Scheduled run receipt shows less than the wallet debit: which number
usage.billable_amount_usd_micros on a run receipt is generation spend only. GET /v1/usage?run_id= debited_usd_micros also counts the Agent's own turns.
Agents
The Sume video agent from code: Agent Completions, scheduled runs, MCP tools and timeouts, programmatic tool calling, and guardrails for paid APIs.
Start with: What is a video agent? How Sume defines and runs one
Sizing generation_spend_cap_usd for an editing Agent Completion
generation_spend_cap_usd is required on every Agent Completion. Price the media jobs, then add headroom: edits of $0.22 to $1.24 suggest caps of $1 to $3.
What to log from a Sume agent run: request ids, never signed URLs
Safe logs for agents calling Sume: request ids, status, sanitized media metadata. Unsafe: API keys, signed URLs, private media URLs, transcripts.
avatars_search: hybrid by default, hybrid=false for exact handles
Sume avatars_search defaults to hybrid ranking for semantic queries. Send hybrid=false to match an exact handle or keyword, and filter for ready avatars.
account_me then balance_get: two reads before an agent spends
Before an unattended agent submits paid work over Sume MCP, call account_me to confirm the workspace and balance_get for the wallet. What each returns.
Claude Code routine connectors: leave Sume's Write off first
A new Claude Code routine includes every connector and can call write tools without asking. Grant Sume read-only first, then widen it on purpose.
Can an AI agent swap the person in my video? Ask Remixy
In a Sume chat, Remixy remakes a clip from a reference and can use H3 Max Recast for a person swap. What to give it, and what it should say before paying.
Dry-run an H3 Max Recast from MCP: dry_run and max_spend_usd
Before a paid person swap, Sume's hosted MCP can preview cost with dry_run and cap spend with max_spend_usd. The calls, their order, and idempotency_key.
Reference remix take kinds: which model renders each shot
A remake in Sume is planned as takes, and each take kind names its model: talking, performed, voice, still, recast, genjutsu, product swap. The map, and why.
Remake a viral clip with my product: edit it or rebuild it?
Sume has two remake routes: models that keep the source footage, and reference remix that rebuilds shot by shot. How the agent decides and when to override.
Claude Code routine API text is untrusted: pass Sume input as data
A routine's fire text arrives wrapped as untrusted data, and the saved prompt must opt in to use it. Sume's run input follows the same rule: data only.
Claude Code routine run is green: did the Sume video get made?
A green routine status means the session exited cleanly, not that the task worked. Read Sume's run receipt: status, outcome, output_error and billed amount.
Copilot Studio computer-use allow list: what it blocks, and spend
The access-control list stops actions on other sites but not opening them. It is not a spend control either, so put a cap on the Sume call itself.
How an agent picks a video model over hosted MCP
An agent reads video-router_models, picks an id, then calls generate_video with an idempotency_key and waits with jobs_wait. Omit the model for sume/auto.
Get a typed podcast clip plan from an Agent Completion
Send a transcript and an output_schema to POST /v1/agent/completions, and get back start and end times for each clip, ready for video-trim.
Why a Scheduled Action run says skipped: previous_run_active
A Sume Scheduled Action recorded status skipped with skip_reason previous_run_active. It means the last run was still going. Here is how to react.
A weekly scheduled agent run for podcast clips: cron, cap and trigger
Set a Sume Scheduled Action to run every week, with a cron schedule, an IANA time zone and a spend cap that a manual or API run can lower but never raise.
Sume Action run 409: action_api_trigger_disabled vs action_inactive
Two different 409s stop POST /v1/actions/{id}/runs before any work starts. What each means, the dashboard fix, and how they differ from action_run_in_progress.
Why a Sume run's billable_amount_usd_micros is not its total cost
usage.billable_amount_usd_micros on a Sume Scheduled run receipt is generation spend only. It leaves out the agent's own LLM turn. Where to read the full bill.
Claude Code routines API trigger vs the Sume Scheduled run API
Claude Code routines' /fire endpoint and Sume's POST /v1/actions/{id}/runs both start a saved agent over HTTP. Compare payload, limits, receipts and webhooks.
Which model runs a Sume scheduled agent, and how to choose it
A Sume schedule stores its own model beside its instructions and cap. Where to pick it, what Agent Completions accepts, and what the changelog says on defaults.
hypit Understand order: one probe, then parallel batches
Sume's hypit Understand order: probe alone, then transcribe, boundaries and tiles in one batch, then notes. Which verbs wait on the transcript.
Poll a Sume Action run in Python: branch on next_action, not status
A small Python loop for a Sume Scheduled run that sleeps on poll_status, retries on retry_later and stops on none, then fetches the result.
Cron or API call: the trigger type of a Sume schedule is fixed
A Sume schedule's trigger_type is set at creation. An API-only schedule can never gain a cadence, and a cron one can never become API-only. What to pick first.
