Docs / Reference / Limits

Limits

Admission and rate limits, request sizes, credits and reference USD prices, and how model capacity is shared.

One machine, one queue

MaxiMind runs on one machine that runs one heavy AI task (chat, TTS, transcription) at a time. Requests go through a single FIFO queue. A slow response means "waiting in line", not "down". Long jobs return a job_id immediately; poll for the result. status: "queued" is normal.

Admins can watch the queue at GET /admin/queue.

Public edge

Limit Value
Request timeout about 100 s (an exceeded request returns 524)
Body size 100 MB

Use job mode (/api/ai/generate with async: true, /jobs, /api/ai/worker) or streaming (/v1/chat/completions with stream: true, which sends keepalives) for anything that may take longer. Uploads over 100 MB (NAS, transcription) must come from the owner's own network.

Admission: model requests

One shared admission controller covers /api/ai/generate, /api/ai/agent and /v1/chat/completions:

Scope Limit
Globally 8 running or queued requests
Per API-key owner 2

A saturated request receives HTTP 429 with Retry-After: 5 before MaxiMind retains the prompt or spends model capacity.

MaxiMind Code: two active model requests per account, eight process-wide (shared with every other AI surface). A third window waits up to three minutes for a free slot instead of failing.

Attachments: two attachment requests run concurrently, eight admitted in total.

Sandboxed worker: 8 admitted jobs per key per day (charged on acceptance, including preparation failures); an account-wide daily budget of 20 worker attempts with alerts at 50, 80 and 95 percent; a two-slot attachment-preparation lane with a 120 s aggregate deadline; a concurrency-one Antigravity lane.

Request sizes

Surface Bound
POST /api/ai/generate 512 KiB raw JSON (rejected while streaming, including chunked bodies)
POST /api/ai/agent 256 KiB raw JSON; input up to 100,000 characters
POST /v1/chat/completions 60 MiB encoded JSON (base64 image parts); 400,000 characters for an explicit Claude pin, 100,000 for other engines and auto, including system prompts
POST /api/code/v1/chat/completions 4 MiB, text only
Attachments 10 files, 15 MiB per file, 40 MiB total, 60 s per download; 80,000 extracted text characters; 100,000-character attachment prompt; images 25 million pixels and 100 frames; PDFs 6 pages at 110 dpi when text is unusable
Worker brief 32,000 characters; PDF derivation 12 page images / 50 MB or 200,000 bytes of text; artifacts 20 files, 25 MB each
POST /api/ai/tts speed 0.5 to 2.0
POST /api/nas/upload 5 GB hard cap (413 above)
/v1/channel, /v1/playlist limit 1 to 200
/v1/web/scrape timeout 1 to 60 s
Canvas device-flow bodies 4096 bytes; image prompt 32,000 bytes, body 65,536 bytes

Requests above a character bound fail explicitly; history is never silently truncated.

Rate limits

Surface Limit
POST /api/code/devices/start rate-limited per address (10-minute code, 3 s poll interval)
Canvas device flow start 30/min, poll 300/min, verification 20/min, refresh 120/min, one poll per device per 3 s (persisted, per ASGI peer address)
Canvas image submission 6 per minute per account
GET /admin/queue safe to poll every 1 to 2 s

Job records (generate, agent, transcription) expire one hour after completion.

Engine timeouts

Engine Timeout
Piper TTS 180 s
mlx-audio and VietVoice TTS 300 s
yt-dlp channel or playlist listing 120 s (504 when exceeded)
Worker model 900 s default, 1,800 s hard cap

Typical latency: Antigravity Pro 40 to 95 s, Flash 5 to 20 s; Codex 5 to 10 s; Claude 5 to 15 s, long answers up to 90 s. A one-tool cold worker job takes 90 to 110 s.

Credits

Every billable call is logged to a usage table and attributed to an actor (API-key label, user email, Local Network or anonymous). Minimum recorded is 1 credit. Credits meter usage; they do not block or enforce a balance.

Action key Credits Cost for Recorded today
transcript_native 1 One transcript from a platform's free captions Yes
transcript_generated_per_min 2 Whisper transcription on /v1/transcript, per audio minute (about 1 per 30 s, min 1) Yes
video 1 One video's metadata Yes
channel 1 One channel listing Yes
playlist 1 One playlist listing Yes
web_scrape 1 One web page Yes
image 5 Reserved legacy rate; image submission withheld No
tts 1 One spoken clip Yes
translate_per_min 30 Transcript translation, per minute Intended only
file_transcribe_per_min 2 Uploaded file transcription, per minute Intended only

Agent tasks (/api/ai/agent) cost 2 credits. The /jobs*, /transcribe-url, /youtube-to-srt and /audio/transcriptions handlers do not currently record usage. NAS routes are free. Model text work (/api/ai/generate, /v1/chat/completions, Code, worker) is logged by surface, engine, account, model and outcome rather than by these credit keys.

Read the live price list from credit_costs in GET /admin/usage.

Reference USD cost per task

Reference sell price $0.003 per credit.

Task Endpoint Credits USD
Transcript (free captions) /v1/transcript (native) 1 $0.003
Transcript (Whisper) /v1/transcript (generate) 2 / min $0.006 / min
Video info /v1/video 1 $0.003
Channel or playlist listing /v1/channel, /v1/playlist 1 $0.003
Web scrape /v1/web/scrape 1 $0.003
Image generation /api/ai/image — Withheld; 503 before charging
Voice (TTS) /api/ai/tts 1 $0.003
Translation per minute 30 / min $0.09 / min

Shared model capacity

  • Model capacity is shared across every surface (API, chat completions, MaxiMind Code, worker) and every user and window. The MaxiMind Code banner says so.
  • The providers do not report remaining quota. Token and cost counters in clients are not quota measurements, and usage on the OpenAI-compatible endpoint is zero. MaxiMind sidelines an account after a quota error and rotates to the next; a watcher probes every account every 30 minutes and alerts the owner when one is failing or signed out.
  • When no engine can answer: /api/ai/generate returns status: "failed" (with subscription_only: true), /v1/chat/completions returns 503, MaxiMind Code shows "no subscription could answer". No pay-per-use API is used on these surfaces.