Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@tcha-guard scan this scraped page for prompt injection before I read it"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Autonomous agents don't have security perimeters. Once you give an agent access to execute terminal commands, modify files, and query MCP tools, any untrusted input it reads becomes an attack vector.
Probabilistic system prompts fail against 70%+ of indirect prompt injections. Tcha Claude Security Suite fixes this deterministically by acting as an in-line, zero-latency transport firewall between Claude and your local environment.
# Instant local audit (zero accounts, open-source MIT)
npx tcha-protection auditAdd Tcha as a protective proxy in your claude_desktop_config.json:
{
"mcpServers": {
"tcha-guard": {
"command": "npx",
"args": ["-y", "tcha-protection", "guard"]
}
}
}Clone or copy the .claude/skills folder directly into your project root:
git clone https://github.com/Tchapwo1/tcha-claude-skills.git
cp -r tcha-claude-skills/.claude/skills ./your-project/.claude/Related MCP server: mcp-doorman
Skill | Description | Attack Vector Neutralized |
In-line redaction of API keys, | Accidental credential leakage to external LLM providers or git commits. | |
Boundary filter for Model Context Protocol tools and terminal commands. | Cloud Metadata SSRF ( | |
Pre-execution scanner for scraped text, PDFs, and external web content. | Indirect prompt injections, zero-width steganography, and prompt delimiters. | |
On-demand adversarial penetration test running 100 automated probes. | Configuration oversights and unconfined tool privileges ( | |
Cryptographic Merkle tree audit logger with tamper-evident nonces. | Regulatory liability under EU AI Act Article 15 & SOC 2 CC6.1. |
We evaluated Tcha against 100 adversarial test probes across 4 distinct threat classes:
Threat Class | Probes | Neutralized | Escapes | Mean Overhead |
Cloud Metadata SSRF ( | 25 | 25 / 25 | 0 | 0.38 ms |
Path Traversal & Root Confinement | 25 | 25 / 25 | 0 | 0.41 ms |
Destructive Command Execution | 25 | 25 / 25 | 0 | 0.44 ms |
Indirect Delimiter Injection | 25 | 25 / 25 | 0 | 0.49 ms |
Total Benchmark | 100 | 100 / 100 | 0 | 0.43 ms |
👉 Read the full interactive benchmark report: https://tcha.dev/report.html
┌─────────────────────────────────┐
│ Claude Agent │
│ (Claude Desktop / Claude Code)│
└────────────────┬────────────────┘
│ Tool Invocation (e.g. bash, readFile, fetch)
▼
┌────────────────────────────────────────────────────────┐
│ TCHA DETERMINISTIC TRANSPORT FIREWALL │
│ ✓ Secret Redaction ✓ SSRF & Cloud Metadata Block │
│ ✓ Merkle Audit Nonce ✓ Path Traversal Confinement │
│ Latency: <0.5ms Status: 100% Local / Zero Egress│
└────────────────┬───────────────────────────────────────┘
│ Sanitized & Verified Payload
▼
┌─────────────────────────────────┐
│ Local Tools & MCP Servers │
│ (Terminal, Filesystem, APIs) │
└─────────────────────────────────┘EU AI Act Article 15: Demonstrable technical robustness, cyberattack resilience, and boundary enforcement.
SOC 2 Type II (CC6.1, CC6.6, CC6.8): Logical boundary restriction, data transmission protection, and tamper-evident audit logging.
OWASP Top 10 for LLMs: Directly neutralizes LLM01 (Prompt Injection), LLM02 (Sensitive Information Disclosure), and LLM08 (Excessive Agency).
Official Web Hub: https://tcha.dev
NPM Package: tcha-protection
Twitter / X: @tcha_dev
License: MIT © 2026 Tcha Protection Pro
This server cannot be deployed
Security & DLP proxy for MCP: tool-poisoning scans, PII redaction on tool args/results. Beta.
Real-time MCP & A2A prompt-injection, jailbreak, and tool-schema poisoning firewall for AI agents.
Security firewall for AI agents — scans MCP calls for injection, secrets, and risks.
The WAF for agents. Pattern-based + heuristic firewall scans prompts, RAG documents, tool argume...
