defer_loading and tool search for Sume's MCP tools
Set defer_loading: true on the Sume MCP tool so OpenAI tool search loads definitions on demand, then use tools_list and tools_schema for details.
Add defer_loading: true to the MCP tool entry that points at https://mcp.sume.com/mcp. OpenAI's MCP guide says this option works together with tool search, so the model does not carry every Sume tool definition on every turn.
Sume's tool set is large: assets, generation, crawl, avatar and timeline tools. Loading all of it up front costs tokens, and OpenAI bills tokens for imported tool definitions. Deferral pushes that cost to the turns that need it.
Deferral and Sume's own discovery tools
Sume ships two layers for the same problem. The first is client-side: your agent platform decides when to load definitions. The second is server-side: tools_list returns names and short descriptions, and tools_schema returns the full schema for one tool (hosted MCP docs).
| Layer | Mechanism | Documented by |
|---|---|---|
| Client | defer_loading: true plus tool search | OpenAI MCP guide |
| Server | tools_list, then tools_schema for one tool | Sume hosted MCP docs |
| Server gate | mcp:read hides write and paid tools | Sume tools and gates |
| Both | Pin allowed_tools to a short list | OpenAI MCP guide |
Request shape
Combine deferral with a short allowed_tools list for the tools a job actually uses. The request below only differs from a normal MCP tool entry in the extra flag.
{
"type": "mcp",
"server_label": "sume",
"server_url": "https://mcp.sume.com/mcp",
"authorization": "<SUME_OAUTH_TOKEN>",
"defer_loading": true,
"allowed_tools": ["tools_list", "tools_schema", "account_me", "catalog_list"]
}Caveats
- Deferral changes when definitions load, not what the server allows. Scope gates still apply.
- Read the OpenAI page for the exact tool-search setup; this post does not restate it.
- If a tool is missing from search results, call
tools_listyourself to see whether the token's scope hides it.
Sources
Related posts
More in Developers
- Wire an OpenAI mcp_approval_response to a Sume spend cap
When a person approves an mcp_approval_request, reuse the approved dollar amount as max_spend_usd on the Sume call so the approval and the cap match.
- require_approval never with tool_names for Sume read tools
Use the object form of require_approval to skip approval only for Sume read tools like account_me, and keep every paid tool on the approval path.
- Restyle burned-in captions without transcribing twice
Pass source_caption_id instead of video_url to re-burn a video under a new Sume caption style. Word timings are reused, so no second speech-to-text runs.
- Resume an Omni batch after a crash with stable idempotency keys
A worker dies halfway through 40 Omni clips. Build Idempotency-Key from the item id so a rerun resubmits safely, and treat 409 and 429 as signals, not failures.
Written by Sume