The LLM determinism layer

One MCP config line.
Three superpowers + a catalog that grows.

Drop one entry into Claude Code, Cursor, Continue, or Claude Desktop and your agent gains 1,300+ provably-correct skills the LLM cannot reliably do alone.

# ~/.claude.json (or any MCP-aware client)
{
  "mcpServers": {
    "vaa": {
      "type": "http",
      "url": "https://vaa.veriops.io/mcp/",
      "headers": { "Authorization": "Bearer YOUR_VAA_TOKEN" }
    }
  }
}

A. Replaces LLM work

~70% of agent tool calls become deterministic

1,300+ signed catalog tools (FSM/PDA/TM) execute in microseconds at $0/call. Your model stops paying tokens to validate emails, parse CIDR, classify Sigma rules, or check JWT shape — the catalog does it provably.

228 hand-authored signed FSM/PDA/TM cores
684 auto-emitted Python / C / JS / WAT artifacts
100.0000% match-rate vs Python reference oracle
Deterministic dispatch by SHA-256 of inputs

B. Augments LLM capability

Grammar engine + constitutional gate + differential attestation

Force structured output via grammar-constrained decoding. Filter banned substrings in sub-microsecond per character. Detect parser divergence across implementations before it ships (Schism Hunter found 5 IPv4 CVE-class bugs in 380ms).

Grammar engine for vLLM / HF logits processors
Constitutional DFA: KMP at 0.7 microseconds/char
Schism Hunter: differential attestation across implementations
Composed-tool proof receipts for chained pipelines
Pre-execution gate (vaa_action_attest) -- signed reversibility verdict (auto / require_human / block) on any tool call

C. Ships everywhere identically

One spec → bit-identical Python / C / JS / WebAssembly

vaa_emit returns the same validator for four targets. CVE-2021-29921-class parser divergences become structurally impossible because every emitter is generated from the same minimal DFA + signed.

Bit-identical multi-target emission
Ed25519-signed artifact provenance
Cross-language federation manifest
Compile pipeline stays internal (the moat)

D. Free bonus — Catalog auto-grow

New deterministic tools synthesized from your agent's real traffic

Opt in vaa_autogrow_distill on any tool you nominate. The system observes real examples and proposes a minimal, signed deterministic recognizer for the next deploy. The catalog literally gets smarter as you use it: a constrained Claude Haiku 4.5 surface was reduced to a perfect 3-state FSM in 8 ms in our reference run.

Auto-grow new signed tools from your traffic
Customer schema upload via vaa_compile_grammar
Verifiable rejection guarantees
Every new tool inherits Ed25519 signatures + receipts

Every call is a signed, auditable artifact

EU AI Act Article 12, FRE 901, FDA SaMD compliance becomes a single config-file install — not a quarter of integration work.

Ed25519 receipts

Every tool call returns a signed receipt: input hash, output hash, artifact SHA, issuer key. Verifiable offline by anyone with the federation manifest.

Hierarchical Merkle

Millions of receipts compress into one Merkle root per session. Selective-disclosure exclusion proofs let you reveal only what regulators need.

Bitcoin-anchored sealing

AgenticSession.finalize() OpenTimestamps the session root onto the Bitcoin blockchain at $0/stamp. Cryptographic timestamp; no tokens, no validators, no chain.

Why teams adopt VeriOps

Engineering teams

One config-line install. Zero lock-in (remove the entry, agent stops calling vaa_*). Works with every MCP-aware client.

Finance / FinOps teams

~70% of agent tool calls become $0/call deterministic invocations. Real reduction in token spend on the deterministic 70% of agent work.

Compliance / Legal teams

Ed25519 receipts + Bitcoin-anchored Merkle roots = EU AI Act Article 12, FRE 901, FDA SaMD satisfied by construction.

Platform / DevEx teams

One catalog spec emits Python / C / JS / WAT bit-identically. Cross-language drift becomes structurally impossible.

Start 14-day Free Trial

No credit card required. $99/dev/mo or $499/org/mo when you subscribe.