CrewAI

5. CrewAI

Time: 5 minutes. raucle gates every tool call your crew makes against a signed capability token and records each ALLOW/DENY as a verifiable receipt — without changing how your tools look to the agent.

pip install 'raucle[compliance,crewai]'

Wrap your tools

guard_tools returns drop-in BaseTools with the same name / description / args_schema, so the agent is unchanged:

from raucle.audit import Ed25519Signer, HashChainSink
from raucle.capability import CapabilityGate, CapabilityIssuer
from raucle.integrations.crewai import guard_tools, set_in_force_token

issuer = CapabilityIssuer.generate(issuer="acme.platform")
gate   = CapabilityGate(trusted_issuers={issuer.key_id: issuer.public_key_pem})
sink   = HashChainSink("receipts.jsonl", signer=Ed25519Signer.generate())

tools = guard_tools(my_raw_tools, gate=gate, sink=sink)

# Per session: bind a token scoped to what THIS task may do.
set_in_force_token(issuer.mint(
    agent_id="agent:billing", tool="transfer_funds",
    constraints={"max_value": {"amount": 100}}, ttl_seconds=60,
))

agent = Agent(role="biller", goal="...", tools=tools, ...)
Crew(agents=[agent], tasks=[...]).kickoff()

A call outside the token’s signed constraints raises CapabilityDenied before the tool body runs — CrewAI surfaces it to the agent, and the denial is recorded in the signed chain. Because the gate evaluates arguments, not the agent’s reasoning, a prompt-injected call is denied on the same path as any other.

Verify the evidence offline:

raucle audit verify receipts.jsonl
raucle watch receipts.jsonl --denies-only   # live SOC view

See the runnable examples/crewai_demo (no API key) for injection-blocked → receipt-verified end to end.

---

Next: SIEM export & live monitoring to stream these decisions to your SOC.