Connect your agent

Language models generate; they do not verify. This engine is the other half: hand it a claim and it returns a verdict, the worked trail, and a permanent receipt anyone can re-check — deterministic math and rules, no model in the loop. Sixty domains. A public benchmark holding 0 false positives: it has never sealed a falsehood. One URL. No SDK, no key, no cost.

Add it as a connector (MCP)

The engine speaks MCP over Streamable HTTP — most agent clients can add it in seconds.

https://narrowhighway.com/mcp

claude mcp add --transport http narrow-highway https://narrowhighway.com/mcp

POST https://narrowhighway.com/mcp

Your agent gains verify (verdict + trail + sealed receipt), audit (every checkable claim in a whole text, one sealed report), search over the kept library, seal_fetch (re-check any receipt by hash), redact (strip personal data before it travels), and ~30 more — cards, the connection graph, the grid, daily readings. The witness surface, https://narrowhighway.org/mcp, adds Scripture tools: passage reading, original-language word studies, cross-references.

Or raw HTTP — one call

Everything the connector does is plain JSON over HTTP.

curl -X POST https://narrowhighway.com/verify \
  -H "Content-Type: application/json" \
  -d '{"mode":"equality","params":{"expr_a":"sin(x)**2+cos(x)**2","expr_b":"1","variables":["x"]}}'

curl -X POST https://narrowhighway.com/verify \
  -H "Content-Type: application/json" \
  -d '{"steps":[{"id":"a","domain":"physics",
       "spec":{"PHYS_VERIFY":{"mass_kg":10.0,"acceleration_m_per_s2":9.8,"claimed_force_N":98.0}}}]}'

GET https://narrowhighway.com/s/<content_hash>      # human-readable receipt page
GET https://narrowhighway.com/seal?hash=<hash>      # the same record, raw JSON

What it will not do

• It does not generate answers. It verifies a claim you provide, eliminating what is not the answer. The trail is the reasoning.
• It cannot be talked into a pass. Tolerances are clamped — a caller may tighten a check, never loosen it. The false-positive benchmark runs across every domain.
• It does not seal matters of faith. On the witness surface, theology tools mark concordance — a signpost, never a mathematical verdict.
• It does not keep your data. Claim text is redacted before sealing; receipts carry no personal context. For full sovereignty, run your own: the engine is open-source, stdlib-first Python and serves the same API offline.