to wrap, or say keep going. Unclear replies are not treated as consent. After synthesis, the app reads a successful refined prompt aloud in manageable chunks and leaves the Markdown on screen. If synthesis fails, it says that the answers are saved and the degraded document is available on screen.

Permissions and browser controls

Permissions you may encounter
Permission or capability Why it is needed If unavailable
Microphone Captures dictated answers for browser recognition or hosted transcription. Voice is withheld or an explicit denial error is shown; typing remains available and hands-free stands down.
Speech synthesis Reads questions, recovery prompts, and the final result in hands-free mode. Single-answer dictation may still work, but the hands-free checkbox is hidden.
Local-network access in Chrome A hosted page may need browser approval before it can contact a loopback model server. Use a hosted provider or run the app locally with the model server.
Local storage Keeps sessions, encrypted credentials, and device preferences. IdeaForge reports failed session saves; export before closing the tab.

With recorder-based transcription, IdeaForge keeps the microphone stream for the current interview after it is first opened, avoiding a new permission prompt for every answer. It releases the stream when the voice controller is torn down, such as when you leave the interview or complete the wrap-up.

What hands-free does when it cannot hear you

Do not rely on hands-free controls in a situation where a missed question or transcription could create a safety risk. The recovery loop is designed not to strand the interface waiting for a tap, but speech recognition and network transcription can still mishear, skip, or fail.