⚠️ Alpha: Core SDK and self-hosting work. Human approval workflows and full dashboard are in active development. See the roadmap.

---

Why AgentWatch?

Organizations deploy AI agents in production every day — customer support, code generation, document processing, and more. But once deployed, nobody can answer: What did the agent do? How much did it cost? Did it get approval before taking action?

AgentWatch is the missing observability layer. Unlike closed-source alternatives:

• 🏠 Fully self-hostable — the entire stack (SDK + API + dashboard) runs on your infra. Not just the SDK — everything.
• 📡 OpenTelemetry-native — no proprietary event format. Your data stays portable.
• ⚖️ MIT licensed — no AGPL copyleft triggers. Enterprise-friendly from day one.
• 🛡️ Human approval workflows — built into the core, not gated behind an Enterprise tier.
• 🔌 Framework-agnostic — one SDK works with OpenAI, Anthropic, LangChain, CrewAI, AutoGen, and LlamaIndex.

---

⚡ Quickstart

pip install agentwatch

OpenAI

import agentwatch
import openai

agentwatch.init(api_key="aw_your_key")  # That's it. All OpenAI calls are now traced.

client = openai.OpenAI()
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello!"}]
)

LangChain

import agentwatch
agentwatch.init(api_key="aw_your_key")

from langchain_openai import ChatOpenAI
llm = ChatOpenAI(model="gpt-4o")
result = llm.invoke("Explain quantum computing")

CrewAI

import agentwatch
agentwatch.init(api_key="aw_your_key")

from crewai import Crew, Agent, Task
crew = Crew(agents=[...], tasks=[...])
result = crew.kickoff()  # Automatically traced

Custom Instrumentation

@agentwatch.session
def my_agent_pipeline():
    result = research()
    return summarize(result)

@agentwatch.tool
def research():
    return search_web("latest AI news")

---

🏠 Self-Hosting

Deploy the entire AgentWatch stack in under 5 minutes:

git clone https://github.com/mukktinaadh/agentwatch
cd agentwatch
cp .env.example .env    # Edit with your secrets
docker compose up -d    # That's it.

Dashboard: http://localhost · API: http://localhost:8000/docs

Point your SDK to your self-hosted instance:

agentwatch.init(
    api_key="aw_your_key",
    endpoint_url="http://your-server:8000"
)

---

📊 Dashboard

AgentWatch ships with a production-ready dashboard:

• Sessions View — All agent runs with status, cost, duration, error count
• Trace Waterfall — OTel-style span visualization with timing bars
• Cost Analytics — Spend by day, model, and agent with trend charts
• Alerts — Configure cost overrun and failure threshold alerts
• Approvals — Human-in-the-loop approval workflows with real-time polling
• Settings — API key management and self-hosting configuration

---

🏗️ Architecture

┌─────────────────┐     OTel Spans      ┌──────────────────┐
│   Your Agent    │ ──────────────────▶ │  AgentWatch API  │
│  + SDK (pip)    │                     │   (FastAPI)      │
└─────────────────┘                     └────────┬─────────┘
                                                 │
                                    ┌────────────┼────────────┐
                                    ▼            ▼            ▼
                              ┌──────────┐ ┌──────────┐ ┌──────────┐
                              │ PostgreSQL│ │  Redis   │ │ Dashboard│
                              │  (data)  │ │ (buffer) │ │  (React) │
                              └──────────┘ └──────────┘ └──────────┘

Approval Flow:
  Agent ──▶ SDK blocks ──▶ API ──▶ Dashboard (human review) ──▶ Agent resumes

---

🆚 Comparison

Feature	AgentWatch	AgentOps.ai	Langfuse	LangSmith
Fully self-hostable (entire stack)	✅	❌	✅	❌
Open source (MIT)	✅	⚠️ SDK only	✅	❌
OpenTelemetry-native	✅	❌	⚠️	❌
Human approval workflows	✅	❌	❌	❌
Cost analytics	✅	✅	⚠️	⚠️
Framework-agnostic	✅	✅	✅	❌
Compliance in open tier	✅	❌	✅	❌

---

🗺️ Roadmap

See our GitHub Projects board for the full roadmap.

Phase 1 (MVP): ✅ OTel traces · ✅ Cost analytics · ✅ Self-hosting · ✅ Dashboard Phase 2: ✅ Human approval workflows · Cross-run analysis · Alert webhooks (Slack) Phase 3: Compliance reporting · Memory management · Enterprise SSO/RBAC · AgentWatch Cloud

---

🤝 Contributing

We welcome contributions! See CONTRIBUTING.md for details.

# Local development
make dev                    # Start Postgres + Redis
cd api && uvicorn agentwatch_api.main:app --reload
cd dashboard && npm run dev

---

📄 License

MIT — see LICENSE. Use it however you want. No AGPL, no copyleft triggers.

---