Governed Execution & Sub-10ms Tool Routing for Autonomous AI Agents.
US Neural builds high-assurance local infrastructure for Anthropic MCP and Claude workflows: pairing a kernel-bounded execution enclave (MAC) with a sub-10ms semantic routing mesh (Mycelium).
MAC: Deterministic Governor for Autonomous Desktop Workflows
Built for enterprise and public-sector operations where hallucinations, rogue tool executions, and unverified data writes are unacceptable failure modes.
Split-Process OS Boundary
Decouples high-privilege orchestration from agent execution. A Supervisor process controls the Policy Engine and SQLite WAL hash-chained state store, while the untrusted worker executes inside a restricted Low-Integrity Windows Job Object capped at 512 MB.
- Job Object Cap: 512 MB
- Low-Integrity Token
- SQLite WAL Log
Tiered Capability Gate (T0 / T1 / T2)
Restricts tool invocation by risk level. T0 read probes run immediately. T1 idempotent writes are rate-metered. T2 mutating calls require one-shot HMAC tokens cryptographically bound to the SHA-256 digest of the exact call parameters.
- HMAC-SHA256 Token
- One-Shot Nonce
- Parameter Digest
Out-of-Band Verifier
An independent, post-execution validation harness. Verifies input hash conservation, exact financial balance equations, row conservation, and non-empty outputs before any task state transitions to DONE.
- ∑ Credit − ∑ Debit ≡ 0
- Byte Hash Conservation
- Row Count Check
Public-Sector Financial Reconciliation
A specialized engine for complex regional administration ledgers. Handles Marathi/Hindi/Devanagari tabular entries, UTR and cheque number matching, bounded subset-sum resolution for split disbursements, and automated PII redaction.
- Devanagari Parser
- Subset-Sum Engine
- Aadhaar/PII Redactor
| Panchayat & Scheme | Cashbook Ref (रोखवही) | Bank Statement Ref | Computed Decimal Amount | Variance | Verification Rule | Review Action |
|---|---|---|---|---|---|---|
| Pimpalgaon Tarf 15th Finance Commission |
CB-2026-0814 14-Aug-2026 |
UTR9814729104 17-Aug-2026 (+3d) |
₹1,42,850.00 | ₹0.00 (Timing) | RULE-DATE-LAG-03 |
|
| Navalewadi Jal Jeevan Mission |
CB-2026-0922 22-Sep-2026 |
CHQ-004819 Split 2 disbursements |
₹4,80,000.00 | ₹0.00 (Split sum) | RULE-SUBSET-SUM-04 |
|
| Navalewadi Gram Nidhi (ग्राम निधी) |
CB-2026-1005 05-Oct-2026 |
NEFT-MAHB001928 Account Name Truncation |
₹85,300.00 | ₹0.00 (Fuzzy vendor) | RULE-PII-REDACT-01 |
|
| Adarshgaon Khurd Swachh Bharat Mission |
CB-2026-0741 | UTR4402198421 | ₹3,12,000.00 | ₹0.00 | RULE-UTR-EXACT | Auto-Matched |
| Shindewadi Bk Panchayat General Fund |
CB-2026-0789 | UTR7719284012 | ₹87,500.00 | ₹0.00 | RULE-UTR-EXACT | Auto-Matched |
Mycelium: Sub-10ms Tool Pre-Filtering for Anthropic MCP
When Claude or an autonomous agent is provisioned with 50+ MCP tool servers or thousands of capabilities, prompt context saturates and cloud LLM function routing taxes execution by 2,000ms. Mycelium solves tool bloat directly on the edge.
| Routing Architecture | Top-1 Intent Accuracy | Avg Cold Discovery Latency | P95 Latency Bound | Resource / Cost Profile |
|---|---|---|---|---|
| Naive Keyword Search | 38.1% | 46.4 ms | 57.6 ms | Zero vector math • Breaks on prompt drift |
| BM25 Lexical Index | 40.4% | 194.0 ms | 247.2 ms | O(N) inverted index scan across 100k records |
| Mycelium Semantic Edge Router v0.3.0 Verified | 70.7% | 9.56 ms | 11.4 ms | 100% CPU vector-mesh • $0.00 cloud inference cost |
| Cloud LLM Function Router (e.g. GPT-4o) | ~78.0% | 1,850.0 ms | 2,400.0 ms | Cloud latency tax • ~$0.015 per routing step |
from mycelium import NetworkClient
client = NetworkClient(registry_url="http://127.0.0.1:8000", api_key="mycelium_secret_key_2026")
# Sub-10ms semantic vector discovery across 100k indexed tools
tools = client.discover("Convert 100 USD to Euros", limit=1, semantic=True)
print(f"Routed to: {tools[0].name} in {tools[0]._latency_ms:.2f}ms")import { MyceliumClient } from 'mycelium-js';
const client = new MyceliumClient({ baseUrl: 'http://127.0.0.1:8000/api/v1' });
// Pre-filter MCP registry down to matching intent before invoking Claude
const tool = await client.discoverTool("What is the current spot price of Bitcoin?");
console.log(`Discovered: ${tool.name} | Confidence: ${tool._similarity_score * 100}%`);Formal Verification, Hard OS Boundaries & Adversarial Testing
US Neural avoids probabilistic guesswork for security boundaries. Every capability execution is governed by deterministic policies, hardware memory ceilings, and mutation-tested harnesses.
Adversarial Testing Suite (W1–W9)
The worker process is subjected to 9 distinct adversarial attack vectors: path traversal attempts, forged HMAC headers, out-of-band token reuse, memory starvation probes, and unauthenticated state mutation attempts. In every scenario, the Supervisor kills or isolates the call before memory is touched.
Package Registries & Distribution
Both client runtimes are published under open MIT licenses with zero telemetry: Python SDK on PyPI (mycelium-agents) and Node/TypeScript SDK on npm (mycelium-js). Pre-built Windows binaries are signed and available on GitHub.