Blog
Sume blog: agent-native video generation, Avatar 1.0 orchestration, and production APIs for builders and coding agents.
Audio version of an article with TTS: cost and the 20,000-char split
A 6,000-character article is one Sume TTS job at 29 cents. A 30,000-character one is two jobs plus a 1-cent concat, $1.45 total. Steps and code included.
Egress firewall for a Claude agent in CI: Sume hosts to allow
Claude Agent SDK v0.2.164 hardens CI egress firewalls. Which Sume hosts a render job needs: api.sume.com, mcp.sume.com and media.sume.com, and what each is for.
claude --strict-mcp-config: run CI with only the Sume server
--strict-mcp-config makes Claude Code use only servers from --mcp-config. A CI recipe for loading just Sume's hosted MCP, with a key, a wait limit and a cap.
Omni takes JPEG and PNG: gate each reference image URL in Python
Google lists JPEG and PNG for Gemini Omni image input. Check each reference_image_urls entry with a HEAD request, then submit the clean list through Sume.
Vercel AI SDK chunkMs timeouts and a 55 s Sume jobs_wait
ai@7.0.136 stops chunkMs and firstChunkMs when the model response ends. stepMs still covers the step, so size it for a Sume jobs_wait slice of up to 55 s.
Veo 3.1 Lite limits: no 4K, no extension, and cheaper options
Veo 3.1 Lite is $0.05 a second at 720p but has no 4K and no extension, and it shuts down October 22. See what the Sume catalog offers in its place.
Max AI video length: Runway Gen-4.5 10 s vs Sume's 30-second rows
Runway's Gen-4.5 runs 2 to 10 seconds and Omni 3 to 10. Sume's Wan 3.0 and Seedance 2.5 reach 30. Longest single clip by model, with the sources.
Agent SDK verbatim_prompts vs Sume Agent Completions input
The Claude Agent SDK verbatim_prompts option stops @path and slash expansion of user text. Sume Agent Completions keeps input as data in a file. Compare them.
claude -p --permission-prompts none: Sume tools in CI
With --permission-prompts none, prompts nobody can answer are denied. Which Sume tools still run? Read-only OAuth ones do; paid ones need mcp:write.
Voiceover or video: which is the bigger cost in a 15-second ad?
A 250-character voiceover is $0.012 on Sume TTS. A 15-second clip is $1.88 on Wan 3.0 at 720p. Where the money goes in a short ad, with the arithmetic shown.
Lint a Video Router request offline: reference limits in Python
Check reference_image_urls (10), reference_video_urls (3), reference_audio_urls (1-5) and the video_url rule in Python before POST /v1/video-router/generate.
Claude Code remote MCP backoff up to 30 s: retrying Sume safely
Headless Claude Code now backs off up to 30 s after remote MCP drops. A retried paid Sume call needs the same idempotency_key or you may pay twice.
Veo 3.1 lists 4, 6 or 8 seconds; Omni on Sume runs 3 to 10
Veo 3.1 lists 4, 6 or 8 seconds. Omni on Sume runs 3 to 10 seconds at 1080p or 4K, priced here to the cent.
How to cap an AI video agent's spend: five limits, in order
A video agent has five separate limits: model budget, turns, per-call max_spend_usd, run cap and plan queue. Which stops a runaway Sume job, and which does not.
MCP tools/list repeating a cursor: Claude Code 2.1.295 and Sume
Claude Code 2.1.295 stops asking a server for the same tools page up to 20 times. Sume's tools/list returns one page with no nextCursor, so it never loops.
Pick a Sume video model in code: audio refs, 1080p, 20 seconds
Filter GET /v1/video-router/models on reference_audios, resolutions and duration_seconds, then submit the survivor with reference_audio_urls (up to five).
Count queued and processing jobs with GET /v1/jobs before a wave
Page GET /v1/jobs with status=queued and status=processing, subtract from concurrency_limit, and submit only that many Seedance 2.5 or Omni clips.
claude -p --max-budget-usd does not cap Sume renders
Claude Code's --max-budget-usd is a client-side estimate of API spend in print mode. Pair it with max_spend_usd on every paid Sume call and leave headroom.
Codex worktrees run agents in parallel: one Sume queue
Codex CLI 0.162.0 adds managed worktree tools. Parallel agents in worktrees still share one Sume workspace: plan concurrency, queue and 429 queue_full apply.
Claude Code 2.1.288: MCP calls that ran twice, and paid Sume calls
2.1.288 fixed MCP tool calls sometimes running twice when a remote result was over 16 MB or unparsable. Why every paid Sume call needs an idempotency_key.
Luma Ray3.2 video-to-video costs 4.8x text-to-video at 540p
Luma's API lists a 5-second 540p video-to-video job at $0.72 and text-to-video at $0.15. The full table, and what a Gemini Omni edit costs on Sume.
Veo 3.1 takes 3 reference images; Omni on Sume takes 10 plus 3 clips
How many reference images does Veo 3.1 accept, and how many does Gemini Omni Flash take on Sume? Limits, tag syntax and a request example.
Runway Gen-4 Turbo 5 credits vs Gen-4.5 12: a 10-second clip costs
Runway's API lists Gen-4 Turbo at 5 credits a second and Gen-4.5 at 12, with credits at $0.01. What 10 seconds cost, next to the Sume rates.
Claude Code hooks on Sume tool calls: check the top-level guards
A Claude Code hook that gates paid Sume MCP calls should read the top-level idempotency_key, dry_run and max_spend_usd, and deny when input is unreadable.
