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.
Gemini CLI 0.62 MCP titles: reading Sume's tool names
Gemini CLI v0.62.0 formats MCP tool call titles as structured signatures. Sume tool ids are underscore names such as generate_image; dotted aliases map to them.
Retrain a cloned voice without breaking old videos
HeyGen keeps a voice ID when a clone is retrained. On Sume, an avatar is referenced by a stable handle, and each text-to-speech job records its voice and model.
Track AI video spend per job: Synthesia Billing API vs Sume job cost
Synthesia added a Billing API and auto top-up. On Sume, every finished job reports usage.cost, so you can keep a per-job ledger without a billing endpoint.
Cloudflare Sandbox SDK 1.0: run the Sume SDK inside one
The @sume-com/sdk has no runtime dependencies and needs only fetch and WebCrypto, so it can run in a sandbox. Pass the key as an env var, server-side only.
Voice replication API audit checklist before you switch
Gemini 3.8 Flash TTS is GA with voice replication and 150+ voices. Before switching providers, audit these items against Sume's live catalog.
Retry hints in the body or a header: reading Sume's Retry-After
Notion repeats Retry-After in response bodies. Sume can send a retry-after header on 429s. Here is how to retry each Sume error code safely.
Fetch tool to media input: which URLs Sume accepts
A model can fetch a page, but Sume media inputs must be public HTTPS image or video URLs, not web pages. Here is what each Sume endpoint accepts.
Trigger.dev public tokens: keep the Sume key server-side
Trigger.dev v4.6.2 hardened authorization for public tokens. Whatever token your browser holds, a Sume API key must never be one of them. Here is the split.
Workers KV jurisdictions: keep Sume job records in region
Cloudflare made Workers KV jurisdictions generally available on Oct 2, 2026. Here is how to key Sume job_id records into a region-scoped namespace.
CrewAI conversational flows: confirm cost before a Sume render
In a CrewAI chat flow, have the step call Sume with dry_run first and ask the user to confirm the cost.
Prompt length limits across Firefly and Sume
Adobe Firefly now takes longer prompts for Image and Video on the web. Sume documents a 5000-character cap for music and no stated cap elsewhere.
Luma callback: 3 retries, 5 s timeout, vs Sume webhooks
Luma retries a failed callback at most 3 times with a 5-second timeout. Sume retries job webhooks up to 10 times with a 10-second timeout and signs each body.
Workflow DevKit replays: a stable Idempotency-Key for Sume submits
Workflow DevKit 4.8.12 fixed replay clock tracking. A replayed step must send the same Sume Idempotency-Key, so derive it from the run id, not Date.now.
Dropped connection mid-render: what happens to the Sume job
A dropped connection never cancels a Sume job. Wait again with jobs_wait on the same ids, or read the job status; never resubmit the paid create.
Parallel tool calls: fan out Sume renders, wait once
ElevenAgents can run multiple tools in one turn. On Sume's hosted MCP, submit several jobs, then wait on all of them with one jobs_wait call of up to 20 ids.
OpenTelemetry spans for MCP tools: tag the Sume request_id
Claude Code v2.1.283 adds MCP tool outputs to OpenTelemetry spans. Put the Sume request_id on the span and keep signed URLs off it.
Computer-use agent or API call: use Sume jobs, not clicks
OpenAI added computer use to its Agents API on Sep 29, 2026. For media generation, an agent should call Sume's API or MCP tools rather than click a dashboard.
Pre-flight tool access: tools_list on Sume's MCP
Notion added a tool that reveals connection-scoped capabilities before requests. Sume's MCP does the same job with tools_list, tools_schema and mcp_health.
Veo is US-region only: what EU teams call on Sume
A Sept 2026 listing says Veo runs only in us-central1, with no EU region documented. Sume has one video endpoint and its docs state no region.
claude plugin validate: a clean .mcp.json entry for Sume
Claude Code v2.1.281 makes claude plugin validate check .mcp.json entries. A Sume entry needs only the URL https://mcp.sume.com/mcp and no secret.
Basin Pipelines: log Sume webhook deliveries and dedupe by job_id
Cloudflare Basin Pipelines streams ingest up to 1 GB/s. Log each signed Sume webhook delivery as one record and dedupe on job_id; Python sketch included.
Sora files lived 1 hour: storing Sume job results
OpenAI's Sora 2 and Videos API shut down Sept 24, 2026, and videos were downloadable for 1 hour. Download and store Sume job results promptly.
HappyHorse video-edit ids and the Sume edit route
Alibaba Model Studio lists HappyHorse 1.1 t2v, i2v and r2v ids and points edits elsewhere. Sume has no HappyHorse ids; its edit route is Gemini Omni Flash 1.1.
Trigger.dev Node 21 warning: which Node runs the Sume SDK
Trigger.dev v4.6.1 added Node.js 21 deprecation warnings. The Sume TypeScript SDK needs Node 18 or later, so tasks on Node 22 or newer are fine.
