The Ultimate Observability, Security, and Analytics Platform for AI Agents
OpenTelemetry + Wireshark + Cloudflare for AI Agents, LLMs, and MCP Servers
AgentWire is a cutting-edge observability and security gateway specifically designed for modern AI Agents, Large Language Models (LLMs), and Model Context Protocol (MCP) servers. Just as Wireshark inspects network packets and Cloudflare provides a protective edge layer, AgentWire sits between your users, agents, and external services to monitor, inspect, analyze, secure, replay, and optimize every single AI interaction.
Whether you're dealing with prompt injection attacks, skyrocketing LLM API costs, or debugging complex multi-agent workflows, AgentWire provides the real-time visibility and control you need to scale your AI infrastructure with confidence.
- 🚦 Live Traffic Monitor: Real-time dashboard similar to a network operations center.
- 🕵️♂️ AI Packet Inspector: Deep inspection of every prompt, context, memory, tool request, and model output.
- ⏪ Replay Engine: Replay any execution with different prompts, models, or temperatures to debug or optimize.
- 📚 Prompt Version Control: Track history, diffs, and success rates of all prompts.
- 🛡️ Security & Guardrails: Detect prompt injections, sensitive data leakage (PII/PHI), and malicious MCP server behavior.
- 💸 Cost Intelligence: Detailed cost analytics broken down by organization, project, model, and user.
- 🔍 Advanced Search: Global search to find specific executions, errors, semantic matches, or security events.
- 🔌 Universal Plugin System: Seamless integration with OpenAI, Anthropic, Gemini, local models, and MCP Servers.
Company X deployed a powerful autonomous customer support agent. Everything worked perfectly in testing, but in production, they faced three major crises:
- Unpredictable Costs: API bills skyrocketed without clear attribution. Was it user spam, inefficient loops, or a runaway agent?
- Security Vulnerabilities: A user successfully executed a prompt injection attack, tricking the agent into leaking internal system instructions.
- Debugging Nightmares: When an agent gave a hallucinatory response, developers had no way to trace the exact context, tool calls, and LLM state at that specific millisecond.
By routing their agent traffic through AgentWire, Company X transformed their operations:
- Instant Cost Attribution: AgentWire's Cost Intelligence pinpointed a specific looping tool call that was burning tokens, saving them 40% on API costs overnight.
- Proactive Security: The Security Scanner automatically intercepted and blocked prompt injections before they even reached the LLM, keeping their system instructions secure.
- One-Click Replays: Developers used the Replay Engine to reproduce the exact state of the hallucination, tweaked the system prompt, and tested the fix instantly against historical traffic.
Result: A secure, cost-effective, and highly predictable AI deployment.
AgentWire is built on a scalable, cloud-native architecture capable of handling massive throughput:
- Backend: ASP.NET Core 10, .NET Aspire
- Database: PostgreSQL (Metadata) + ClickHouse (High-volume packet storage) + Redis (Caching)
- Frontend: Next.js, React, Tailwind CSS, React Flow, Apache ECharts
- Infrastructure: Docker, Kubernetes, Helm, Terraform
See docs/architecture.md for more details.
(Documentation coming soon)
AI Agents LLM Observability AI Security Prompt Injection MCP Servers Model Context Protocol AI Gateway LLM Analytics AI Traffic Monitor Generative AI Agentic AI AI Infrastructure Machine Learning LLMOps AIOps OpenTelemetry for AI Wireshark for AI Cloudflare for AI
This project is licensed under the Apache 2.0 License - see the LICENSE file for details.