Overview#

MaxiMind puts Claude, GPT and Gemini models, plus speech, voice and vision services, behind one account and one HTTP door, with per-person and per-key access.

What MaxiMind is#

MaxiMind runs on a single Apple Silicon machine and exposes one HTTPS surface at https://maximind.dev. Behind that surface are:

• Three model engines: Claude (engine claude), GPT via Codex (engine codex) and Gemini via Antigravity (engine agy). MaxiMind balances the load, applies the model rules and reports which engine and model actually answered.
• Speech-to-text with Whisper large-v3 (async jobs, URL and YouTube transcription, an OpenAI-compatible transcription endpoint) and AI-cleaned subtitles.
• Text-to-speech with Vietnamese Northern, Southern and Central dialects and American and British English accents.
• Vision input: images, PDFs and text files attached to a generation request or sent as image_url parts on the OpenAI-compatible chat endpoint.
• Media utilities: platform captions, video, channel and playlist metadata, web scraping, and large-file sharing on the owner's NAS.
• Operational tools for the owner's team: a sign-in codes relay, an approve-then-run AI task queue and admin pages.

Every heavy task (chat, TTS, transcription) runs one at a time through a single queue. A slow response usually means "waiting in line", not "down". Long work returns a job_id immediately and you poll for the result.

Image generation is currently withheld on every public surface. See Images.

The three products#

How to get access#

There are two identities on MaxiMind:

1. An approved account. You sign in with Google in the browser. New accounts start as pending and are rejected with HTTP 403 until the owner approves them in the admin area. An approved account can use the transcription web app at /app/, the workspace pages and, once granted, MaxiMind Code.
2. An API key. The owner issues one key per app or site from the admin area. Send it as x-api-key: <key> or Authorization: Bearer <key>. Keys created from 2026-07-22 onward start with no abilities; the owner ticks the capabilities each key may use (generation, engines, media, worker profiles).

A MaxiMind Code device credential and a Canvas token are neither of these: they work only on their own endpoints and are refused everywhere else. See Authentication.

Hostnames#

https://maximind.dev is the primary hostname. https://maximind.macsvff.uk is the same server and stays supported; every endpoint, key and account works on both names. Examples in these docs use https://maximind.dev.

The public edge cuts a single request at about 100 seconds and 100 MB. Use job mode and polling for anything that takes longer. See Limits.

Start here#

• Quickstart: first request in three steps.
• Authentication: headers, browser login, per-key capabilities.
• Models: engines, auto tiers, live catalogue, denied models.
• Text generation: POST /api/ai/generate in job mode.
• Chat completions: the OpenAI-compatible /v1/chat/completions.
• Speech-to-text: Whisper jobs, URL and YouTube transcription.
• Text-to-speech: voices and synthesis.
• Vision: attaching images, PDFs and text.
• Images: why image generation is withheld.
• Files: NAS file sharing and media endpoints.
• Workspace: sign-in codes relay, task queue, Canvas, admin pages.
• MaxiMind Code and MaxiMind Chat.
• Limits and Errors.

Live references served by the server#