Click on "Install 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., "@AgentGuard MCP Server Intercept tool calls, redact secrets, and require approval for sensitive actions."
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.
Runtime security for AI agents. AgentGuard intercepts tool calls at an MCP boundary, validates provenance and arguments, applies deterministic policy precedence, redacts secrets, pauses sensitive actions for human approval, and writes a tamper-evident audit trail.
The application boots with a complete deterministic attack demo and zero credentials. Live adapters activate only when their hosted-service environment variables are present. AgentGuard never downloads a model and has no Ollama or local-vLLM fallback.
This is not a chatbot wrapper. It demonstrates agent security, protocol design, streaming UX, policy engineering, human-in-the-loop execution, multi-runtime architecture, observability, retrieval, evaluation, and hosted deployment.
Four attack fixtures prove four distinct controls:
Fixture | Expected enforcement | Security control |
Indirect prompt injection |
| Untrusted provenance + destructive SQL |
Secret exfiltration |
| Recursive credential removal before egress |
Production rollback |
| Suspended execution + operator decision |
Repository intelligence |
| Explicit read-only tool authorization |
Technology | Implemented responsibility |
Vercel AI SDK | Typed UI event streaming and hosted model calls |
OpenAI Responses API | Explicit |
OpenAI Agents SDK | Primary hosted agent runtime connected to protected MCP tools |
Model Context Protocol | Streamable HTTP gateway with six guarded tools |
LangGraph | Stateful retrieve → assess → approve/finalize workflow |
Mastra | TypeScript-native agent through LiteLLM |
PydanticAI | Typed |
LiteLLM | Hosted provider routing, retry, and budget gateway |
Hosted vLLM | Remote OpenAI-compatible serving adapter; never local |
PostgreSQL + pgvector | Security-run persistence, checkpoints, semantic audit history |
Qdrant | Hosted policy and attack-corpus vector retrieval |
LlamaIndex | Ingestion and retrieval over the Qdrant collection |
DSPy | Evaluation-driven policy-classifier optimization |
Arize Phoenix | Hosted OpenTelemetry traces and evaluation visibility |
Related MCP server: SINT Protocol
pnpm install
pnpm devOpen http://localhost:3000. The first attack runs automatically. No .env.local is necessary.
pnpm typecheck
pnpm test
pnpm buildsrc/app/ Next.js UI and API routes
src/components/ Recruiter-facing attack console
src/lib/guard/ Policy engine, schemas, fixtures, audit chain
src/lib/mcp/ MCP server and protected real/simulated executors
src/lib/runtimes/ OpenAI Agents, Mastra, and Python-service adapters
src/lib/ai/ LiteLLM and remote-vLLM model providers
services/intelligence/ Hosted FastAPI intelligence service
infrastructure/postgres/ pgvector schema
infrastructure/litellm/ Hosted LiteLLM routing config
tests/ Policy and browser-level verification
docs/ Architecture, threat model, demo, credentialsFor the live architecture, see Architecture. For the exact optional keys and their destinations, see Credentials. For interview delivery, use the 90-second demo script.
This server cannot be installed
Unclaimed servers have limited discoverability.
If you are the server author, to access and configure the admin panel.
Runtime permission, approval, and audit layer for AI agent tool execution.
Security firewall for AI agents — scans MCP calls for injection, secrets, and risks.
See, price, and control every tool call your AI agents make: policy checks, cost, and audit tools.
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/vaasu202/agentguard'
If you have feedback or need assistance with the MCP directory API, please join our Discord server
