MCP server for orchestrating parallel Claude Code worker swarms to manage long-running coding tasks
Claude Swarm exposes only one tool (orchestrator_init) with a partially-defined schema and minimal description. The tool name lacks action clarity (should be 'init_orchestrator' or 'start_orchestration'), the description is vague about what 'orchestration session' means, and critical parameter relationships and constraints are undocumented. The input schema exists but lacks descriptions for all parameters, a violation of the rubric baseline that 100% of A+ tools have parameter descriptions. Output schema is not documented at all. The tool's WRITE risk is noted but no error recovery guidance is provided. This server appears to be a stub or prototype; the actual orchestration logic is delegated to external Claude Code workers, but the tool definition does not expose that complexity to the LLM.
Initialize a new long-running orchestration session. Call this first with a task description to set up the feature list, progress tracking, and environment.
Tool name 'orchestrator_init' lacks clear action verb. Should start with 'init_', 'start_', or 'create_' to signal intent. Current name is ambiguous, does it initialize a session, start workers, or configure the environment?
Tool description (69 chars) is too vague. Phrases like 'long-running orchestration session' and 'set up the feature list' are unclear. LLM cannot determine when to call this vs other hypothetical tools. Does NOT state what an 'orchestration session' is, what happens if called twice, or what the return value contains.
All input parameters lack descriptions. 'projectDir', 'taskDescription', and 'existingFeatures' are defined in schema with types and format hints, but no human-readable descriptions explaining what values are expected, constraints, or format. Rubric baseline: 100% of A+ tools have param descriptions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 38 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 46 | - | v1 |
No output schema or return value documented. LLM has no way to know what fields the tool returns, whether it includes session_id, worker_list, or status. This breaks downstream tool composition and forces the LLM to reason about unknown structures.
No error handling or recovery guidance. Tool marked as WRITE risk but does not document failure modes (e.g., 'invalid projectDir', 'taskDescription too long') or what the LLM should do if initialization fails.
Parameter 'existingFeatures' is optional but dependency is undocumented. When should the LLM pass this vs omit it? Does omitting it trigger auto-analysis? Rubric baseline: undocumented parameter dependencies cause silent misuse.
Parameter 'projectDir' accepts a string but no format, validation rule, or example is documented. Should it be absolute? Relative? Does the tool create the dir if missing? Rubric baseline: 'specify expected format, range, and allowed values directly in parameter description.'