sbo3l-crewai-keeperhub 1.2.0

CrewAI Python tool that gates KeeperHub workflow execution through SBO3L's policy boundary. Composable with sbo3l-sdk.

sbo3l-crewai-keeperhub

CrewAI Python tool that gates KeeperHub workflow execution through SBO3L's policy boundary. Composable with sbo3l-sdk.

Why this exists alongside langchain-keeperhub (Devendra's PyPI pkg)

	langchain-keeperhub (Devendra)	sbo3l-crewai-keeperhub (this)
Framework	LangChain	CrewAI
What it wraps	KH webhook execution	SBO3L policy gate → KH webhook execution
Decision step	(agent decides)	(SBO3L decides; signed receipt)
Budget enforcement	no	yes
Audit chain	no	yes (hash-chained Ed25519 log)
Multi-agent crew composition	n/a	one policy boundary across all crew agents
ENS / Turnkey TEE / MCP bridge	yes	no (not duplicated)

Composable: use Devendra's tool for the raw KH binding + ours as the policy gate that decides whether the raw call should fire. Or use ours alone for the full gate-then-execute path. The CrewAI angle adds: one SBO3L policy boundary shared across N specialized crew agents, with a single hash-chained audit log spanning every agent's KH executions.

Install

pip install "sbo3l-crewai-keeperhub[crewai]" sbo3l-sdk

The [crewai] extra pulls in crewai so the typed Sbo3lKeeperHubCrewAITool (a crewai.tools.BaseTool subclass) is importable. Without it, only the framework-agnostic sbo3l_crewai_keeperhub_tool() factory is available.

5-line setup

from sbo3l_sdk import SBO3LClientSync
from sbo3l_crewai_keeperhub import Sbo3lKeeperHubCrewAITool

client = SBO3LClientSync("http://localhost:8730")
tool = Sbo3lKeeperHubCrewAITool(client=client)
# pass `tool` into your CrewAI Agent's tools=[...] list

Or framework-agnostic:

from sbo3l_crewai_keeperhub import sbo3l_crewai_keeperhub_tool
descriptor = sbo3l_crewai_keeperhub_tool(client=client)
# descriptor.name + descriptor.description + descriptor.func

Wire path

Same as the sibling LangChain / TS packages — SBO3L decides, KH adapter executes on allow, tool returns kh_execution_ref. See the LangChain README's "Wire path" section for the full sequence. All packages share the daemon as the policy boundary, so kh_execution_ref matches across SDK languages.

Multi-agent crew composition

CrewAI's strength is N specialized agents collaborating on a shared task. Wire one Sbo3lKeeperHubCrewAITool instance into every agent's tool list and you get:

• One policy boundary for the whole crew
• One KH workflow_id (advisory) tagged across every execution
• One hash-chained audit log spanning researcher → executor → auditor agents

See examples/crewai-keeperhub-demo/agent.py for a runnable 3-agent crew using this pattern.

On kh_workflow_id_advisory

The _advisory suffix is intentional: today the daemon's env-configured webhook URL is the source of truth for actual routing. The per-call workflow_id you pass is surfaced in the envelope for context tagging / audit logs, not as a routing override. See KeeperHub/cli#52 for the proposed contract that would make per-call routing safe.

API

sbo3l_crewai_keeperhub_tool(
    *,
    client: SBO3LClientLike,
    workflow_id: str | None = None,           # default: DEFAULT_KH_WORKFLOW_ID
    name: str = "sbo3l_keeperhub_payment_request",
    description: str = ...,
    idempotency_key: Callable[[dict], str] | None = None,
) -> SBO3LKeeperHubToolDescriptor

# Or, with crewai installed:
Sbo3lKeeperHubCrewAITool(
    client=...,                # required, kw-only
    workflow_id=None,          # default: DEFAULT_KH_WORKFLOW_ID
    name=None, description=None,
)
# Real crewai.tools.BaseTool subclass — drops into Agent(tools=[...]).

License

MIT