Blog
Sume blog: agent-native video generation, Avatar 1.0 orchestration, and production APIs for builders and coding agents.
claude plugin validate --strict in CI: what it checks in a Sume entry
Run claude plugin validate with --strict so warnings fail the build. The MCP checks that hit a Sume entry: undeclared keys, bad URLs, literal credentials.
Where a Claude Code plugin keeps Sume state: CLAUDE_PLUGIN_DATA
A plugin's data directory survives updates, while its root changes. Where to put a Sume call log or cache, and when uninstalling deletes it.
Plugin headersHelper cannot read user_config: where the Sume key goes
Claude Code rejects ${user_config.*} inside an MCP headersHelper. Use the headers field for a Sume API key, or fetch the key inside the helper script yourself.
Name your Sume plugin right: claude- and anthropic- prefixes fail
claude plugin validate errors on plugin names that pass as Anthropic's own. Pick a name for a Sume hosted MCP plugin that validates, and see what still loads.
Plugin lists Sume twice? In Claude Code the later declaration wins
Claude Code loads a plugin's .mcp.json first, then mcpServers from plugin.json; a name declared later replaces an earlier one. Keep one Sume entry.
Claude Code plugin userConfig: prompt for the Sume API key once
A plugin can ask for a sensitive value and reuse it as ${user_config.KEY} in an MCP header. A Sume plugin manifest that keeps the key out of settings.json.
Ship a Sume Claude Code plugin off by default: defaultEnabled false
A manifest can set defaultEnabled to false, so a Sume plugin with paid tools starts disabled until someone turns it on. What persists after updates.
Hook matcher for a plugin-bundled Sume server: mcp__plugin_ names
A Sume server shipped inside a Claude Code plugin gets a longer tool prefix than one added by hand. Write the PreToolUse matcher for it and test it first.
GPT Image 2.5 inpainting with mask_url on Sume: steps and cost
Edit part of an image with mask_url and up to 16 references on openai/gpt-image-2.5. Billed about $0.066 an image on Sume. Steps and the limits.
Sume image_size vs aspect_ratio: 1080x1350 and GPT size limits
image_size wins over aspect_ratio on Sume's image API. On Nano Banana 1080x1350 becomes 4:5 plus target pixels; GPT needs multiples of 16. Rules and checks.
How Sume bills an image batch: list price times 1.25, worked out
Sume's image bill is the provider list price times 1.25, times n. Worked batches for Seedream, Nano Banana 2.1, Qwen and Ideogram 4.5, plus how to check it.
Seedream 5.0 Lite search and thinking: what Sume's API passes
ByteDance describes online search and thinking for Seedream 5.0 Lite. Sume's row lists 9 ratios at about $0.04375 and empty provider.options. What that means.
Python: find Sume image models that list a ratio like 8:1 or 4:5
A 15-line Python script reads GET /v1/images/models and prints every Sume image model that lists a given aspect ratio, so you stop guessing before a 400.
Phonon-2 is CC-BY-4.0: what attribution means for shipped transcripts
Phonon-2 allows commercial use under CC-BY-4.0 with attribution. What that asks of an app that ships transcripts, and where hosted Sume STT differs.
Moroccan Arabic speech to text: open-source Voxtral model vs Sume STT
A Moroccan open-source stack pairs language ID with Voxtral ASR for dialect code-switching. What Sume STT offers for Arabic audio, and what it lacks.
Transcribe 100 ten-minute clips: Scribe v2, gpt-transcribe, MAI, Sume
1,000 minutes of audio costs about $3.67 on Scribe v2 list, $4.50 on gpt-transcribe, $9 on MAI intro pricing, and $10 on Sume STT. The sums and the caveats.
Fireflies Talk dictation vs transcribing a recorded file by API
Fireflies Talk is free, unlimited desktop dictation. A recorded file needs a different tool. How to tell which you need, and what Sume STT costs for the file.
Does dictation audio leave your device? Fireflies, Phonon-2, Sume
Fireflies Talk keeps finished dictations local but sends audio to its STT provider. Phonon-2 runs on device. Sume STT is hosted. A sourced privacy comparison.
Phonon-2 (164 MB, English) or hosted Sume STT for a 10-minute file?
Phonon-2 is a free 164 MB English-only on-device model. Sume STT is hosted, multilingual, about 1 cent a minute. When each fits, with sourced numbers.
JSON2Video credits per second and 4K at 4x vs Sume Timeline
JSON2Video deducts one credit per video second and four times that for 4K. Sume Timeline charges $0.10 per rounded output minute. What a minute costs on each.
Placid 1 credit and Bannerbear 2 per video second vs Sume
Placid charges one credit per second of video and Bannerbear two per second of animation. Sume Timeline charges $0.10 per output minute. Convert them.
Inworld buys Ultravox, voices move to TTS-2: pin voice and model IDs
Inworld's Ultravox deal moves built-in voices to Realtime TTS-2 at no extra cost. Why to pin voice and model ids wherever you generate speech.
Cartesia sonic-3-latest and sonic-3-preview aliases: fix model strings
Cartesia lists sonic-3-latest and sonic-3-preview as deprecated aliases. What each maps to, which ids Sume's TTS Router accepts, and a safe migration.
TypeScript 7.0.2 with @sume-com/sdk 0.2.0: the tsconfig that compiles
We compiled a Sume SDK 0.2.0 job-status call with TypeScript 7.0.2. It needs nodenext resolution, types set to node and a module-type package.json.
npm audit signatures on @sume-com/sdk 0.2.0: what it proves, what not
Run npm audit signatures on @sume-com/sdk 0.2.0 before deploy. It checks the registry signature; the registry lists no provenance attestation for this version.
