MCP tool search: how a long Sume tool list loads in Claude Code

Claude Code's MCP page says tool search is on by default: rather than putting every MCP tool description into the model's context, it loads them when the model needs them. That suits Sume's hosted server, whose tool list is long, because a video task only needs a handful of tools.

The page was read 2026-09-29; the Sume tool names come from its MCP docs.

What does Claude Code's MCP page say?

A few details matter for a large server.

Which Sume tools does a video task need?

Most of the tool list is for other jobs: the crawl_* family, assets_*, timeline_* and more. A video task usually touches five.

• tools_list and tools_schema to discover and inspect a contract
• generate_video to submit, with payload.model omitted to route to sume/auto
• jobs_wait to wait on the returned job ids
• jobs_result to read the finished media URLs
• balance_get or generation_admission_preview to check funds first

How do I help the model find them?

Name the tools in the prompt. With tool search the model discovers tools by need, and a prompt that says which Sume tool to use for which step removes guesswork. Have it call tools_schema with a tool name before its first paid call so it reads the exact contract, including idempotency_key and dry_run.

You can also pre-approve a short list in Claude Code's permissions instead of leaving the whole server open; the site's page on allowing MCP tools shows the rule syntax.

What if a result is too large?

Sume's job results are short JSON with media URLs, so they sit well under Claude Code's default 25,000 token output limit. A crawl or catalog listing can be larger; ask for a narrower call rather than raising the limit. To do many calls in one step, script_run runs a short JavaScript program on Sume's side with sume.call and returns a calls[] journal and jobs[].