Framework Adapter Comparison

Edictum ships eight framework adapters. This guide helps you choose the right one and understand the tradeoffs.

Quick Comparison

Cost column reflects benchmarks from edictum-demo using each framework's default model. N/A indicates no published benchmark data.

Which Adapter Should I Use?

• Need full PII interception? -- Use LangChain, Agno, Semantic Kernel, Nanobot, or Google ADK. These adapters can replace the tool result before the LLM sees it.
• Cheapest per-task cost? -- Semantic Kernel ($0.008 per task in benchmarks).
• Simplest integration? -- Claude SDK or Agno. Both require minimal wiring.
• Need ADK plugin/callback enforcement? -- Use Google ADK adapter.
• Using CrewAI? -- CrewAI adapter is the only option. Note that CrewAI hooks are global (applied to every tool across all agents in the crew).

Per-Adapter Snippets

LangChain

from edictum import Edictum, Principal
from edictum.adapters.langchain import LangChainAdapter
from langgraph.prebuilt import ToolNode

guard = Edictum.from_yaml("rules.yaml")
adapter = LangChainAdapter(guard=guard, principal=Principal(role="analyst"))
wrapper = adapter.as_tool_wrapper()
# Pass to: ToolNode(tools=tools, wrap_tool_call=wrapper)

OpenAI Agents

from edictum import Edictum, Principal
from edictum.adapters.openai_agents import OpenAIAgentsAdapter
from agents import function_tool

guard = Edictum.from_yaml("rules.yaml")
adapter = OpenAIAgentsAdapter(guard=guard, principal=Principal(role="assistant"))
input_gr, output_gr = adapter.as_guardrails()
# Pass to: @function_tool(tool_input_guardrails=[input_gr], tool_output_guardrails=[output_gr])

CrewAI

from edictum import Edictum, Principal
from edictum.adapters.crewai import CrewAIAdapter

guard = Edictum.from_yaml("rules.yaml")
adapter = CrewAIAdapter(guard=guard, principal=Principal(role="researcher"))
adapter.register()
# Hooks are now globally registered for all CrewAI tool calls

Agno

from edictum import Edictum, Principal
from edictum.adapters.agno import AgnoAdapter

guard = Edictum.from_yaml("rules.yaml")
adapter = AgnoAdapter(guard=guard, principal=Principal(role="assistant"))
hook = adapter.as_tool_hook()
# Pass to: Agent(tool_hooks=[hook])

Semantic Kernel

from edictum import Edictum, Principal
from edictum.adapters.semantic_kernel import SemanticKernelAdapter

guard = Edictum.from_yaml("rules.yaml")
adapter = SemanticKernelAdapter(guard=guard, principal=Principal(role="analyst"))
adapter.register(kernel)
# Filter is now registered on the kernel instance

Google ADK

from edictum import Edictum, Principal
from edictum.adapters.google_adk import GoogleADKAdapter

guard = Edictum.from_yaml("rules.yaml")
adapter = GoogleADKAdapter(guard=guard, principal=Principal(role="analyst"))
plugin = adapter.as_plugin()
# Pass to your ADK Runner: plugins=[plugin]

Claude SDK

from edictum import Edictum, Principal
from edictum.adapters.claude_agent_sdk import ClaudeAgentSDKAdapter

guard = Edictum.from_yaml("rules.yaml")
adapter = ClaudeAgentSDKAdapter(guard=guard, principal=Principal(role="sre"))
hooks = adapter.to_hook_callables()
# Use in your agent loop — see bridge recipe in Claude SDK adapter docs

Nanobot

from edictum import Edictum, Principal
from edictum.adapters.nanobot import NanobotAdapter

guard = Edictum.from_yaml("rules.yaml")
adapter = NanobotAdapter(guard=guard, principal=Principal(role="user"))
governed_registry = adapter.wrap_registry(agent.tool_registry)
# Replace: agent.tool_registry = governed_registry

Known Limitations

LangChain

• as_middleware() is sync-only. If an asyncio event loop is already running (Jupyter, FastAPI), use as_tool_wrapper() (handles nested loops via ThreadPoolExecutor) or as_async_tool_wrapper(). See LangChain adapter docs.

OpenAI Agents

• Guardrails are per-tool via @function_tool(), not per-agent. Input and output guardrails are separate functions; the adapter correlates them using insertion-order (FIFO), which assumes sequential tool execution. See OpenAI Agents adapter docs.

CrewAI

• Hooks are global -- they apply to every tool across all agents in the crew. There is no per-agent hook scoping. See CrewAI adapter docs.
• Side-effect only -- on_postcondition_warn callbacks fire for side effects (logging, alerting) but cannot replace the tool result. Postcondition redact/block effects set PostCallResult.result for wrapper consumers but cannot modify what CrewAI passes to the model.

Agno

• Tool callables must accept keyword arguments (the adapter spreads the args dict with **arguments). See Agno adapter docs.

Google ADK

• Use plugin path for runner-wide enforcement and callback path for live/streaming mode. See Google ADK adapter docs.

Semantic Kernel

• By default, context.terminate = True on block stops all auto-invocations in the current turn, not just the blocked tool. Set terminate_on_deny=False to allow remaining tool calls to proceed. See Semantic Kernel adapter docs.

Claude SDK

• Side-effect only -- the hook callables (to_hook_callables()) cannot replace the tool result. PII detection is logged but not intercepted before reaching the model. Postcondition redact/block effects set PostCallResult.result for wrapper consumers but cannot modify the SDK's result flow. A warning is logged at adapter construction when postconditions declare these effects. See Claude SDK adapter docs.

Nanobot

• GovernedToolRegistry.execute() always returns a string (matching nanobot's ToolRegistry rule). Non-string results from the inner registry are converted via str(). See Nanobot adapter docs.

OpenAI Agents (postcondition enforcement)

• The output guardrail can .allow() or .reject_content() but cannot substitute the tool result. With one pending tool call, postcondition action: block on pure/read tools returns .reject_content(). When calls overlap, the native output guardrail skips postcondition evaluation because it cannot identify which tool produced the output. Use the wrapper path with explicit call IDs for guaranteed postcondition enforcement under concurrency. Postcondition action: redact is not supported because native guardrails cannot transform tool results.