Claude Code subagent MCP server: scope Sume tools to one
Put a Sume MCP server in one subagent's mcpServers field so video and image tools stay out of the main chat. Frontmatter for both routes, plus limits.

Yes: a Claude Code subagent can have its own MCP server. List it under mcpServers in the subagent's frontmatter, either as the name of a server you already configured or as an inline definition that only that subagent sees. An inline server connects when the subagent starts and disconnects when it finishes, and the main conversation never loads its tool descriptions.
That is from Anthropic's subagents page, read 2026-09-29. Sume's hosted MCP endpoint is https://mcp.sume.com/mcp, per the MCP quickstart. Below are both routes for a subagent that makes video and image clips.
How do I define an inline Sume server for one subagent?
Save this as .claude/agents/clip-maker.md. The sume key is the server name; the entry uses the same schema as a .mcp.json server, and http is one of the supported types. The header authenticates with an API key from an environment variable, which keeps the key out of the file.
---
name: clip-maker
description: Makes short video and image clips with Sume. Use for any clip request.
mcpServers:
- sume:
type: http
url: https://mcp.sume.com/mcp
headers:
Authorization: Bearer ${SUME_API_KEY}
---
Use the Sume tools. Preview a paid call with dry_run first,
send max_spend_usd, and wait with jobs_wait on the job id.How do I reuse a server I already added?
If you ran claude mcp add --transport http sume https://mcp.sume.com/mcp and signed in, reference it by name. A string reference shares the parent session's connection. Add a tools allowlist to narrow the subagent to a few tools. Anthropic's page says tools and disallowedTools accept mcp__<server> and mcp__<server>__* patterns.
---
name: clip-maker
description: Makes short video clips with Sume.
tools: Read, mcp__sume__generate_video, mcp__sume__jobs_wait, mcp__sume__jobs_result
mcpServers:
- sume
---What can and can't a subagent do with MCP?
Anthropic's page answers these questions about subagents and MCP tools.
| Question | Answer |
|---|---|
| Does a subagent see the parent's MCP tools? | It inherits the MCP tools available in the main conversation, narrowed by filters |
| How do I keep it away from MCP tools? | List only built-in tools in tools; the page's Read, Grep, Glob, Bash example says the subagent cannot use any MCP tools |
Do plugin subagents support mcpServers? | No; the field is ignored for plugin subagents |
| Is an inline server always loaded? | Not from a project's .claude/agents/ until you trust that folder |
| Do background subagents keep MCP tools? | Yes; a background subagent keeps every MCP tool |
Can the subagent wait for a video job?
Yes, but one jobs_wait call holds at most 55 seconds (default 50). Sume's jobs docs say to repeat the wait with the same ids when it returns wait_slice_expired, never to resubmit the paid create. Put that rule in the subagent's prompt. jobs_wait also takes up to 20 ids in job_ids, so one wait can cover a batch.
Paid tools such as generate_video need an idempotency_key, and dry_run and max_spend_usd are optional gates, per Sume's MCP tools and gates. A subagent does not add its own spend limit, so tell it to send the cap.
Why did my subagent file not load the server?
Anthropic's page says field names are camelCase and must match the table exactly, and that Claude Code ignores a field it does not recognize without reporting an error. A misspelled mcpServers gives you a subagent with no Sume tools and no warning.
An untrusted project folder also skips inline servers until you accept the trust dialog, and managed MCP restrictions such as allowedMcpServers and deniedMcpServers cover subagent servers too. For sharing one server across a team instead, see Claude Code MCP project scope.
Sources
Related posts
More in Developers
- MCP URL mode elicitation in Claude Code vs OAuth consent
Claude Code 2.1.281 added MCP URL-mode elicitation, a browser flow a server can request. Sume's browser step is OAuth consent, a different mechanism.
- Claude skill allowed-tools: pre-approve Sume MCP tools
A skill's allowed-tools field pre-approves tools for one turn and does not restrict the rest. How to list Sume's job tools and leave paid ones prompting.
- Add a video tool to Claude Sonnet 5.5 in the Messages API
Define a generate_video tool with input_schema, run it against Sume's /v1/videos when Claude Sonnet 5.5 returns tool_use, and return the job id as tool_result.
- Claude tool_result is_error: return Sume API errors
When a Sume REST call fails inside your Claude tool loop, send the error back as a tool_result with is_error true and say what to try next. Code and table.
Written by Sume