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.
silence_split_seconds: tune caption line breaks from STT segments
Sume STT sentence segmentation can split on silence. silence_split_seconds takes 0.2 to 3 and returns gapless segments you can use as caption lines.
Time a 360p Omni draft vs 720p: separate queue wait from run time
Google says the 360p draft is up to 60% faster. Measure it on Sume with a Python script that splits pending time from in_progress time.
script_run for one TTS clip per sentence: limits to set first
Fan out one tts_create per sentence inside Sume script_run, with timeout_seconds, max_calls and max_paid_calls set, then wait on the child jobs with jobs_wait.
AI video API: rate limit or concurrency, which stops a batch first
Concurrency binds long before request rate on Sume. Free allows 120 writes a minute but 1 running job; Pro 300 and 4. How to size a batch against both.
wait_timeout_seconds 30 is not a 30-second video
The 30 in wait_timeout_seconds is how long your HTTP request may block, not how long a clip may run. A 30-second video job still needs a poll or webhook.
One source image, three Ideogram 4.5 edits in parallel (Python)
Run three different Ideogram 4.5 edits from the same source image at once on Sume, each with its own Idempotency-Key, instead of chaining them.
Audio detach errors: unsupported_media_type, source_not_found
Each Sume audio detach refusal code and its one-line fix: off-host URL, other workspace, not a video, empty range, no audio track, source too long.
SDK waitForJob reads per minute: the 2-second floor, and 10 jobs
The Sume SDK polls a job at least every 2 seconds, and a longer next_poll_after_seconds wins. That is up to 30 reads a minute per job. A webhook removes them.
Trim an AI-generated clip without a media import
Generated artifacts already live on media.sume.com, which is the host video-trim accepts. Use media-imports only for clips from outside Sume.
Which video models accept 30 seconds on Sume: read the catalog first
Seedance 2.5, Wan 3.0, h3-max-recast and Genjutsu list 30 s on the Sume Video Router; MiniMax H3 stops at 15 s. Read the catalog before you set duration.
AI video generator app: four server calls, key never on the client
An AI video app on Sume needs four server-side calls: submit, poll, download and balance. The key stays on your server, never in the browser or mobile app.
Last frame of an Omni clip: fixing frame_time_out_of_range
Asking Sume video-frames for t equal to the clip length fails with frame_time_out_of_range. Take the last frame at duration minus a frame, in Python.
Estimate a Sume video clip in Python: list x 1.25, rounded up
A runnable Python script that prices a clip for Wan 3.0, MiniMax H3, H3 Max and Gemini Omni 1.1 Flash with integer micros, so the cents match the bill.
Launching 24 ad variants on Pro: queued jobs, then queue_full
A Pro workspace runs 4 paid generations at once and accepts 24 in total. Plan a TikTok, Reels and Shorts variant launch around those numbers and 429 queue_full.
Check an image URL before you pay for image-to-video: Python preflight
A bad first-frame URL fails the job with an input download error. A short Python check of status, content type and size catches most bad URLs first.
Video job stuck in queued on Sume: limits by plan and queue_full
A queued video job is not a failure. Sume runs 1, 4, 8 or 20 jobs at once on Free, Pro, Startup and Scale, and queues the rest. 429 queue_full: queue full.
Last frame of an AI video clip: fix frame_time_out_of_range
To chain AI video clips, take the last frame of clip one as the first frame of clip two. Ask video-frames for a time below the duration, not equal to it.
OpenAI Realtime GA migration: drop the OpenAI-Beta header
Beta Realtime integrations must move to GA and stop sending OpenAI-Beta: realtime=v1. A short checklist, a code scan for the header, and where Sume jobs fit.
ElevenLabs TTS output_format: 192 kbps needs Creator, PCM needs Pro
The ElevenLabs text to speech reference ties 192 kbps MP3 to Creator and PCM or WAV to Pro. A format table, a fallback chooser in Python, and the seed range.
Browser voice app that starts Sume jobs: keep the key on your server
Voice apps run in the browser over WebRTC, but Sume keys belong on a server. A route handler that holds the key, allowlists models, reuses idempotency keys.
OpenAI TTS: 13 built-in voices vs 9 legacy ones, pick the model first
OpenAI's guide lists gpt-4o-mini-tts with 13 built-in voices and legacy tts-1 and tts-1-hd with 9. Formats, custom voice consent and disclosure, as a checklist.
Draft with GPT Image 2.5 Flare, finish with Sunburst: a two-pass edit
OpenAI pairs Flare with fast generation and Sunburst with editing precision. A Python two-pass on Sume's Image API that drafts, then refines the first result.
GPT Image 2.5 on ElevenLabs: 14 ratios plus auto. Sume lists 17
ElevenLabs offers 14 fixed ratios plus auto for GPT Image 2.5. Sume's normalized list has 17 plus auto. Read what each model accepts before sending one.
Gemini 3.8 Live audio: wrap 24 kHz PCM in WAV, resample to 16 kHz
Gemini 3.8 Live takes 16-bit 16 kHz PCM in and returns 24 kHz out. A Python WAV wrapper, an ffmpeg resample command, and the Sume detach settings that match.
whisper-1 or gpt-transcribe for subtitles: what OpenAI assigns to each
OpenAI recommends gpt-transcribe, gpt-4o-transcribe-diarize for speakers, whisper-1 for translation and subtitles. Plus the 25 MB limit and a chunking script.
