Skip to main content
Mako has 23 built-in tools that she calls autonomously during inference. Tool execution happens entirely server-side in the gateway — you receive the final answer, not the raw tool calls.

​
Tool categories

​
How tool calling works

When you send a message to the chat completions endpoint, the gateway:
  1. Sends tool definitions to the model alongside your message. The model sees all 23 tools and their parameter schemas.
  2. Checks the response for tool calls. If the model outputs a structured tool call (or embeds one in XML/JSON within its text), the gateway parses and executes it.
  3. Feeds results back. The tool result is formatted into human-readable text and added to the conversation as a tool message.
  4. Loops. The model sees the tool result and can either respond or call another tool. This repeats for up to 8 rounds.

​
Auto-chaining

The gateway adds intelligence on top of the model’s tool calls:
  • Transaction forensics: When a tx hash is detected, the gateway automatically fires get_transaction, get_transaction_receipt, and decode_tx_logs in parallel — giving the model the complete picture (basic details, execution outcome, and decoded token transfers) without requiring multiple turns.
  • Address profiling: When an address is detected, lookup_address fires alongside balance and contract checks to resolve ENS names and known protocol labels. If the user asks about “history” or “activity”, get_address_history is also triggered.
  • Contract inspection: When the user mentions “source”, “code”, “abi”, or “verified” alongside an address, get_contract_source fires automatically.
  • Tweet auto-reading: When web_search returns results containing Twitter/X URLs, the gateway automatically calls read_tweet on those URLs without waiting for the model to request it.
  • Page auto-extraction: The top non-social-media URL from search results is automatically extracted via web_extract for additional context.
  • Stall detection: If the model responds with “let me check that” without actually calling a tool, the gateway nudges it with a direct instruction to make the appropriate tool call.

​
Tool timeout

All tools have a 15-second timeout. If the underlying RPC call or API request takes longer, the tool returns an error and the model continues with whatever information it has.

​
All tools