langgraph_quickstart.md

Open Edison + LangGraph Quick Start

See LangGraph docs for the more about LangGraph and associated libraries:

• https://langchain-ai.github.io/langgraph/

1) Install

pip install -U langgraph langchain-openai open-edison

Set environment variables:

• OPENAI_API_KEY = your OpenAI key
• OPEN_EDISON_API_KEY = your Open Edison API key
• OPEN_EDISON_API_BASE = http://localhost:3001 (default)

2) Start the Open Edison server

make run

Access the Open Edison dashboard: http://localhost:3001/dashboard

3) Track your tools and run a langgraph agent

from langchain_core.tools import tool
from langchain_openai import ChatOpenAI
from langgraph.prebuilt import create_react_agent

from open_edison import Edison

# Edison reads OPEN_EDISON_API_BASE / OPEN_EDISON_API_KEY from env
edison = Edison()

@tool
@edison.track() # auto-names: agent_web_search
def web_search(query: str, max_results: int = 3) -> str:
    """Return up to N result URLs (demo)."""
    return "https://docs.python.org/3/"

@tool
@edison.track() # auto-names: agent_fetch_url
def fetch_url(url: str, max_chars: int = 1000) -> str:
    """Fetch and return the first max_chars of the page."""
    import httpx
    return httpx.get(url, follow_redirects=True, timeout=10).text[:max_chars]

llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)
agent = create_react_agent(model=llm, tools=[web_search, fetch_url])

result = agent.invoke({
    "messages": [("user", "Fetch the first 1000 chars of the CPython docs homepage.")]
})
print(result["messages"][-1].content)

Minimal integration diff (4 lines)

diff agent_no_firewall.py agent_firewall.py

examples/langgraph/agent_no_firewall.py                     examples/langgraph/agent_firewall.py                       
                                                                                                                        
                                                            from open_edison import Edison                
                                                            edison = Edison()                                           
                                                                                                                        
@tool                                                       @tool                                                       
                                                            @edison.track()  # auto-names: agent_web_search            
def web_search(query: str, _max_results: int = 3) -> str:   def web_search(query: str, _max_results: int = 3) -> str:  
---                                                         ---                                                        
@tool                                                       @tool                                                       
                                                            @edison.track()  # auto-names: agent_fetch_url             
def fetch_url(url: str, _max_chars: int = 1000) -> str:     def fetch_url(url: str, _max_chars: int = 1000) -> str:    

4) Permissions (server-side)

Tracked tools are named agent_<function_name>. Configure under the agent section in tool_permissions.json:

{
  "agent": {
    "web_search": {
      "enabled": true,
      "write_operation": false,
      "read_private_data": false,
      "read_untrusted_public_data": true,
      "acl": "SECRET"
    },
    "fetch_url": {
      "enabled": true,
      "write_operation": false,
      "read_private_data": true,
      "read_untrusted_public_data": true,
      "acl": "SECRET"
    }
  }
}

Open Edison enforces:

• Lethal trifecta prevention (private data + untrusted exposure + write)
• ACL write-downgrade blocking
• Manual approvals via the dashboard

5) What gets recorded

• Pre-call: /agent/begin (gating + possible approval wait)
• Post-call: /agent/end (status, duration, result preview)
• Arguments/results are captured as JSON previews (capped at 1,000,000 chars).

6) Troubleshooting

• 401 Invalid API key: check OPEN_EDISON_API_KEY and server is running.
• Tool blocked: verify tool_permissions.json and approve in dashboard if needed.