Evaluating a video MCP server: five checks, with Runway as context
Runway shipped MCP and plugin channels on Sep 17, 22, 23 and Oct 1. Before you connect any video MCP server, check scopes, idempotency, spend caps and waits.

Before you connect a video MCP server to an agent, check five things: whether paid tools are separated from read tools, whether paid calls need an idempotency key, whether you can preview or cap spend, how long waits are bounded, and whether the agent can list tool schemas. Runway's changelog shows four new connection channels in two weeks, so these checks matter more each month.
| Date | Entry |
|---|---|
| Sep 17, 2026 | Grok Bot plugin |
| Sep 22, 2026 | Read-only Brand Kits through MCP |
| Sep 23, 2026 | Cursor Marketplace and DaVinci Resolve plugin |
| Oct 1, 2026 | OpenAI Dots integration, available on all plans |
1. Read and write are separate
A server that lets any connected agent spend money on first contact is a risk. Sume's hosted MCP defaults to read-only visibility under OAuth mcp:read; mutating and paid tools stay hidden until the session has mcp:write or uses an API key. Runway's Brand Kits entry is described as read-only, which is the same instinct applied to a single feature.
2. Paid calls carry an idempotency key
Agents retry. A paid tool without a dedup key turns a dropped connection into a second charge. On Sume, idempotency_key is required on every write and paid tool, so a retry can reuse the same key instead of paying twice.
3. Preview and cap the spend
Look for a dry run and a ceiling. Sume offers dry_run=true for an admission and cost preview without submitting, and max_spend_usd, which is enforced only when you provide it. The second point matters: a cap you never set protects nothing, so set it in the agent's instructions.
4. Waiting is bounded
A held HTTP request dies at the edge eventually. Sume's jobs_wait holds at most 55 seconds per call (50 by default), accepts 1 to 20 job ids, and returns wait_slice_expired when the slice ends so the agent repeats the wait instead of resubmitting the paid create. A server that promises a ten-minute wait in one call is promising something transport cannot deliver.
5. The agent can discover the contract
tools_list and tools_schema let an agent read the live contract before it acts, and the docs advise against assuming parity with the HTTP API. Run both against any candidate server and read the safety metadata for yourself. The full inventory is in MCP tools and gates.
Sources
Related posts
More in Integrations
- fal lists Seedream 5.0 and Krea 2: find what Sume's image catalog has
fal.ai/models names GPT Image 2.5, Flux 2, Nano Banana 2, Ideogram 4, Krea 2 and Seedream 5.0. A Python check of which names appear in Sume's /v1/images/models.
- LangGraph interrupt or wait: resume a graph from a Sume webhook
A LangGraph interrupt is for a human decision, not a video render. Pause for approval, then resume the graph from a signed Sume webhook keyed by job_id.
- Mastra eager tool execution: dry-run Sume calls first
Mastra 1.71 can start a tool once its own arguments are complete. For paid Sume generation that means a stable idempotency_key and a dry run before any spend.
- Mastra 1.72 crash recovery and leases: checkpoint the Sume job id
Mastra 1.72 adds multi-worker task leases and crash recovery. Store the Sume job id before waiting so a recovered worker polls instead of paying twice.
Written by Sume