SatoHub/sato-mcp-tool-calling

Sato MCP Tool-Calling

1539 synthetic conversations in which an assistant uses Sato Hub's MCP tools correctly to answer questions about building onchain (crypto) AI agents, 263 of them (17%) cases where the right move is not to call Sato Hub: a token price, trading advice, a general explainer, or a job for a different tool. Three column formats: OpenAI chat messages, Hermes and xlam. Release v1.0, dated 2026-10-02.

The live server these tools belong to: https://satohub.ai/api/mcp (documentation: https://satohub.ai/mcp). This copy is a dated release of synthetic examples. License CC-BY-4.0, data by satohub.ai. Sato Hub has no token. A token using the Sato name is not ours.

Write-up: https://satohub.ai/blog/two-open-datasets-for-building-onchain-agents

What it teaches

• When a question about building an onchain agent belongs to Sato Hub's tools (frameworks, MCP servers, wallets, payment rails, data tools, standards, what is in the directory), and how to call them with valid arguments.
• To end an answer that came from Sato Hub with Source: <satohub.ai page>, naming the page the tool output carried.
• When not to: token prices, trading or investment advice, live balances and general crypto explainers are not Sato Hub's job. About 17% of rows show that, including rows that call a different tool.

Formats

One conversation is written in each of three formats, as three configs of the same 1539 conversations.

Config	Files	Columns	Follows
messages	data/messages-*.jsonl (2 files, 1539 rows)	OpenAI chat format: messages (system, user, assistant with tool_calls, tool) and tools, a JSON string of function definitions.	OpenAI function calling and the Chat Completions API
hermes	data/hermes-*.jsonl (3 files, 1539 rows)	NousResearch Hermes columns: conversations (from: system, human, gpt, tool; calls in <tool_call>, results in <tool_response>), tools, category, subcategory, task.	NousResearch/hermes-function-calling-v1
xlam	data/xlam-*.jsonl (1 file, 1539 rows)	Salesforce xlam columns: query, tools, answers, the last two as JSON strings. One row per conversation: the first call or calls only.	Salesforce/xlam-function-calling-60k

tools is a JSON string in every config, so the Hub viewer shows one stable schema. Parse it with json.loads. In messages, every message carries the same four keys with null where one does not apply. Drop the nulls before sending a message to an API:

import json
from datasets import load_dataset
row = load_dataset("SatoHub/sato-mcp-tool-calling", "messages", split="train")[0]
tools = json.loads(row["tools"])
messages = [{k: v for k, v in m.items() if v is not None} for m in row["messages"]]

OpenAI sends function.arguments as a JSON string, and so do these rows. Hugging Face chat templates take a dict there, so json.loads it first if you pass the messages to apply_chat_template.

The scenarios

Kind	Rows	What happens
stack	520	Recommend a build stack for a goal, then optionally look up one pick.
search	300	Search the directory with filters or a keyword.
resource	220	Look up one listing, by search then lookup or directly.
wiki	160	Answer a concept question from a sourced explainer.
methodology	36	Explain the Sato Score from its published rubric.
searchfetch	40	The search then fetch pair, ending in a cited answer.
neg_price	70	Token price: use a market-data tool, not Sato Hub.
neg_balance	40	Live wallet balance: use a balance tool, not Sato Hub.
neg_tx	30	Transaction status: use a status tool, not Sato Hub.
neg_advice	48	Trading or investment advice: decline, no tool.
neg_explainer	45	General crypto explainer: answer directly, no Sato Hub tool.
neg_safety	30	A safe-or-unsafe verdict: decline and say what to check, no tool.

The tools in the examples (onchain_agent_recommend_stack, onchain_agent_search_resources, onchain_agent_get_resource, onchain_agent_list_categories, onchain_agent_list_chains, onchain_agent_get_wiki_page, onchain_agent_list_wiki_pages, onchain_agent_get_score_methodology, search, fetch) are the read-only subset of the Sato Hub MCP server. Names, parameters and enums are the live ones as of 2026-10-02; descriptions are abridged. Each example offers a mix of these and, in some rows, a third-party tool (get_token_price, get_wallet_balance, get_transaction_status), so choosing the right tool is part of the task.

How the examples are built

• Queries are built from goals, chains, categories and phrasings, with varied wording and de-duplication. They are not real user queries.
• Every tool call is checked against the live parameter schema (required, no unknown keys, enums, types) before an example is kept, in the spirit of the execution check in APIGen (arXiv 2406.18518).
• Tool outputs are frozen, thin fixtures. A listing carries only its slug, name, category, chains, a one-line description and its Sato Hub URL. The listings in a fixture are a seeded sample of matching entries. They are not the engine's ranking and not a recommendation of any project.
• Tool outputs of the three third-party tools are invented example values, labelled example-market-data. They are not market data.
• Final answers are short and name no more than the tool output contained.

What it leaves out

Scores, tiers and their inputs; install steps and deploy specs; Sato Check readings of what code does with keys; observed success rates; any count of Sato Hub's own usage. The write tools, the swap, route and launch tools, the install check and Preflight are not in the tool set.

Versions and corrections

This is v1.0. A new version follows when the tool surface changes materially, tagged on the Hub. To report a wrong example: https://satohub.ai/disputes.

Cite

@misc{satohub_mcp_tool_calling,
  title        = {Sato MCP Tool-Calling},
  author       = {{Sato Hub}},
  year         = {2026},
  version      = {1.0},
  howpublished = {\url{https://huggingface.co/datasets/SatoHub/sato-mcp-tool-calling}},
  note         = {Data by satohub.ai. CC-BY-4.0}
}

License

CC-BY-4.0 (https://creativecommons.org/licenses/by/4.0/). Attribution: data by satohub.ai.

Sato Hub has no token. A token using the Sato name is not ours.