MCP server for 17 AI agents working as a team with quality guardians and SNCP project memory access
CervellaSwarm MCP Server exposes 3 tools for reading and searching SNCP project state. All three tools have descriptions and are read-only (safe for safe-to-retry semantics). However, the server has significant gaps: (1) Input schemas are inferred from the provided JSON but NOT explicitly visible in the TypeScript source code shown, only tool names and descriptions are evident from 'packages/mcp-server/src/sncp/tools.ts' reference. (2) Parameter descriptions are minimal. 'sncp_read_ripresa' accepts a 'project' param described only as 'SNCP project name (e.g. cervellaswarm)', under 60 chars, weak on format/constraint detail. (3) Output schemas are completely undocumented, no documentation of what fields are returned by any tool, forcing LLMs to guess structure. (4) The tool names use domain-specific prefixes ('sncp_') which are clear but lack action verbs at the start ('read_' and 'search_' are present in two tools, but naming is inconsistent, 'list' vs 'search' semantics differ but both are somewhat generic). (5) No error handling guidance, no rate limiting details, no mention of permissions. Tool definitions are minimalist, suitable for a niche domain tool but falling short of production quality. The fact that actual TypeScript implementation is not fully visible (only file paths given) means schema validation is impossible.
List all CervellaSwarm SNCP projects with their available files and status.
Read the PROMPT_RIPRESA (session context) for a CervellaSwarm project. Contains: last session summary, decisions made, next steps.
Search across all CervellaSwarm SNCP project files for a query string. Searches PROMPT_RIPRESA, roadmaps, and other .md files.
Output schemas completely undocumented. No specification of what fields are returned by sncp_read_ripresa, sncp_list_projects, or sncp_search. LLMs cannot infer what data to extract or how to chain tool calls.
Input schemas not verifiable from provided source. TypeScript implementation file (packages/mcp-server/src/sncp/tools.ts) is referenced but not shown in detail. Cannot confirm parameter types, descriptions, or constraints are actually present in code.
Parameter descriptions are minimal and lack format/constraint details. 'project' param in sncp_read_ripresa is described as 'SNCP project name (e.g. cervellaswarm)', example value embedded (anti-pattern), no minLength justified, no examples of valid projects beyond the instance name.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | 2026-07-28+ | v2 |
sncp_list_projects has empty input schema ({}) but no description of what it returns or when to use it. 'List all CervellaSwarm SNCP projects with their available files and status' tells what it does but not return format, pagination, or result limits.
No error handling guidance. What happens if a project does not exist? What if search returns no results? Are there rate limits? Can tools be retried safely? Descriptions do not guide LLM recovery.
No pagination or result limits documented for sncp_list_projects or sncp_search. If searches return hundreds of matches or project lists are large, context window exhaustion risk. Baseline pattern expects limit + page/offset parameters.
Tool naming is domain-specific but not action-verb-first. 'sncp_read_ripresa', 'sncp_list_projects', 'sncp_search' start with domain prefix instead of clear action. While 'read_' and 'list_' and 'search_' are action verbs, the SNCP domain coupling may reduce clarity for agents unfamiliar with CervellaSwarm terminology.