MCP C# SDK 2.0 for a .NET client: check the protocol header with Sume
The C# MCP SDK 2.0 targets the 2026-07-28 spec and keeps 1.x APIs compiling. What to test before pointing a .NET client at Sume, which lists 2025 versions.
Lambda get_remaining_time_in_millis: stop polling a Sume job in time
A Lambda that polls a Sume job should check get_remaining_time_in_millis before each sleep and return the job id so a later invocation can resume. Python.
AWS Lambda's 3 s default vs POST /v1/images' 30 s sync wait
POST /v1/images waits up to 30 seconds by default, but a Lambda defaults to 3. Send mode async with an Idempotency-Key and return the job id instead.
A 170-word avatar script in 34-word sentences plans 70 s: refused
Why a 170-word script of five 34-word sentences needs 70 seconds in Avatar 1.0, past the 60-second cap, and how to split it into two 35-second videos.
Avatar script: 33 words = one 12 s clip, 34 words = two 7 s clips
A Sume Avatar 1.0 sentence of 33 words fits one 12-second clip. Add a 34th word and it splits into two 7-second clips. The numbers, with code.
Hedra's developer API is in private beta; the Sume routes open to call
Hedra lists its developer API as a private beta. Sume documents its avatar routes publicly: lip sync, Fabric, talking video, motion control. Routes and limits.
Bun 1.4.2 fixes an AsyncLocalStorage leak: per-request Sume keys
Bun 1.4.2 fixes an AsyncLocalStorage leak from 1.4.1. Carry a per-request Sume Idempotency-Key in the store, and check your Bun version first.
FLUX 3 Image is on BFL's page: which FLUX id does Sume call?
BFL lists FLUX 3 Image, FLUX.2 Max and Klein. Sume's image router calls FLUX.2 pro and flex. Ids, prices at $0.0375 and $0.0625, ratios and edit limits.
Vidu Q4 Preview aspect ratios vs Sume video rows: who offers 21:9, 4:3
Vidu Q4 reference-to-video lists 1:1, 9:16, 16:9, 3:4 and 4:3. On Sume Seedance and H3 add 21:9, Kling 3 has three ratios, and Omni only two. No row lists 4:5.
Timeline render of 61 seconds bills 2 minutes: trim 1 s, pay half
Sume's timeline render is $0.10 per started output minute: a 61-second video costs $0.20, a 60-second one $0.10. Table from 6 s to 10 min, with the arithmetic.
A 340-product catalog: four Sume bulk queues of 100, 100, 100 and 40
A bulk queue holds 1 to 100 items, so 340 product videos need four queues. Split the list, key each queue, and remember workspace concurrency still applies.
A 58-word six-sentence avatar script becomes a 10 s and a 12 s clip
How Sume Avatar 1.0 plans a 58-word script of six sentences: 27 words in 10 seconds, then 31 words in 12 seconds, 22 seconds total, and why the cut falls there.
Next.js 16.4 Cache Components: keep Sume job polls out of use cache
Next.js 16.4 enables Cache Components in new apps. We built a Sume status route both ways: a plain GET handler stayed live, a use cache helper froze one answer.
Five Sume image models are text-only. What to use with a photo
Soul, Imagen 4 Fast and Ultra, Recraft V4 and Qwen Image Max take no reference images on Sume. Price and ratio of each, plus the cheapest edit-capable swap.
Codex background tasks keep their turn's permissions: paid Sume calls
Codex 0.161.0 background tasks retain the permissions of the turn that started them. What that means for a Sume render started from a background task.
Sume 429: retry-after header first, then the body field, then backoff
How long to wait after a Sume 429? Use the retry-after header, then error.retry_after_seconds, then jittered backoff. A Python stdlib function with checks.
Vidu Q4 Preview takes 3 audio references: which Sume rows take audio
Vidu Q4 Preview accepts up to 3 mp3 voice references of 3 to 12 s. On Sume, Seedance 2.x, Wan 3.0 and the MiniMax H3 rows take audio references; Omni does not.
Vidu Q4 Preview result URLs last 24 hours: save the file the same day
Vidu says Q4 Preview output URLs stay valid for 24 hours. Copy the file the day you render it, and see what Sume documents about its job result URL.
YouTube Shorts: 3 minutes at 1080p, rendered as 180 s on Sume
YouTube's help page says Shorts run up to 3 minutes at up to 1080p. A 180-second Sume Timeline render bills 3 minutes at $0.10, so $0.30. Settings inside.
Clock changes vs clock drift: what breaks Sume webhook checks
Sume webhook timestamps are Unix seconds, so time zone changes cannot break the 5-minute replay check; a drifting server clock can. A check to tell them apart.
Kling Image 3.0 lists native 4K: Sume's image resolution tiers
Kling's page lists Kling Image 3.0 with native 4K. Sume's image API takes resolution 512, 1K, 2K or 4K per model; check the catalog before asking for 4K.
Vidu @subject tags vs Sume Omni <IMAGE_REF_0> reference tokens
Vidu Q4 reference prompts use @[subject_name]. On Sume only Gemini Omni Flash 1.1 documents prompt tokens: <IMAGE_REF_0>, <VIDEO_REF_0>.
Python: poll a Sume bulk queue, then list the item indexes to redo
A short Python loop that polls GET /v1/format-run-queues/{id} with backoff, waits for completed, and lists failed or canceled item indexes for a second queue.