Claude Code Subagent Check
Paste a .claude/agents/*.md subagent definition and get every structural mistake flagged in seconds — a name that isn't kebab-case, a description too vague for the orchestrator to delegate to it reliably, a tools entry that's empty, duplicated, or unrecognized, a model alias this tool doesn't know, and a missing or near-empty system prompt.
Structural checks, not a delegation test. This tool never runs your subagent and never calls an LLM — it only looks at the text you paste and flags known failure shapes in it. It does not know whether Claude Code's orchestrator will actually delegate to this subagent on a given request; it only tells you the file is well-formed enough to have a fair shot.
1. Paste your subagent definition
The full contents of the .md file, frontmatter and system prompt together. Nothing here is uploaded — everything runs in this page.
Findings
Paste a subagent definition above, or load the example, then press Check.
Usage is optional, stays in this browser, and counts only that a check ran or a summary was copied. Nothing you paste — no subagent name, description, tools, or prompt text — is ever recorded, uploaded, or sent to an analytics service. Uncheck to turn it off. Do Not Track and Global Privacy Control disable counting automatically.
Common questions
What does this tool check in a subagent definition file?
The YAML frontmatter (name, description, tools, model) and the Markdown system-prompt body that follows it. Name: present, kebab-case, and matching the filename you give. Description: present and not too vague to trigger reliable delegation (length, a concrete trigger cue, distinctive vocabulary). Tools: no empty or duplicated entries, and each name checked against a documented list of built-ins as a caution. Model: checked against documented aliases and the shape of a full model id, also as a caution. Body: not missing or near-empty.
Does this tool contact a network or Anthropic's servers?
No. Every check is a deterministic, structural check against the text you paste, entirely in your browser. It never uploads anything and never calls an LLM.
Does a clean result mean the orchestrator will delegate to my subagent correctly?
No. This tool checks the shape of the file, not Claude Code's actual delegation decisions at runtime. A description can pass every heuristic here and still not get picked for a given request, and a description that trips a caution can still work fine in practice -- every finding is worth a look, not a verdict.
Why is an unrecognized tool or model name only a caution, never an error?
Because MCP tool names and custom tool names are open-ended, and Anthropic can add a new built-in tool or model alias at any time. This tool checks against a documented, dated reference list -- a name outside it is worth a second look for a typo, not proof of a mistake.