MCP server that exposes a tool catalog with search, describe, and execution capabilities. Allows AI assistants to discover and execute tools against downstream APIs via MCP protocol.
Agent Atlas provides 3 tools with clear purpose and reasonable schema documentation. Tool naming follows verb_noun convention (SearchTools, DescribeTool, ExecutePlan), which is good. All tools have descriptions ranging from 180-320 characters, meeting the baseline average (194 chars). Input schemas are visible and typed. However, parameter descriptions are generic and lack actionable detail; output schemas are not explicitly documented; and error handling guidance is absent. The ExecutePlan tool presents a significant composition concern: it combines multiple responsibilities (dry-run validation, plan execution, multi-mode operation) into a single tool, violating the single-responsibility principle.
Get full metadata for a specific tool including its request/response schema, required downstream permissions, and access entitlement hints.
Execute an Atlas plan against one or more downstream API tools. Use mode='dryRun' to validate and preview the execution plan without making HTTP calls, or mode='run' to execute it. The caller's JWT is forwarded to downstream APIs; Atlas does not enforce downstream permissions.
Search the Agent Atlas tool catalog. Returns matching tools with their metadata including method, path, safety level, and required downstream permissions (informational).
Parameter descriptions lack actionable detail and format constraints. Example: 'query' param in SearchTools is described only as 'Optional free-text query...' with no specification of expected length, format, or examples of valid queries. 'safety' param mentions three enum values ('read', 'write', 'destructive') but does not state these as a constrained enum in the schema or describe when to use each filter.
Output schemas are not documented. SearchTools and DescribeTool return tool metadata, but the structure (fields, types, whether paginated, total counts) is inferred from the tool description, not formally specified. This forces LLMs to guess what fields to expect in responses and makes chaining unreliable.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 57 | 2026-07-28+ | v2 |
| 2026-03-09 | D | 54 | - | v1 |
ExecutePlan combines multiple responsibilities: dryRun validation, execution, mode selection, and environment switching. This violates single-responsibility design. An LLM selecting ExecutePlan may pick the wrong mode or expect to receive validation details that are not returned. Should split into separate tools: validate_plan (dryRun mode) and execute_plan (run mode).
No error handling or recovery guidance. The description states 'Atlas does not enforce downstream permissions' but does not explain what error will occur if the caller lacks permission, how to recover, or what the LLM should do next. Missing errors for invalid toolId, malformed plan JSON, timeout, and permission denial.
SearchTools 'accessFilter' parameter depends on JWT claims but is not documented in the parameter description. The description mentions 'requires JWT claims' in parentheses but does not explain what claims, what happens if they are absent, or how the filter behaves when passed 'granted' vs 'missing' when claims are not present.
ExecutePlan 'plan' parameter is documented as 'JSON object' but accepts a structured plan DSL (with 'steps', 'type', 'toolId', etc.). The description provides an inline example but does not formally specify the schema of the plan structure, required fields, valid step types, or constraints on nesting depth. LLMs must reverse-engineer the plan schema from the single example.
SearchTools and DescribeTool do not document pagination. If results exceed a threshold, is pagination supported? Are there limit/offset parameters? Does DescribeTool return the full request/response schema or a summary? Without clarity, LLMs cannot predict whether large tool catalogs will be returned in full or truncated.