Static source inference · medium confidence · evidence: Streamable HTTP
Current-spec patterns detected
Summary
This server has basic tool definitions but lacks critical quality standards for LLM-safe agent tooling. While all 8 tools have names and descriptions, parameter schemas are incomplete, descriptions are superficial, output schemas are entirely undocumented, and error handling provides no recovery guidance. The descriptions (10-24 characters) fall well below the 50-200 char baseline for LLM-optimized tools. No parameters include type information in the tool registration (inferred from code review, not explicit schema). The HTTP transport is good, but definition quality is significantly below production standards.
Tool descriptions are 10-24 characters; far below the 50-200 char LLM-optimized baseline. 'Get account information' and 'Get order book offers' provide no WHEN/WHY context or disambiguation from similar tools. LLMs cannot select tools reliably without descriptive guidance.
Parameter schemas lack explicit type definitions. Code inspection shows 'account', 'ledger_index', 'hash', 'tx_blob', 'taker_gets', 'taker_pays' are passed as strings, but the tool registration does not expose JSON Schema with 'type' fields.
Expand all tool descriptions to 60-150 characters. Format: 'Get [resource] from [source]. Returns [key fields]. Use when [context/difference from similar tools].' Example: 'Get account trust lines on XRPL. Returns list of currencies the account has open. Use to check collateral before trading.'
Add explicit JSON Schema to every tool with full parameter types, descriptions, enums, and min/max constraints. Example: {'account': {'type': 'string', 'pattern': '^r[a-zA-Z0-9]{25,34}$', 'description': 'XRPL account address (r...)'}}, {'ledger_index': {'type': 'string', 'enum': ['validated', 'closed'], 'description': 'Ledger version (validated=latest confirmed, closed=previous ledger)'}}.}
Document output schemas for every tool. Create a schema definition or prose description. Example: account_info returns {account, Balance (drops), Flags, PreviousTxnID, PreviousTxnLgrSeq, Sequence, ...}. Specify types: Balance is a number (drops), Flags is a number (bitfield), Sequence is a number.
Add pagination parameters to account_lines, account_nfts, account_transactions: 'limit' (1-100, default 20), 'marker' (opaque cursor from previous response for next page). Return {items: [...], marker: '...' | null, total_count: number}.
Add destructive operation safeguards: (a) Add 'dryRun' boolean parameter to submit_transaction (default false). (b) When dryRun=true, validate tx_blob and return validation result without submitting. (c) Annotate submit_transaction with destructiveHint: true in tool metadata. (d) Return clear confirmation message: 'Ready to submit transaction. Cost: X drops. Call again with dryRun=false to confirm.'
No output schemas documented anywhere. LLMs cannot infer what fields to expect in the response (e.g., account_info returns account, balance, flags, sequence, but this is not declared). Agents must guess at response structure or fail downstream.
Parameter descriptions are missing or trivial. 'The account address' and 'Ledger index (defaults to validated)' are bare minimum, no format hints, no examples of valid values, no guidance on what happens if the parameter is wrong. Descriptions must answer: format, range, constraints, and what happens on missing input.
No error handling or recovery guidance. When a handler raises an exception (e.g., invalid account, network timeout), the error response is a raw exception string. 'Invalid account format. Use valid XRPL addresses (r...)' is actionable; raw stack traces are not.
submit_transaction is a destructive write operation but has no destructive hint, confirmation step, or dry-run option. An agent could accidentally submit real transactions without a safety gate. tool annotation 'destructiveHint' is missing.
No pagination support on list-like tools (account_lines, account_nfts, account_transactions). These queries can return hundreds of results. No limit parameter, no page/offset, no next_cursor, agents cannot fetch results in batches.
All read tools (account_*) lack metadata about which fields are required vs. optional, and accept 'ledger_index' with vague guidance. No mention of what valid ledger_index values are ('validated', 'closed', a block number). This forces LLMs to guess or fail.
book_offers has three required parameters but no description indicates which fields (taker_gets, taker_pays) are required vs. optional, or what format they accept (currency codes, asset objects, full JSON). The description does not disambiguate this tool from potential similar queries.
book_offers
Implement structured error handling. Replace bare exception strings with typed error objects: {error: 'INVALID_ACCOUNT', message: 'Account address must match r[a-zA-Z0-9]{25,34}', recovery: 'Use a valid XRPL account address or call account_search first.'}. Include error classification: retryable (network timeout), user-fixable (invalid input), or fatal.
Add format constraints and examples to parameter descriptions. Example: 'Ledger index: use 'validated' for latest confirmed ledger, 'closed' for previous ledger, or a specific ledger sequence number (integer). Example: 'validated' or '12345678'.'
For read-only tools, add tool annotations (per MCP 2026-07-28): readOnlyHint: true. This signals to the client that no state is modified.
Add timeout handling. Wrap xrpl.client calls with explicit timeouts (e.g., 10s) and return clear timeout errors: 'XRPL node request timed out after 10s. Node may be overloaded. Retry after 5 seconds or switch to another node.'
Create a result limit policy: cap account_lines, account_nfts, account_transactions at 20 results by default. State in description: 'Returns up to 20 most recent results. Use the marker parameter to fetch additional pages.'
Add parameter validation inside handlers. Before calling xrpl.client, validate: (1) Account format matches XRPL pattern. (2) Ledger index is 'validated', 'closed', or a valid number. (3) Hash is valid hex. Return actionable errors instead of letting xrpl-py raise exceptions.