Developers
Building on the Sume API: keys and hosts, jobs and runs, polling, webhooks, idempotency, errors and rate limits, the SDK, the CLI, and MCP auth.
Ideogram 4 download: Hugging Face gate, login and first image
To run Ideogram 4 locally: accept the gate on Hugging Face, log in with hf, pip install the repo, run run_inference.py. The flags and the nf4 or fp8 choice.
claude plugin validate: read the hooks and calls lines of a mod
Before installing a Claude Code mod, run claude plugin validate and read its hooks and calls lines. Which combinations matter when Sume's MCP is connected.
Claude Code mod: stop paid Sume calls after N in a session
Write a Claude Code mod that counts paid Sume MCP calls with a tool.call hook, denies call N+1, and fails closed. Code, matcher, and what it cannot cap.
Are Claude Code mods safe with a Sume API key in your env?
Claude Code mods run unsandboxed and can read env vars and settings files. What that means for a Sume API key, the CLI config file and an OAuth session.
allowManagedModsOnly in Claude Code: does hosted Sume MCP still load?
allowManagedModsOnly keeps users' own Claude Code mods from loading. What it leaves alone, how a policy mod reviews the rest, and the Sume MCP connection.
Claude Code mod hook skipped and a paid Sume call still ran
A Claude Code mod hook that throws or times out is skipped, so the call runs anyway. Add .catch to fail closed, and know what a mod still cannot guarantee.
Agent crashed after submitting: find the Sume runs it started
List a Format's runs, read trigger.idempotency_key, and rebuild which rows already ran after a fresh agent session or crash, without paying for a duplicate.
AI video API fallback: retry on another model when a job fails
Chain seedance-2.5, seedance-2 and kling-3 on Sume: poll status_url, read the job error category, and resubmit the brief to the next model.
Mix Seedance, Kling and Omni clips in one video: shared aspect ratio
16:9 and 9:16 are the aspect ratios Seedance 2.5, Kling 3 and Gemini Omni Flash 1.1 all list. Set the Timeline output to match and plan before render.
H3 Max Recast job failed: what Sume refunds and what to retry
A failed Recast job releases its hold on Sume. Read the error category, fix input errors, retry queue errors with the same Idempotency-Key, never double-submit.
H3 Max Recast webhook: submit with mode webhook, verify the HMAC
Recast jobs run for a while. Submit h3-max-recast with mode webhook, then verify Sume's sume-v1 HMAC signature before you download the swapped video.
AI video person swap rejects long takes: find shots over 15 s
H3 Max Recast on Sume needs 5 to 30 seconds with no single shot over 15. Find the long takes with ffmpeg scene detection, then cut them before you pay.
Dub one Short into 8 languages: Python fan-out and the total cost
Detach and transcribe once, then run one TTS job and one render per language. A Python fan-out and the per-Short bill, from Sume's catalog rates.
H3 Max Recast seed: fal has one, Sume does not send it
fal's Recast API takes and returns a seed. Sume's Video Router accepts none, so each run is a new take. How to re-roll, what to vary, and what it costs.
Pydantic model to Sume output_schema: extra forbid, no defaults
Turn a Pydantic v2 model into a valid Sume Format output_schema: extra=forbid, nullable instead of defaults, and the SumeMediaFile reference. Tested.
Zod 4 toJSONSchema to Sume output_schema: nullable, not optional
z.toJSONSchema works for a Sume Format output_schema if you use nullable instead of optional. A tested table of what passes and what the validator rejects.
Zod 4 discriminated union for Sume job and run webhooks (TypeScript)
Parse Sume job.* and format.run.terminal webhooks with one Zod 4 discriminatedUnion: typed branches, degraded runs, oversized receipts. Tested with Zod 4.
Virtual try-on API: which Sume call returns an image, which a video
Need a try-on photo or a try-on clip? On Sume the two catalog try-on Formats return video; a still comes from the image API. The table, plus one call for each.
AI voice reads Spanish with an English accent: two causes to check
Spanish read with an English accent usually means the language field was omitted or the voice is tagged for another language. How Sume's TTS 409 works.
Latin American Spanish text to speech: es or es-MX on Sume?
Sume's TTS language is a free string and its voice library tags voices with plain es. What that means for Mexican, Argentine or Spain Spanish, and how to test.
FLUX 3 bounding box to a mask_url: Python region edit on Sume
FLUX 3 Image boxes use [top, left, bottom, right] on a 0-1000 grid. Convert one to an RGBA mask with Pillow and run the region edit on Sume's GPT Image 2.5.
Prefect 3 task retries for a Sume job: same Idempotency-Key
Retry a Sume image job in Prefect 3 without paying twice: a tested flow with retry_condition_fn, delay list and an order-derived idempotency key.
DBOS Python durable workflow for a Sume job: resume after a crash
Submit and poll a Sume image job in a DBOS workflow: step retries, order-derived Idempotency-Key and workflow id, tested with DBOS 3.2.0 on SQLite.