Runway API Veo 3.1: 40 credits a second with audio, 20 without
Runway's API prices Veo 3.1 at 20 credits a second silent and 40 with audio, at $0.01 a credit. An 8-second example, and what audio costs on Sume.
Send a signed webhook.test with POST /v1/webhooks/test-deliveries
Point POST /v1/webhooks/test-deliveries at a new HTTPS endpoint to get a signed webhook.test before any video job runs, then check the sume-v1 signature.
Seedance 2.5 in 16:9, 9:16 and 1:1: one idempotency key per ratio
Fan one prompt out to three aspect ratios on /v1/videos, derive a separate Idempotency-Key for each ratio, and rerun the script without paying twice.
HappyHorse 1.1 [Image 1] tags vs Sume's <IMAGE_REF_0> in prompts
HappyHorse reference-to-video counts images from [Image 1]; Sume's Omni counts from <IMAGE_REF_0>. How to map a prompt without an off-by-one.
VS Code 1.141 background shells and a long Sume render
VS Code 1.141 tracks background shells for agent sessions. Start a Sume render, keep the job id, and use sume jobs watch instead of repeating the paid create.
Runway Enhance Frame Rate (24 to 120 fps) vs Sume's fps conform
Runway's Enhance Frame Rate converts video to 24-120 fps at 1 credit per 2 seconds. Sume's trim only conforms to 24, 25, 30 or 60 fps.
Omni Flash defaults to 720p and 16:9: set both on every request
Google's Omni docs default to 720p and 16:9; Sume's Auto defaults to 720p and 8 seconds. Why to send resolution, ratio and duration every time, with prices.
Runway Model Router vs sume/auto: how each picks a video model
Runway's Model Router saves a cost, latency or quality config and reuses it by configId. Sume's sume/auto picks the family and never says which. Differences.
Stop a Seedance 2.5 batch on SumeInsufficientCreditsError
Submit a batch with createVideoGeneration and one idempotency key per item. On a 402, stop the loop, top up, and rerun the same script without paying twice.
Claude Code closes ws MCP messages over 16 MiB: Sume sends URLs
Claude Code 2.1.295 closes a websocket MCP connection on a message over 16 MiB. Sume hosted MCP returns media.sume.com URLs for results, not file bytes.
HappyHorse 1.1 image-to-video: 300 px, 20 MB and ratio 1:2.5 to 2.5:1
Alibaba's HappyHorse 1.1 image-to-video accepts JPEG, PNG or WEBP of at least 300 px, up to 20 MB. Prepare a first frame, or use an Omni image_url on Sume.
Find exhausted Sume webhook deliveries and redeliver them in Python
Page GET /v1/jobs for completed and failed jobs, keep webhook_delivery.status exhausted, and call POST /v1/jobs/{id}/webhook/redeliver for each.
Codex custom Responses provider: can it point at Sume?
Codex CLI 0.162.0 adds custom Responses-compatible provider config. Sume's Agent Completions is not a model endpoint; use Sume's hosted MCP from Codex instead.
Stopped a Sume MCP call in Claude Code? The job still runs
Stopping a long Sume MCP wait in Claude Code does not stop the render. The job keeps running and billing; find the job id, then wait again, read or cancel.
Omni reference videos: read Sume capabilities, not Google's note
Google's Omni page says multi-video reference is unsupported; Sume's docs list up to three reference clips. Read the catalog entry before you send them.
Gemini Omni Flash on Runway is 720p only; Sume sells 360p to 4K
Runway's Omni Flash model outputs 720p at 10 credits a second. Sume offers four resolutions of the same Google model. Which to use for drafts and finals?
claude --fallback-model when overloaded: Sume paid calls
--fallback-model switches Claude Code to another model when the main one is overloaded. Sume's spend caps and idempotency keys are independent of the model.
claude -p --max-turns exits with an error: Sume job in flight
When --max-turns ends a CI run, a Sume job may still be rendering. Keep the job id in the output, resume with jobs_wait, and never repeat the paid create.
Gemini Omni Flash 1.1 price ladder: 4K is 10x the 360p rate
Omni Flash is $0.03 to $0.30 per second on fal; Sume bills list x 1.25. Price of the 8-second default, a 10-second max clip and a minute at each resolution.
Vidu task states mapped onto Sume job statuses in Python
Map Vidu task states (created, queueing, processing, success, failed) to Sume queued, processing, completed and failed so an old poller keeps working.
Ideogram 4.5 price by quality: fal list vs Sume, per 1,000 images
Ideogram 4.5 is $0.03, $0.06 or $0.22 per image on fal by quality, size-independent. Sume bills list x 1.25: $0.0375, $0.075, $0.275. Per-1,000 table.
Edit a 20-second AI video: Luma, Aleph 2 or Omni's 10-second input?
Omni edits clips of 10 seconds or less; Runway Aleph 2.0 edits 2 to 30 seconds; Luma lists video-to-video at 10, 15 or 20. Editing longer clips on Sume.
MCP tool description length: Claude Code 16,384 vs Sume
Claude Code 2.1.295 raised the tool-search description cut from 2,048 to 16,384 characters. What that means for a Sume MCP connection, and how to check it.
Claude Code 2.1.295 MCP changes: a checklist for Sume
Three MCP changes in Claude Code 2.1.295 (Oct 8) and what each means for a connection to Sume's hosted server. A short table to run through after upgrading.