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.
Codex disabled_tools: block Sume's paid MCP tools
List Sume's paid tool ids under disabled_tools in the Codex config.toml so the agent cannot call them, even on an API-key session that sees every tool.
Gemini CLI trust: true on a Sume MCP server: what it skips
Gemini CLI's trust setting bypasses every tool confirmation dialog. What that means for Sume's paid tools, and which Sume gates still apply.
Codex mcp_oauth_callback_port and the Sume OAuth login
Pin the Codex OAuth callback port for a remote Sume MCP login, and how the login flows from Codex to the Sume consent page on mcp.sume.com.
OpenAI Agents SDK MCPServerManager with a Sume server
Run Sume next to another MCP server in the OpenAI Agents Python SDK: connect Sume over streamable HTTP, check it with mcp_health, and read tools_list.
Gemini CLI MCP timeout default vs Sume's 55-second jobs_wait
Gemini CLI's MCP timeout defaults to 600,000 ms. Sume's jobs_wait holds at most 55 seconds per call: repeat it on wait_slice_expired, never resubmit.
AI SDK isLoopFinished vs the 20-step cap for Sume video jobs
ToolLoopAgent stops after 20 steps by default. How many jobs_wait calls a Sume video render needs, and the per-call guards to keep if you lift the cap.
AI SDK MCP client close() in onEnd: Sume jobs keep running
Closing the AI SDK MCP client in onEnd ends the tool connection, not the Sume jobs already submitted. Track job ids, then read or cancel them.
Claude Code list_changed and reconnect: refresh Sume tools
Claude Code refreshes an MCP server's tools on list_changed and lets you reconnect from /mcp. After changing Sume's Write consent, verify with tools_list.
Codex MCP enabled = false: pause the Sume server safely
Turn the Sume server off in Codex with enabled = false instead of deleting it, and rotate the API key if it ever showed up in logs or chat.
Can I continue a Sume Agent Completion with thread_id?
Not yet. Every Agent Completion runs in a fresh thread, so a follow-up call cannot reuse thread_id. Pass earlier results back in the next request instead.
OpenAI response_format json_schema on a Sume scheduled run
Sume accepts an OpenAI-shaped response_format as an alias for output_schema on schedule runs. Sending both returns 400, and the schema must be strict.
catalog_list shows a capability that tools_list does not
On Sume's hosted MCP, catalog_list lists public API capabilities and can name ones with no MCP tool. tools_list is the session's real tool contract.
MCP tool name with a dot or underscore: tools.list vs tools_list
Sume MCP tool ids use underscores; a dotted alias such as tools.list is canonicalized on call. Retired aliases map to generate_image and generate_video.
Sume music job says sume/music-auto: which engine ran?
job.model echoes the id you requested; job.request.routed_model names the engine that ran, such as lyria-3.5. Read both fields on the job envelope.
Lambda 90-minute timeout: does it change AI video jobs?
AWS raised Lambda's async timeout to 90 minutes on Managed Instances. A Sume job still fits best as submit, then poll or webhook, and sync waits stay 30 s.
Workflows keeps state 7 days now: store your Sume job ids
Cloudflare Workflows on Workers Paid now keeps finished instance state 7 days by default. Save the Sume job id and result URL in your own store.
Workflows .subscribe() vs Sume job events: push or pull?
Cloudflare Workflows can stream instance events with subscribe(). Sume has no SSE stream for jobs: use a signed webhook, or poll status and the events snapshot.
Start a Cloudflare Workflow from a Sume webhook with ctx.exports
Verify the Sume webhook signature in a Worker, then start a Workflow with ctx.exports and no workflows binding. Includes a short Worker handler.
Workflows instance limits vs Sume queue headroom: which binds?
Cloudflare lists 50,000 concurrent Workflow instances (Paid, April entry). Sume reports your own generation headroom; size fan-out by that formula.
GitHub Actions re-run: which idempotency key for Sume jobs?
GITHUB_RUN_ID stays the same on a re-run while GITHUB_RUN_ATTEMPT increments. Build the Sume Idempotency-Key from the run id so a re-run does not bill twice.
cancel-in-progress killed my workflow; does the Sume job stop?
Cancelling a GitHub Actions run does not cancel a Sume job. Cancellation only works before generation starts, so store the job id and cancel it explicitly.
Zapier accepted my Sume webhook but the Zap ran minutes later
Zapier says it may return 200 and delay webhook processing by minutes. Sume sees a delivered 200, so keep a status poll as the backup for run and job results.
Zapier Catch Hook test trigger with Sume Send test
Zapier lists the three most recent webhooks from the past hour. Use Sume's Send test to put a signed webhook.test sample there before a real run exists.
Airtable has no webhook signature check: verify Sume first
Airtable's When webhook received trigger cannot verify signatures and caps payloads at 100kb. Verify the Sume signature in a relay and forward a small object.
