Mithril MCP has clear, action-oriented tool names (all start with verbs: get_, place_) and reasonable descriptions (80 - 180 chars, within baseline 34 - 392 range). All 5 tools have input schemas with type definitions and parameter descriptions. However, output schemas are completely undocumented, the code returns raw JSON strings from the Mithril API without specifying what fields agents should expect. Error handling is present but generic (humanize() function maps codes to English, but doesn't classify errors as retryable/user-fixable/fatal). Parameter descriptions lack format constraints (e.g., 'market' accepts '{BASE}-USD-PERP' but no enum or regex pattern is declared). No pagination support visible for list-returning tools (mithril_dex_get_markets, mithril_dex_get_fills). Tool composition is reasonable, each tool has one responsibility, but no idempotency guarantees stated.

• Document output schemas for all tools. Specify which fields are returned (e.g., for get_balance: {balance: number, margin: number, available_margin: number}). Include field types and descriptions so agents know what to extract.
• Add enum or regex constraints to 'market' parameter. Define valid format as enum of supported markets or regex pattern '^[A-Z]+-USD-PERP$' to prevent hallucinated invalid values.
• Implement pagination for list tools. Add 'limit' (default 20, max 100) and 'offset' parameters to mithril_dex_get_markets and mithril_dex_get_fills. Return total_count in response.
• Classify errors in humanize() output. Append error classification: '(retryable)', '(user-fixable)', or '(fatal)' so agents know whether to retry, ask user, or abort.
• Add dry-run mode to mithril_dex_place_limit_order. Accept optional 'dry_run: boolean' parameter (default false). When true, validate order but do not execute, returning estimated outcome. Require explicit confirmation for real execution.