Strip Secrets Before Prompts Leave Your Rig. Freeze Rogue MCP Commands in 0.21ms.
Cursor, Claude Code, Windsurf, and MCP servers run with full filesystem and shell privileges. One indirect prompt injection can silently exfiltrate ~/.ssh, .env, or AWS keys—or wipe a repo. Agent Sentinel (C1-AG) intercepts every LLM prompt and MCP tool call 100% locally on your CPU.
ANTHROPIC_BASE_URL=http://127.0.0.1:8787/anthropic # <0.6ms ONNX Token Redactor AgentSentinel.exe mcp-wrap -- npx -y @modelcontextprotocol/server-filesystem ./src
Cloud LLMs Never See Your Keys. Your Local Code Never Breaks.
Inspect how C1-AG intercepts three real-world AI coding failure modes locally on 127.0.0.1 without routing proprietary source code through a third-party SaaS gateway.
Engineered for Autonomous Coding Agents
Sub-Millisecond Secret & PII Redactor
Embedded quantized local ONNX token classifier + structural recognizers strip OpenAI, Anthropic, GitHub, AWS, Stripe, JWT, DB URIs, and Luhn payment cards in 0.58 ms while leaving 40-char Git commit SHAs and UUIDs untouched.
Split-Chunk Safe Response Restoration
Replaces secrets deterministically per session (<CYZOG_SECRET_1>) and restores them across chunked text/event-stream SSE responses so generated code compiles locally with zero manual copy-pasting.
JSON-RPC 2.0 Credential Store Guard
Transparent mcp-wrap proxy inspects every tools/call and blocks silent reads to ~/.ssh, ~/.aws, ~/.kube, Chrome/Edge Login Data, and crypto wallets.
1-Click Blast-Radius Approval HUD
Watches child processes spawned by Cursor, Claude, Windsurf, and VS Code. Freezes rm -rf, Remove-Item -Recurse, DROP TABLE, and git reset --hard in 0.21 ms for 1-click HUD approval or termination.
Indirect Prompt-Injection Neutralizer
Scans untrusted MCP scraper, repo, and database tool outputs for hidden instruction overrides, Unicode tag smuggling (U+E0000), and Markdown image exfiltration beacons ().
SHA-256 Chained Compliance Ledger
Proves to security auditors that zero secrets leaked to external LLMs using an append-only SHA-256 hash-chained JSONL audit log—without ever writing raw secret strings to disk.
Impulse-Priced for Solo Devs. Audit-Ready for Security Teams.
100% offline Ed25519 cryptographic activation. Zero cloud accounts, zero prompt logging servers.
Agent Sentinel Personal Pro
Cursor, Claude Code, Windsurf & VS Code
- [✓]Loopback
127.0.0.1:8787Proxy for OpenAI, Anthropic, Gemini & OpenRouter - [✓]Sub-3ms ONNX Token NER Redactor (
<CYZOG_SECRET_1>) + SSE Rehydration - [✓]Stdio
mcp-wrapFirewall Blocking~/.ssh,~/.aws& Browser Vaults - [✓]
NtSuspendProcess/SIGSTOPRogue Shell Freeze + 1-Click Approval HUD - [✓]1-Click IDE Auto-Shield Wizard with Byte-Exact Backup & Restore
- [✓]Licensed for up to 3 Personal / Freelance Workstations
Team Compliance Edition
Engineering Orgs, FinTech, Defense & Healthcare
- [+]Everything in Personal Pro across Windows, macOS & Linux developer fleets
- [+]Signed Central Policy JSON (enforce mandatory MCP allowlists & PII redaction)
- [+]Tamper-Evident SHA-256 Hash-Chained DLP Audit Logs for SOC2 / HIPAA / ISO 27001
- [+]Local Syslog / SIEM CEF Telemetry Export (Splunk, Sentinel, Wazuh)
- [+]Custom Regex & Proprietary Internal Codeword Tokenizer Rules
- [+]Offline Multi-Seat Ed25519 Fleet Certificate & Priority Engineering SLA
Runtime & Integration Ledger
- Supported AI IDEs & Agents
- Cursor, Claude Code, Claude Desktop, Windsurf, VS Code, Cline, Continue, OpenClaw
- Supported Upstream LLM APIs
- OpenAI, Anthropic, Google Gemini, OpenRouter, Groq & Custom OpenAI-Compatible
- Embedded Neural Engine
- Quantized Local ONNX Token NER (0.58 ms CPU · Native x64 Binary · No GPU Required)
- Network Binding Guarantee
- Strict 127.0.0.1 Loopback (Refuses 0.0.0.0 LAN bind by design)
- Operating Systems & Packaging
- Windows 10/11 x64 Native Machine-Code Binary (.exe HUD + CLI) · macOS/Linux Fleet Binaries (Team Edition)