crewai-tools-deepkeep

DeepKeep AI Firewall integration for CrewAI native guardrails and execution hooks.

This package is designed as an external CrewAI integration, not a patch to crewai_tools. It follows CrewAI's native extension points:

• Task.guardrail / Task.guardrails for task output validation and retry feedback.
• Execution hooks for pre/post LLM and tool-call enforcement.
• Optional CrewAI tools for agent-driven moderation workflows.

Installation

pip install crewai-tools-deepkeep

Set credentials:

export DEEPKEEP_API_KEY="dk_..."
export DEEPKEEP_BASE_URL="https://api.example.deepkeep.ai"

Task Guardrail

Use this when you want CrewAI to validate task output and retry the task when DeepKeep blocks it.

from crewai import Agent, Task
from crewai_deepkeep import DeepKeepGuardrail

deepkeep = DeepKeepGuardrail(
    post_model="output-firewall-id",
)

task = Task(
    description="Answer the user safely.",
    expected_output="A safe answer.",
    agent=agent,
    guardrail=deepkeep.check_output,
    guardrail_max_retries=3,
)

DeepKeepGuardrail.check_output returns CrewAI's expected guardrail tuple:

(True, validated_or_modified_output)
(False, "feedback for the agent retry")

Runtime Hooks

Use this when you want DeepKeep to enforce policy around every model and tool boundary.

from crewai_deepkeep import DeepKeepHooks, install_deepkeep_hooks

install_deepkeep_hooks(
    DeepKeepHooks(
        pre_model="input-firewall-id",
        post_model="output-firewall-id",
    )
)

This registers checks for:

• PRE_MODEL_CALL
• POST_MODEL_CALL
• PRE_TOOL_CALL
• POST_TOOL_CALL

Blocking decisions raise CrewAI HookAborted with source="deepkeep". Redact/modify decisions replace the intercepted message, response, tool input, or tool result where CrewAI supports replacement.

Manual Hook Registration

If you prefer explicit registration:

from crewai.hooks import InterceptionPoint, on
from crewai_deepkeep import DeepKeepHooks

hooks = DeepKeepHooks(pre_model="input-firewall-id", post_model="output-firewall-id")

on(InterceptionPoint.PRE_MODEL_CALL)(hooks.check_model_input)
on(InterceptionPoint.POST_MODEL_CALL)(hooks.check_model_output)
on(InterceptionPoint.PRE_TOOL_CALL)(hooks.check_tool_input)
on(InterceptionPoint.POST_TOOL_CALL)(hooks.check_tool_output)

Optional Tools

Tools are available for workflows where the agent should explicitly call DeepKeep:

from crewai_deepkeep.tools import DeepKeepModerateInputTool

agent = Agent(
    role="AI Security Guard",
    tools=[DeepKeepModerateInputTool(model="input-firewall-id")],
)

For production enforcement, prefer task guardrails and hooks over agent-callable tools.

Recommended CrewAI Docs Positioning

If submitting this to CrewAI upstream, position it as an external integration page, not as a request to add DeepKeep classes into CrewAI core:

DeepKeep integrates with CrewAI's native task guardrails and execution hooks. It is distributed as an external package, so CrewAI core does not need to maintain vendor-specific API code.

Environment Variables

• DEEPKEEP_API_KEY: DeepKeep API key.
• DEEPKEEP_BASE_URL: DeepKeep API base URL, without a trailing slash.

Repository Placement

Recommended upstream repository:

https://github.com/Deepkeepai/crewai-tools-deepkeep