Expiring API keys: Sume key metadata and rotation habits
OpenAI added enforced key lifetimes in Sep 2026. Sume's docs list key id, name, prefix, scopes and last-used time, so rotate on a schedule you keep.
Thin vs full webhook payloads: how Sume's job events work
Stripe made thin events generally available for API v1. Sume's job webhooks carry the result in the payload, are keyed by job_id, and can be redelivered.
Backgrounded MCP tool lost progress: resume with Sume jobs_wait
A Claude Code fix covers MCP progress dropped when a tool moves to the background. Do not rely on progress for Sume jobs: re-issue jobs_wait in slices.
Audit logs without file names: what to log for Sume jobs
Claude's Compliance API Activity Feed stopped returning file names. For Sume jobs, log request ids and job ids, never media URLs or transcripts.
Animate a still by API: first-frame jobs on Sume
fal lists FLUX 3 as animating one still into video. On Sume you send the still as a first frame to a video model; this Python script submits and polls.
Workers post-quantum Web Crypto: Sume webhooks stay HMAC-SHA-256
Cloudflare Workers Web Crypto now supports ML-KEM and ML-DSA. Sume webhook signatures are still HMAC SHA-256, and verifyWebhook runs on Workers with WebCrypto.
Vendors swap GPUs; keep one video job shape
Luma said on Jul 23, 2026 it runs video-to-video inference on AMD and Tensorwave. Your client should not care: one Sume job shape covers every model.
Gemini video understanding 88% fewer tokens vs Sume Video inspect
Gemini reports up to 88% fewer tokens on long video. Sume Video inspect and Reference ingest take another route: stills, transcript and a manifest.
Pydantic AI Workspace sandboxes: fetch Sume artifacts
Pydantic AI's Workspace abstraction runs tools locally or in sandboxes. Inside a sandbox, fetch Sume artifact URLs, and give Sume inputs as public HTTPS URLs.
Attach a terminal to a run: ant sessions connect vs Sume jobs
The ant CLI can attach to a Managed Agents session. For Sume generation jobs, use sume jobs watch and MCP jobs_wait instead, and never resubmit a paid job.
Slow OAuth consent and MCP timeouts: Sume's Write toggle step
MCP Python SDK v2.3.0 stops counting interactive OAuth logins against request timeouts. Here is the consent step with Sume's Write toggle that benefits.
GitHub Actions and Sume: secret setup and a smoke test
Store SUME_API_KEY as a GitHub Actions secret, never echo it, and run sume account get --json as a read-only smoke test in a manually triggered workflow.
Seedance 2.5 720p on fal, 1080p on Sume: read the catalog
fal lists Seedance 2.5 Image to Video at up to 720p and 30 seconds. Sume docs list seedance-2.5 at 4-30 s and 480p/720p/1080p. Check the catalog before pinning.
Wan 3.0 reads PDFs: what Sume's video request accepts
Wan 3.0 adds document and webpage inputs. Sume's video request takes prompt, image, video and audio fields only, so turn a PDF into a prompt first.
Blind-test Sonic 3.6 against 3.5 on your own script
A vendor's blind-test percentage is not yours. Render the same lines with two catalog versions through the TTS Router, shuffle them, and let listeners vote.
How do I know a new TTS model is available on Sume?
Every week brings a new voice model. The live answer is one GET request: list the TTS Router catalog. What an unknown id returns, and how to avoid hard-coding.
19-second AI video with references: Wan 3.0 limits on Sume
A 19 s reference-to-video job on Wan 3.0 allows 10 images, 5 videos and 5 audio clips, with video and audio each capped at 15 s total. Price for 19 s.
22-second AI video from a first and last frame on Sume
A 22 s clip that must land on a given last frame can use Wan 3.0 or Seedance 2.5 on Sume, the two end-frame models that run past 15 s.
23-second AI video with audio references on Wan 3.0 and Seedance
Audio references are accepted by Wan 3.0 (up to 5 clips, 15 s total) and Seedance 2.5. A 23 s clip price on Wan and the rules for audio inputs.
Sume bulk queue 404 format_run_queue_not_found: three causes
A Sume bulk queue poll returned 404 format_run_queue_not_found. The id is wrong or the queue is another owner's. How to tell which, and what to do next.
Claude batch custom_id is 64 characters: keep SKU keys valid
Anthropic batch custom_id allows 1 to 64 letters, digits, underscore and hyphen. How to sanitize SKUs, avoid collisions and carry the key into Sume input.
MCP error text showed a Bearer token: rotate the Sume API key
Claude Code 2.1.286 masks credentials after Bearer or Basic in MCP errors. If an old log shows a Sume key, replace it and keep keys out of logs.
Claude Code scope re-authenticate prompt vs unattended Sume runs
Claude Code 2.1.288 prompts to re-authenticate when a server asks for more OAuth scope. Unattended runs cannot answer it, so grant write up front.
Claude Code MCP startup wait: Sume tools missing on the first -p turn
Headless Claude Code bounds how long turn one waits for MCP servers with CLAUDE_CODE_MCP_STARTUP_WAIT_MS. Name Sume tools and check mcp_health first.