Transcribe a three-hour recording when the API caps at ten minutes
Streaming sessions end at an hour; Sume STT jobs take up to 600 seconds. Split with Timeline audio, transcribe 18 chunks, and stitch word times back together.
MAI-Transcribe-2-Streaming Realtime API: events vs Sume job URLs
Microsoft's streaming transcriber uses a WebSocket with delta, intermediate and commit events. Sume STT takes a file URL and returns a job. A side-by-side.
Unit test Sume webhook signature checks in pytest (Python)
A pytest file for HMAC webhook verification: tamper, rotation header, stale timestamp and empty secret, written against Sume's sume-v1 scheme.
GPT Image 2.5 seed: can you get the same image twice?
Sume lists seed in the schema but does not serve it: seed returns 400 unsupported_parameter. How to keep a look repeatable with references and masked edits.
Claude Code mod: ask before a paid Sume call (and claude -p)
A Claude Code mod can hold a paid Sume MCP call with $.ui.ask and show max_spend_usd in the question. In claude -p nobody answers, so it refuses. Code inside.
Is AI avatar video real time? How long a Sume job takes
A Sume avatar video is a job, not a live stream: it queues, renders, and you poll or take a webhook. What the sync wait caps at, and a Python polling loop.
Live AI avatar API: a Tavus conversation vs a Sume job
A live avatar API creates a room you join. Sume's Avatar API creates a job you poll. Field-by-field map of Tavus create conversation and Sume talking-video.
Log x-sume-request-id and Idempotency-Key on every call (Python)
A requests response hook that writes one JSON log line per Sume call: x-sume-request-id, idempotency key, error code and rate-limit headers. Tested.
Transactional outbox for paid API calls in Python (Sume)
Write the Sume request and its Idempotency-Key in the order's transaction, drain later: a tested Python outbox that survives crashes, 429s and 409s.
Translate an SRT and burn it in: Sume caption cues, limits, Python
Sume takes no SRT upload, but caption cues take the same text and times. A Python converter, the 200-cue and 60-second limits, and which fonts apply.
Contract-test Sume API responses against openapi.json (pytest)
Validate recorded Sume responses against the OpenAPI schema with jsonschema, including the OpenAPI 3.0 nullable fix. A tested pytest file and fixtures guide.
Use a local photo as a GPT Image 2.5 reference: it needs a URL
Sume's input_references take public HTTPS image URLs only; localhost and private URLs are rejected. Three ways to turn a file on disk into a usable reference.
Perplexity Decisions API as a publish gate for Sume output
Check a finished Sume Format image with Perplexity's Decisions API before it ships: base64 data URL, one yes/no question, a threshold, and a human-review lane.
FLUX 3 Image on Replicate: safety_tolerance 0-4 vs Sume
Replicate's FLUX 3 Image form has safety_tolerance 0-4, grounding, output_quality and 768sq-4k. Which of those inputs Sume's image API has, and what it returns.
Chinese text to speech API: set language zh or it reads as English
Sume TTS only guesses Korean and Japanese when the language is missing. For Mandarin send language zh and pick a voice tagged zh, then test one line.
Claude Code token usage vs Sume spend: two meters, one session
Claude Code's token totals and Sume's generation spend are billed in different places. Log tokens with a mod, read Sume's wallet with usage_get, and compare.
Batch GPT Image 2.5 from a CSV in Python: async jobs, safe retries
Render one GPT Image 2.5 packshot per CSV row on Sume: mode async, an Idempotency-Key per SKU, a small worker pool, and the 429 queue_full rule.
Pocket TTS API: run it yourself or call a hosted TTS API
Kyutai's Pocket TTS installs with pip and serves from localhost. If you want a hosted API with job URLs instead, here is the Sume request and what changes.
AI image API timeouts in Python: httpx above 30 s, 200 or 202
Sume's image endpoint holds a request up to 30 seconds, then returns 202 with a job. A runnable httpx example with a 40 second client timeout.
Canary 10% of video jobs to Sume before cutover: sticky bucketing
Moving video traffic off a shut-down API: hash a stable key into a percent bucket so each customer stays on one backend, and raise Sume's share in steps.
How to compare AI video models fairly: one prompt, three models, 480p
Submit one prompt to Seedance 2.0 Mini, Wan 3.0 and MiniMax H3 at 480p and 5 seconds on Sume, then judge the clips blind. A runnable Python test harness.
A dry-run flag for Sume API calls: print the request, skip the spend
Add DRY_RUN to code that calls the Sume API: build the body, key and spend cap, print them, and send nothing. Review a batch before it costs money.
Extend a video Veo didn't make: which API takes an uploaded clip
Veo 3.1 extension only accepts Veo-made videos. Here is what Gemini Omni and Sume accept instead when you need to continue a clip you uploaded.
Extend an AI clip by hand: last frame to the next clip, in Python
No extend endpoint? Pull a clip's last frame with Sume's video-frames API and feed it as the first frame of the next job. Python code under 30 lines.
Where to find a voice ID for Sume text to speech: voices_list and voi_
Sume TTS wants a voice UUID or a voi_ library id, never a voice name like alloy. Where to copy it from, and the error you get for a name.