End-user id on jobs: OpenAI safety identifier vs Sume metadata
OpenAI's Realtime guide asks for an OpenAI-Safety-Identifier header. Sume stores caller metadata on the job but does not send it to the provider. Use both.
Patching Supabase Postgres 17.11 vs Sume's 10-attempt webhook budget
Supabase's September 25 Postgres 15.19 and 17.11 releases fix 44 CVEs. A restart can outlast Sume's ten 30-second webhook attempts, so plan a redeliver.
Split a narration script by model character limit: Python
ElevenLabs lists a 40,000 character limit for Flash v2.5 and 10,000 for v4 and v4 Turbo. A Python splitter that cuts on sentences, with a concat step on Sume.
Test a faster-and-cheaper claim with your own timings and usage.cost
Luma's news page says Ray3.14 is 4x faster and 3x cheaper. A short Python script turns your own job timings and usage.cost values into two ratios you can trust.
MCP TS SDK 2.3 enforces one server per request: where state lives
TypeScript SDK 2.3.0 enforces one server instance per request. For a media tool that means job state belongs in job ids, as Sume's jobs_wait does.
MCP tasks/cancel vs Sume jobs_cancel: cancel works only before start
TypeScript SDK 2.3.0 adds tasks/get and tasks/cancel. Sume's jobs_cancel is narrower: it succeeds only before generation starts, then returns 409.
Next.js 16.3.8 fixes ISR cache poisoning: keep job pages dynamic
Next.js v16.3.8 fixes cache poisoning in SSG and ISR and Draft Mode leaks. Why a Sume job status page should stay dynamic and uncached, whatever the version.
Next.js 16.3.8 dev-server MCP disclosure and where the Sume key lives
Next.js 16.3.8 fixes a low-severity dev-server MCP disclosure and a high-severity image SSRF. How to keep a Sume API key server-side as you upgrade.
Instagram Reels API: a 100-posts-per-24-hours publish budget
The Instagram content publishing API limits an account to 100 API-published posts per moving 24 hours. A tested Python queue that spreads batch output under it.
Sora replacement smoke test: one prompt, three Sume models, in Python
A short Python script that sends the same 5-second prompt to seedance-2.5, wan-3.0 and gemini-omni-flash-1.1 on Sume and prints status, cost and URLs.
Mcp-Name header rules: rate-limit paid render tools at the gateway
MCP 2026-07-28 requires Mcp-Method and Mcp-Name headers on Streamable HTTP POSTs. A gateway can rate-limit paid render tools by name without reading the body.
Ask for the aspect ratio first: an MCP input_required round trip
MCP multi round-trip requests let a tool answer input_required to ask for an aspect ratio or spend approval before a render, with state in requestState.
MCP tasks extension and Sume job statuses: mapping for render tools
In MCP 2026-07-28 tasks are an extension polled with tasks/get. Map task handles, polling, update and list onto Sume job ids, status reads and jobs_cancel.
Windsurf now redirects to Devin Desktop: where Sume MCP setup lives
Windsurf redirects to Devin Desktop, and Cascade was removed in v3.9.19. Re-add Sume's hosted MCP URL there and verify it with mcp_health and tools_list.
Reproduce the same AI voiceover later: model id, voice, settings
To redo a narration line months later you need the model id, voice, language, format and settings. Sume's completed TTS job records them. A short routine.
Claude Files API is GA: review Sume outputs by job id and media URL
The Claude Files API left beta on Aug 19. When Claude reviews generated media, store the Sume job id and media.sume.com URL, not raw provider links.
Claude rejects forced tool_choice: steer generate_video by description
Claude Sonnet 5.5 and Opus 5.5 return a 400 for tool_choice any or tool, and thinking cannot be disabled. Steer Sume tool calls with descriptions and a dry run.
Luma callback_url or polling: which Sume job mode matches
Luma's API docs list keyframes, loop and callback_url for ray-2. On Sume the equivalent choice is job mode: async, sync up to 30 s, subscribe or webhook.
Which Node versions to test the Sume SDK on: 22, 24 and 26
Node 26.10.0 is Current, 24.21.0 and 22.23.3 are LTS. A small CI matrix and smoke test for code that calls the Sume API with fetch and WebCrypto.
Python MCP SDK v1 now gets security fixes only: use v2 for new clients
The Python MCP SDK v1.x line gets security fixes only; v2 added MRTR and the 2026-07-28 spec. What that means when you connect a client to Sume's hosted MCP.
Performance Max text limits: 30, 90, 90 and 25 characters, counted
Performance Max headlines run 30 characters, long headlines 90, descriptions 90, business name 25. A short script counts generated copy before upload.
Claude Code's 60 s MCP timeout: Sume sync waits 30 s, video polls
Claude Code 2.1.287 caps MCP tool calls at 60 seconds per server. Sume's 30 s sync wait and 55 s jobs_wait fit under it; a video render needs repeated waits.
Sort your MCP tool allowlist: stable order keeps the prompt cache warm
MCP 2026-07-28 asks servers for ttlMs, cacheScope and a deterministic tool order. On the client, sort your media tool allowlist the same way. Python snippet.