🔍 MAVT — Multi-Agent Visibility Tool

The missing DevTools for multi-agent AI systems.

You wouldn't ship a backend without logs. Why are you shipping agents blind?

Multi-agent systems are the future of AI — but right now, debugging them feels like reading smoke signals. MAVT gives you full observability: every agent call, every decision step, every inter-agent message, visualized in real time.

The problem

You build a multi-agent workflow. Something breaks. You ask yourself:

• Which agent failed — and why?
• What did agent A actually say to agent B?
• Where in the chain did the task go wrong?
• Why is this running so slow?

You open your terminal. You see... nothing useful.

MAVT fixes this.

What you get

| | | |---|---| | 🔁 Agent-to-agent traces | See every message passed between agents, in order | | 🧠 LLM turn inspector | Full prompt/response history with token counts and latency | | 📊 Live topology graph | Visual execution graph, updating in real time | | 🛠 Tool call traces | Every tool invocation, input, output, and duration | | 🧩 Memory panel | Watch agent memory reads and writes as they happen | | 📜 Run history & replay | Browse past runs and replay them step by step | | 🔔 Webhook alerts | Get notified on token budget exceeded, agent stuck, critic failures |

Get started in 60 seconds

npm install -g agent-visibility
agentscope

Open your browser → http://localhost:4242

Your agents are now fully observable.

Framework adapters

Drop-in Python adapters — two lines of code, full visibility.

LangChain

from adapters.langchain import AgentscopeCallback

AgentExecutor(agent=agent, tools=tools,
              callbacks=[AgentscopeCallback(goal="My task")])

AutoGen

from adapters.autogen import track

scope = track(agents=[orchestrator, researcher, coder], goal="My task")
# ... run your agents ...
scope.finish()

CrewAI

from adapters.crewai import AgentscopeListener

listener = AgentscopeListener(goal="My task")
crew = Crew(agents=[...], tasks=[...], step_callback=listener)
result = crew.kickoff()
listener.finish()

Webhook alerts

Configure alerts directly from the dashboard sidebar:

| Alert | Trigger | |---|---| | 💸 Token budget | Agent uses ≥ X% of its token budget | | ⏳ Agent stuck | Agent silent for more than N seconds | | ❌ Critic fail rate | Critic failure rate exceeds X% |

Alerts appear as toast notifications in the UI and are POSTed to any webhook URL you configure.

Why observability is non-negotiable

"If you can't measure it, you can't manage it."

AI agents are making real decisions in production systems today — in customer service, in code generation, in enterprise workflows. Without visibility:

• You can't debug failures
• You can't trust outputs
• You can't scale safely
• You can't explain decisions to stakeholders

MAVT is the foundation layer your agent stack is missing.

Roadmap

• [x] Live topology graph
• [x] LLM turn inspector
• [x] Tool call traces
• [x] Memory panel
• [x] Run history & replay
• [x] LangChain adapter
• [x] AutoGen adapter
• [x] CrewAI adapter
• [x] Webhook alerts
• [ ] OpenTelemetry export
• [ ] Cloud-hosted dashboard
• [ ] Cost tracking per agent

Contributing

Issues, PRs, and framework integrations are very welcome. If you're using MAVT with a framework not listed above — open an issue and let's add it.

Star history

If MAVT saves you a debugging session, consider leaving a ⭐ — it helps other developers find the tool.

About the author

Built by Hitarth Bhatt — AI product leader with 10+ years shipping AI systems at scale. MAVT grew out of a real frustration: the more powerful multi-agent systems become, the harder they are to see inside.

MIT License · npm · Issues