TinyCode has 14 tools with basic schemas and descriptions, but significant gaps prevent higher scores. All tools have descriptions (10-200 chars, mostly adequate), and input schemas are present with type definitions. However, parameter descriptions are sparse or missing context. The bash, write, and edit tools lack error guidance for destructive operations. Output schemas are undocumented. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. The spawn_agent, wait_agent, and close_agent tools are well-described but lack pagination/result limits. Overall structure is functional but falls short of production-grade quality.

• Add tool annotations to all tools: mark read/grep/find/ls/load_skill with readOnlyHint=true; mark write/edit/bash with destructiveHint=true; mark read/ls/grep/find with idempotentHint=true. This enables MCP clients to apply safety policies.
• Document output schemas for all tools. E.g., read returns {content: string, size: number, encoding: string}; bash returns {stdout: string, stderr: string, exitCode: number, duration_ms: number}. Include chaining IDs (e.g., file paths for downstream read/edit calls).
• Enhance parameter descriptions with format/constraint hints. E.g., path: 'File path relative to project root (no ../ traversal allowed)'; pattern in grep: 'ECMAScript regex pattern (e.g., ^import.*from)'; include: 'Glob pattern (e.g., src/**/*.ts)'.
• Add error recovery guidance to write/edit/bash. E.g., bash: 'If command fails, check stderr for syntax errors or missing dependencies. Use grep to verify file state before retrying.' write: 'If path is invalid, use ls to verify directory structure first.'
• Implement result limits and pagination for agent tools. E.g., list_agents returns max 100 agents with a next_cursor; wait_agent truncates reports >10KB with a link to full artifact. Document these limits in descriptions.