Pipedream 512KB body limit and Sume run webhook payloads
Pipedream limits HTTP trigger bodies to 512KB by default. Sume run webhooks can carry up to 1 MiB, so plan for a slim relay or the result_url fetch.
Pipedream 30-second timeout: call Sume in async mode, not sync
Pipedream HTTP workflows time out at 30 seconds by default. Sume sync waits up to 30 seconds too, so submit async and take the result by webhook or poll.
Retool Workflow webhook needs X-Workflow-Api-Key: use a relay
Retool Workflows authenticate webhooks with an X-Workflow-Api-Key header or query parameter. Sume documents no custom delivery headers: relay after verifying.
Google Cloud Workflows callback needs an IAM token: relay Sume
A Cloud Workflows callback URL needs the workflows.callbacks.send permission and a Bearer token. Sume webhooks cannot carry one, so relay after verifying.
n8n Wait node under 65 seconds: how to poll a Sume job
n8n keeps waits under 65 seconds in memory and saves longer ones to the database. Poll a Sume job with next_poll_after_seconds, or switch to a webhook resume.
Vercel rewrite to an external API times out at 120 seconds
Vercel proxied rewrites to an external destination time out at 120 seconds with ROUTER_EXTERNAL_TARGET_ERROR. A Sume async submit returns at once.
D1 free-tier limit now fails queries: dedupe Sume webhooks safely
Cloudflare D1 queries on Workers Free now fail past the daily row limits. Dedupe Sume webhooks on request_id, return a 5xx, and use redeliver after the reset.
Webhook rate limits: Zapier, Airtable, Pipedream vs a bulk of 100
Zapier, Airtable and Pipedream each publish a webhook intake limit. Compare them with a Sume bulk queue of up to 100 items and pick a receiver that fits.
Remove filler words from a talking video with an API
Sume has no one-call filler remover. Transcribe with video inspect for word timings, then cut the clean ranges with video trim at $0.02 per job.
Face swap API: quality has no default, unlike avatar videos
Avatar videos default to plus when quality is omitted. Sume's Beta face swap has no default: quality is required and takes standard, plus or max.
Video analysis API with a YouTube link: 422 unsupported_source
Sume's video analyses reject YouTube URLs with 422 unsupported_source. Import the video you own to media.sume.com first; rules for the other bad inputs.
What a Sume video analysis returns for each scene
Each scene in a stored video analysis has start and end seconds, summary, visual, shot type, on-screen text, confidence and keyframes. Field-by-field list.
Video analysis limits: max_scenes 2-40 and the 90 s and 300 s caps
Sume's video analysis takes max_scenes from 2 to 40 (default 24), warns past about 90 seconds and rejects past 300. Cost, runtime and polling advice.
Submitting 20 avatar videos at once: what each Sume plan accepts
Sume accepts as many paid jobs as concurrency plus queue allows: 6 on Free, 24 on Pro, 48 on Startup, 120 on Scale. Past that a submit returns 429 queue_full.
Paper edit by API: build a rough cut from transcript lines
Premiere 26.5 added Paper Edit. To do the same by API, transcribe with sentence segments, pick the lines you keep, then render those ranges as a Timeline.
Search footage for a spoken phrase with an API: words with timing
DaVinci Resolve 21 lists IntelliSearch. To find a spoken phrase in a clip by API, transcribe with video inspect and search words[] in your own code.
Remove filler words from a video by API: cut at word timestamps
Descript's API lists Remove Filler Words as an Underlord edit. the Sume docs list no such op; here is how to cut um and uh yourself from words[] and a Timeline.
Reframe a video around the speaker by API: what a crop can do
Descript's Center Active Speaker (Beta) follows whoever talks. Sume's video filter crop is one fixed rectangle per clip, so cut per speaker and crop each part.
Get the rendered MP4 back from a video editing API job
Descript's API lists no rendered file download without publishing. In Sume, a finished timeline_render job result carries video_url; here is the poll flow.
MCP tools for editing video: the hosted Sume list, trim to timeline
Descript and Runway both ship MCP servers. The hosted Sume MCP exposes video_trim, video_filter, audio_detach, video_frames_create and the timeline tools.
Join voiceover clips inside one timeline render with audio.parts
Timeline 1.0 takes up to 20 gapless audio.parts slices with source_in and duration. Skip the timeline-audio job when the join is only for one render.
Timeline render warnings: padded or looped short sources explained
A Timeline 1.0 slot longer than its source renders with a soft warning, not a failure, and /plan cannot predict it. Read warnings[] and probe clips first.
Timeline 1.0 limits: 200 slots, 1800 s, 20 audio parts, 8 fades
The numbers that cap one Timeline 1.0 render: 1 to 200 video slots, 1 to 1800 s of audio, 20 audio parts, 12-slot single strategy and 8 chained fades.
Extract audio from a long video: the 900 s output cap and range
Audio detach takes sources up to 1800 s but outputs at most 900 s. For a longer track pass range and detach in two calls. Errors included.