Framework Integrations

Cordum sits between your agent framework and the tools/data the agent acts on. For each supported framework, the integration is a thin adapter that:

1. Wraps the framework's tool-invocation surface so every tool call routes through the Cordum MCP bridge.
2. Lets the Safety Kernel evaluate the call before it executes.
3. Surfaces approvals and denials back to the agent as tool-call results (so the loop continues cleanly).

Supported frameworks

Framework	Adapter	Status	Tutorial
LangChain	cordum-adapters[langchain]	✅ Shipped	LangChain guard
LangGraph	cordum-adapters[langchain] + graph wiring	✅ Shipped	LangGraph 5-min
CrewAI	cordum-adapters[crewai]	✅ Shipped	CrewAI safety gates
AutoGen / AG2	cordum-adapters[autogen]	🚧 In progress	Coming soon — see the epic
OpenAI Agents SDK	cordum-adapters[openai-agents]	🚧 In progress	Coming soon
LlamaIndex	cordum-adapters[llama-index]	📋 Planned	—
Temporal	Cordum pack	📋 Planned	—

Install the adapters

pip install cordum-adapters[<framework>]

Available extras: langchain, crewai, autogen, autogen-classic (pyautogen 0.2), openai-agents, all, dev.

Common wiring

Every adapter needs three things:

from cordum_agent_adapters.mcp_client import McpStdioClient

client = McpStdioClient(
    command=["cordum-mcp-bridge"],
    env={
        "CORDUM_GATEWAY_URL": "https://localhost:8081",
        "CORDUM_API_KEY": "<your-key>",
    },
)

Then hand client to the framework-specific builder:

• LangChain: build_langchain_tools(client) → list of BaseTool.
• CrewAI: build_crewai_tools(client) → list of BaseTool subclasses.
• AutoGen (classic): build_autogen_tools(client) → (functions, function_map).
• AutoGen (AG2 0.4+): build_ag2_tools(client) → list of FunctionTool.
• OpenAI Agents: build_openai_agent_tools(client) → list of FunctionTool.

What governance looks like in the loop

When Cordum denies a tool call (policy violation, scope filter, rate limit), the adapter translates the deny into a framework-native error:

• LangChain / CrewAI / AutoGen: the tool returns a string prefixed [POLICY DENIED] … — the LLM sees the deny in the next turn and can try a different approach.
• OpenAI Agents: same string format, returned from on_invoke_tool.

When Cordum requires a human approval, the adapter:

1. Returns an approval pending indicator to the LLM.
2. Surfaces the pending approval in the Cordum dashboard.
3. The LLM retries after the human resolves (or on a configurable timeout).

Conversation audit

Every framework adapter integrates with the Cordum audit chain via CordumConversationLogger. Every tool call, every turn's metadata, every approval resolution lands in the tenant's SIEM event stream with hash-chained integrity.

from cordum_agent_adapters.audit import CordumConversationLogger

logger = CordumConversationLogger(client=client)
# pass `logger=logger` to any adapter builder

Tutorials

• LangChain guard — full end-to-end LangChain agent with Cordum governance.
• LangGraph 5-min — drop Cordum into a LangGraph state machine in five minutes.
• CrewAI safety gates — CrewAI crew running governed tool calls.
• AutoGen multi-agent — multi-agent conversation with governance at every handoff.

See also

• Agent Protocol (CAP) — the wire protocol the adapters speak.
• Safety Kernel — what the adapters ultimately call into.
• cap — SDKs (Go, Python, Node, C++) if you want to build a framework adapter yourself.