Passing data to a scheduled agent run: input is data, not instructions
What the input field on a Sume Scheduled run does, its 64-property and 2 MiB limits, and why caller text cannot rewrite the saved instructions.
Storyboard to finished video over Sume MCP: the tool order
The order a Sume MCP client should follow: stills, inspect, wordless clips, probe, then a timeline dry run and render. Where each step bills and what to skip.
Sume Action run 403 insufficient_scope: old keys and service accounts
A 403 on POST /v1/actions/{id}/runs means the key lacks actions:write, predates the scope, or is a service-account key. Why scopes cannot be added and the fix.
sume skills export: review the Sume agent skill before you install it
Run sume skills export to read the packaged Sume skill's source before sume skills install writes it into .agents/skills or .claude/skills. Commands and gates.
Verify a Sume run webhook in Python: replay window and empty secret
A Python verifier for the sume-v1 signature on a Sume run webhook: raw body, five-minute replay window, constant-time compare, and no empty secrets.
Agent Completions 403: service_account_agent_completions_unsupported
A Sume service-account key cannot create Agent Completions. It fails with 403 insufficient_scope and a reason code; use another key with the right scopes.
CrewAI 1.15: per-run state isolation for parallel Sume jobs
CrewAI 1.15 scopes runtime state per run so concurrent crews do not mix. Use that with one Sume Idempotency-Key and job id per crew run.
Higgsfield MCP has no credit cap: cap spend with Sume max_spend_usd
Higgsfield's MCP guide says there is no built-in credit spending cap. Sume's MCP tools take dry_run and max_spend_usd. How they work, and what they skip.
AI Act Article 50(1) chatbot notice for an agent built on Sume
Article 50(1) asks providers to make sure people know they are talking to AI. If you front Sume Agent Completions with a chat box, the notice is yours to add.
GitHub Actions schedule delayed at the top of the hour: what to do
GitHub says scheduled workflows can be delayed, notably at the start of every hour. Offset the cron minute and call a Sume Scheduled run with a dated key.
GitHub Actions schedule disabled after 60 days: a Sume Scheduled fix
A public repo's scheduled workflows are disabled after 60 days without activity. How Sume Scheduled's active and inactive status differs.
Scheduled agent run skipped: no webhook arrives, read the 200 receipt
A skipped Sume Scheduled run returns 200 with skip_reason previous_run_active and sends no webhook. Read the receipt instead of waiting on one.
Vercel Hobby cron runs once a day, up to 59 minutes late
Vercel Hobby cron runs once a day with hourly precision, so a job can start up to 59 minutes late. When to move paid work to a Sume bulk run or Scheduled agent.
Flow Agent batch edits and asset tidying: keep a Sume manifest
Flow Agent can batch edit and organize assets inside Flow. Sume has no asset organizer; a manifest of job ids and media URLs does the job. Code included.
audio_detach MCP tool: idempotency_key, jobs_wait, jobs_result
Hosted MCP exposes audio_detach. It needs an idempotency_key, mcp:write under OAuth, then jobs_wait and jobs_result to read the new audio_url.
Can ChatGPT make videos from photos? Image to video now
Not with Sora, which OpenAI discontinued. ChatGPT can send a photo's public URL to an image-to-video tool over MCP and get a clip back.
Can ChatGPT make videos with sound? Audio and voice
Yes, through a video tool whose model generates audio. A voice that speaks your script is a second step: text to speech, then lip sync.
Can ChatGPT make long videos? Clip limits and how to join
Not in one clip: generated clips top out at 15 or 30 seconds. ChatGPT can join many clips over one audio track into a render of up to 30 minutes.
Can ChatGPT make videos better quality? Upscaling a clip
With an upscaling tool, yes: ChatGPT can send a clip's public URL to a video upscaler over MCP. For new clips, ask for a higher resolution.
Can ChatGPT make slideshow videos? Stills, voice, one MP4
Yes, through tools: ChatGPT can make the slides, voice the script, and have a timeline tool hold each still for its line in one MP4.
Can ChatGPT make TikTok videos? Vertical clips and limits
ChatGPT can write the hook and, with a video tool over MCP, have a 9:16 clip rendered. Posting it to TikTok is a separate step you do.
Can ChatGPT make music videos? Song, clips, one render
With tools, yes: ChatGPT can have a track generated, a short clip made per section, and the clips cut over the song in one MP4 render.
Can ChatGPT make music? Getting a song or beat as audio
To get a song or a beat as an audio file, give ChatGPT a music tool: in developer mode it can call one over MCP and hand back a link to the track.
Can ChatGPT make product videos from a product photo?
Yes, with a video tool: in developer mode ChatGPT can send your product photo to a video model as the first frame and return a link to the clip.