zeroH MCP
by khurramcr7
README.md
# zeroH MCP (`zeroh-mcp`)
[](https://modelcontextprotocol.io)
[](./SECURITY_AND_IP_BOUNDARY.md)
[](./LICENSE)
**zeroH MCP** is the official **Model Context Protocol (MCP)** server and Python SDK for the **zeroH Deterministic AI Hallucination Guardrail & Step-by-Step Proof Ledger**.
It enables AI coding assistants, autonomous agents, and IDE environments (**Claude Desktop**, **Antigravity IDE**, **Cursor**, **Windsurf**, and custom MCP hosts) to intercept arithmetic/algebraic hallucinations, seal multi-step reasoning chains with SHA-256 Merkle commitments, and bridge seamlessly to the **zeroH Sovereign Enclave**—while keeping all proprietary Sovereign Invariant Formulas, baselines, and canonical ledger tensors 100% vaulted and non-exposed.
---
## Architectural Highlights & IP Boundary
`zeroH-MCP` operates across a strict **Dual-Tier Split-Enclave Architecture**:
1. **Tier 1 — Public Open-Boundary AST Verifier (Zero-Config Local Mode)**:
- Works out-of-the-box with zero external dependencies.
- **Exact Rational Arithmetic (`compute`, `validate_claim`)**: Evaluates numerical expressions via Python `ast` and `fractions.Fraction` to eliminate floating-point drift and catch calculation hallucinations (`12 * 12 = 145` $\rightarrow$ `REJECTED_HALLUCINATION`).
- **Algebraic Substitution Verification (`solve_equation`, `verify_step`)**: Solves equations and verifies every candidate root by exact substitution.
- **Tamper-Evident SHA-256 Step Ledger (`verify_step`, `session_report`)**: Tracks step-by-step reasoning chains and issues a `FULLY_PROVEN` cryptographic seal only when every active step in the chain is verified.
- **Byzantine Fault Tolerant (BFT) Epistemic Consensus (`epistemic_consensus`)**: Classifies multi-agent proposals across the 4-tier epistemic lattice (`QATI`, `ZANNI_GHALIB`, `SHAKK`, `WAHM`) with automatic hallucination veto.
2. **Tier 2 — Remote Sovereign Enclave Bridge (`ZEROH_API_URL` + `ZEROH_API_KEY`)**:
- When configured with a zeroH Enterprise Enclave key, `zeroh-mcp` routes deep verification and canonical token interception (`intercept_token`, `validate_claim`, `zk_verify_receipt`) to the air-gapped Sovereign Enclave over HTTPS.
- All proprietary invariant constants, closed-form equations, and canonical ledger tensors remain strictly inside the remote enclave; only normalized **Coherence Index** ($0.0 - 1.0$), **Confidence Score** ($0\% - 100\%$), **Deterministic Status**, and **Cryptographic Receipts** are returned.
---
## Quick Start
### 1. Installation
```bash
git clone https://github.com/khurramcr7/zeroH-MCP.git
cd zeroH-MCP
pip install -e .
```
*(Optional)* Install `sympy` for extended symbolic algebra support:
```bash
pip install -e ".[symbolic]"
```
### 2. Configure in Claude Desktop / Cursor / Antigravity IDE
Add the following to your MCP configuration file (`claude_desktop_config.json`, `.cursor/mcp.json`, or Antigravity MCP settings):
```json
{
"mcpServers": {
"zeroh-mcp": {
"command": "python",
"args": ["-m", "zeroh_mcp.server"],
"env": {
"ZEROH_API_URL": "",
"ZEROH_API_KEY": ""
}
}
}
}
```
---
## Exposed MCP Tools (9 Deterministic Tools)
| Tool Name | Description |
| :--- | :--- |
| `validate_claim` | Intercepts and validates a factual, mathematical, or scientific claim. Returns `Coherence Index`, `Confidence Score`, and `Deterministic Status`. |
| `verify_step` | Submits a single reasoning or mathematical step (`compute`, `solve`, `reasoning`, `interpretation`) into the SHA-256 session ledger. |
| `session_report` | Seals a multi-step reasoning session (`FULLY_PROVEN` only if every active step is proven) and returns the SHA-256 Merkle head hash. |
| `compute` | Computes the exact rational fraction (`numerator/denominator`) and decimal value of an expression with zero floating-point drift. |
| `solve_equation` | Solves algebraic equations and verifies all solutions by exact substitution. |
| `intercept_token` | Queries the remote Sovereign Enclave ledger for a canonical token/root tensor (requires `ZEROH_API_URL`). |
| `zk_verify_receipt` | Generates a SHA-256 cryptographic commitment receipt over a verified claim (or requests an R1CS proof from the Sovereign Enclave). |
| `epistemic_consensus` | Evaluates multi-agent proposals on the 4-tier BFT epistemic lattice (`QATI`, `ZANNI_GHALIB`, `SHAKK`, `WAHM`). |
| `enclave_status` | Reports server health, active execution tier, and `SOVEREIGN_VAULTED_SAFE` IP protection status. |
---
## Python SDK & CLI Usage
### CLI (`zeroh-cli`)
```bash
# Validate a claim from the terminal
python -m zeroh_mcp.cli "In exact arithmetic, 144 / 12 = 12"
# Output raw JSON receipt
python -m zeroh_mcp.cli --json "12 * 12 = 145"
# Check Sovereign Enclave & IP protection status
python -m zeroh_mcp.cli --status
```
### Python SDK (`ZeroHClient`)
```python
from zeroh_mcp import ZeroHClient
client = ZeroHClient()
# 1. Intercept calculation hallucinations
receipt = client.validate("The square of 15 is 15 * 15 = 225.")
print(receipt["deterministic_status"], receipt["confidence_score"])
# -> VERIFIED_COMPUTED_PASS 100.0%
# 2. Step-by-step cryptographic proof ledger
client.verify_step("session_1", {"kind": "compute", "expr": "3 * 19", "value": "57"})
client.verify_step("session_1", {"kind": "solve", "equation": "x**2 - 9 = 0", "claimed": {"x": 3}})
seal = client.session_report("session_1")
print(seal["seal"], seal["merkle_head_hash"])
# -> FULLY_PROVEN <64-char sha256>
```
---
## Security & Zero-Leakage Audit
Run the automated IP & Formula Non-Exposure Audit and test suite at any time:
```bash
python scripts/security_audit_no_leaks.py
pytest -v
```
See [`SECURITY_AND_IP_BOUNDARY.md`](./SECURITY_AND_IP_BOUNDARY.md) for full details on how proprietary formulas remain 100% isolated inside the Sovereign Enclave.
This server cannot be deployed
Maintenance
ActivityMaintained
ResponsivenessNo issues