Integrations CrewAI

CrewAI

Experimental: this works and is tested, but may change and gets less hardening than the stable core.

tracekit.integrations.crewai gates every tool call a CrewAI crew makes, through CrewAI's global tool hooks: before each call it asks the Tracekit signer for a decision (an ask waits for a person), and after the call it records the outcome.

Install

pip install tracekit-ai "crewai>=1.15.27,<1.16"

Tested in CI against a stand-in for CrewAI's tool hooks, not the real library. The case study ran CrewAI 1.15.27 with hand-written hook glue, not this module.

Setup

import tracekit; tracekit.instrument() before the crew runs registers both hooks, and records the model calls CrewAI makes through the OpenAI, Anthropic and Google Gen AI SDKs (autotrace). By hand:

from crewai.hooks import register_after_tool_call_hook, register_before_tool_call_hook
from tracekit.integrations.crewai import hooks
from tracekit.sdk.client import Client

run = Client().run(agent="support-crew")
before, after = hooks(run)
register_before_tool_call_hook(before)
register_after_tool_call_hook(after)

What it does

• Blocks by return value. CrewAI runs a tool when a before hook raises, so the before hook never raises: it returns False, which blocks the call, on anything but an allow or an approved ask. That covers a deny, a rejected or unanswered approval, a refused consume, a signer refusal, an outage under a closed fail mode and any error of its own. Under an open fail mode, a call runs unrecorded during an outage, with a warning.
• Tells the model why. For a blocked call the after hook replaces CrewAI's "Tool execution blocked by hook" with Tracekit's reason (the rule ids, or what the approval came to).
• Records failures as failures. CrewAI hands the after hook a failed tool's exception as the text Error executing tool: ...; that is recorded as status: error. Other results are recorded as the text the model reads, and in full for a tool the policy lists as untrusted (provenance).
• Waits for people. An ask waits up to 5 minutes (the signer's cap for one wait) before the call; then it is blocked with the approval's state.

What the model said

Not recorded by the integration: CrewAI's after-LLM hook sees a text answer only, never a response that asks for a tool, and no reasoning; its tool hooks carry no call id. Record the model calls with tracekit.instrument() or the LLM gateway; they pair with the calls by arguments only. Which frameworks and providers return readable reasoning, and how to turn it on: reasoning.

Not covered

• CrewAI's hooks carry no tool-call id, so each call gets one of its own (crewai-<hex>). Under autotrace or the LLM gateway the reconciler pairs it with the model's tool use of the same turn by its arguments (paired_by_args on run.final). Two calls of the same tool with the same arguments in one turn can't be told apart: both are reported (reconcile.hook_missing and reconcile.fabricated). So are calls whose decide lands after another agent's model call in the same run (async tasks, or several crews in instrument()'s one process run): give each its own run.
• On CrewAI's text (ReAct) path, for models without native tool calling, CrewAI retries a failing tool inside one call of its hooks: the retries run under that call's one decision, and only the last outcome is recorded. A failure there ("I encountered an error while trying to use the tool") is recorded as an error, as on the native path.
• Delegation between agents (delegate_work_to_coworker, ask_question_to_coworker) is decided like any tool: give those tools a class in your policy to gate them.
• A tool that runs other than through CrewAI's tool execution (a tool calling another tool directly in Python) is not seen.