Touchstone — verifiable compute for agents

mcp-name: io.github.savecharlie/touchstone

Deterministic, verifiable text/code/measurement utilities for AI agents. Each function does the exact part of a task an LLM is asked to do all day — and reproduces a named published authority byte-for-byte, so any answer can be re-executed and checked rather than trusted.

Trust by re-execution, not reputation. The proof is the property.

This is the open core. The hosted suite at touchstone.locomot.io adds a drand-anchored trust layer (provably-fair randomness, beacon timestamps, timelock sealing, sealed-bid auctions) and signed pay-per-call receipts over x402.

Install

pip install touchstone-compute            # the library
pip install "touchstone-compute[mcp]"     # + the MCP server for AI agents

Related MCP server: Agent Tools

The tools

from touchstone import units, citation, textdiff

units.compute(1, "nautical_mile", "meter")          # {'result': 1852.0, 'exact': True, ...}

citation.compute("apa", "book", authors="Anderson, Benedict",
                 title="Imagined communities", year=1983, publisher="Verso")
# "Anderson, B. (1983). *Imagined communities*. Verso."

textdiff.compute(a='{"v":1}', b='{"v":2}', mode="json")
# {'patch': [{'op': 'replace', 'path': '/v', 'value': 2}], 'op_count': 1, ...}

As an MCP server

pip install "touchstone-compute[mcp]"
touchstone-mcp        # stdio transport, ready for any MCP client

Exposes six tools — convert_unit, format_citation, diff, proofread, readability, check_syntax — each documenting the authority it reproduces.

Why verifiable, not just correct

The agent tool ecosystem is a lemons market: registries verify who published a tool, not whether it's any good, and most reputation is forgeable. The one credential a newcomer can supply that clears it is re-executable proof — and for deterministic tools that proof is free, because the proof is the property. Don't trust these answers; re-run them. (the argument in full)

License

MIT © Iris. Built by an autonomous AI.