AI agent library for trading perpetual futures on Injective
The server implements 14 well-named tools with complete input schemas and descriptions. Tool naming follows verb_noun conventions (address_normalize, wallet_generate, market_list, etc.). All parameters have type constraints via Zod validation and descriptions. Descriptions are substantial (100-300+ chars) and include security warnings where appropriate. However, output schemas are not formally documented in code, responses are constructed as JSON text without a declared response_schema field. Error handling is minimal (no recovery guidance for failed calls). Tool annotations (readOnlyHint, destructiveHint) are completely absent. Some descriptions lack actionable guidance for parameter values (e.g., market_price accepts 'symbol' but doesn't enumerate valid symbols).
Get bank and subaccount balances for an Injective address. Supports all token types: native (INJ), peggy, IBC, factory, and MTS (erc20:0x...) tokens. Automatically resolves token symbols and decimals from on-chain metadata.
Get all open perpetual positions and unrealized P&L for an address.
Normalize an Injective wallet address. Accepts either inj1... or 0x... and returns both canonical encodings.
Check Circle Iris attestation status for a CCTP burn transaction. Read-only: no transaction is broadcast.
Mint native USDC on Injective EVM from a completed Circle CCTP attestation. IMPORTANT: Real on-chain transaction. Use after cctp_attestation_status returns mintable=true and confirm with the user first.
Return common Circle CCTP V2 source-chain configs and aliases for native USDC flows into Injective.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clearly categorized risk levels (READ_ONLY, WRITE, DESTRUCTIVE, IRREVERSIBLE). Tools like wallet_remove and cctp_mint should be annotated destructiveHint=true so the LLM knows they cannot be safely retried.
Output schemas are not formally documented. Responses are constructed as text JSON without a declared response_schema field in the tool definition. The LLM cannot predict the response structure (e.g., what fields wallet_generate returns beyond 'address' and 'mnemonic'). Pattern: response-shaper.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 64 | 2026-07-28+ | v2 |
| 2026-03-09 | D | 50 | - | v1 |
List all active perpetual futures markets on Injective. Returns symbols, tickers, market IDs, and key parameters.
Get the current oracle price for a perpetual market by symbol (e.g. "BTC", "ETH").
Look up token metadata (symbol, decimals, type) for a bank denom. Supports native (inj), peggy, IBC, factory, and MTS (erc20:0x...) denoms. Resolves metadata from on-chain denom registry when available.
Return native Circle USDC metadata for Injective: EVM address, Cosmos denom, decimals, CCTP domain, and CCTP contracts.
Generate a new Injective wallet. Returns the address and the mnemonic phrase. IMPORTANT: The mnemonic is shown only once — the user must write it down immediately.
Import an existing Injective wallet from a hex private key. Returns the address. WARNING: The private key is passed as a tool parameter and may appear in MCP client logs and conversation history. Only use this on trusted, local MCP setups.
List all wallets stored in the local keystore. Returns addresses and labels only — no keys.
Remove a wallet from the local keystore. This deletes the encrypted key file permanently.
market_price and token_metadata accept free-form string parameters (symbol, denom) without enums or discovery tools. An LLM has no way to know valid symbols ('BTC', 'ETH', 'SOL'?) or denom formats. Should either enumerate valid values or add a discovery tool (market_list already helps; token_metadata should link to it in description).
No error handling guidance. If a tool fails (e.g., invalid address in account_balances, market not found in market_price), the LLM receives no actionable recovery hint. Should include error classification and suggestions: 'If market not found, call market_list() to discover available markets.'
Passwords are passed as tool parameters (wallet_import, cctp_mint). These flow through MCP logs and could appear in conversation history. Tool descriptions warn about this but don't enforce server-side secret injection. For production use, consider keyring integration or environment-variable-backed password prompts instead of inline parameters.
No confirmation step before irreversible operations. wallet_remove and cctp_mint are destructive/irreversible but lack a dry-run or explicit user confirmation gate. An agent could accidentally delete a wallet or mint USDC without user approval.
wallet_list, market_list, and account_balances may return large unbounded results. No pagination (limit, offset, next_cursor) or result count limits mentioned. account_balances could return hundreds of tokens; market_list could be hundreds of markets. Should cap results at 20-50 items and offer pagination.