gatecore-crewai

gatecore-crewai provides three CrewAI tools for GateCore marketplace discovery and procurement, plus helpers that execute a procured capability through GateCore's signed gateway. The adapter contains transport and configuration only; GateCore services enforce identity, policy, trust, pricing, funding, delegation, and settlement.

The MCP tools support public full-mode GateCore endpoints only. read mode passes the server's fail-closed procurement error through unchanged. off mode is unsupported because it does not replace the adapter's legacy placeholder identity with the MCP credential identity.

This package intentionally duplicates the small adapter configuration, execution, and transport modules. A shared adapter-core package is a future decision that waits until a third framework adapter establishes the common surface.

Install from the GateCore monorepo

python -m pip install -e ./libs/sdk -e ./libs/adapters/gatecore-crewai

Install from PyPI (effective on publish)

Effective once gatecore-sdk and this package have a real PyPI release — see .github/workflows/publish-pypi.yml. Until then, use the monorepo install above.

python -m pip install gatecore-crewai

Configuration

Discovery and procurement require:

GATECORE_MCP_URL=https://mcp.example.com/mcp
GATECORE_MCP_KEY=gcmk_...

execute_procured and aexecute_procured validate these variables lazily on first use:

GATECORE_GATEWAY_URL=https://api.example.com
GATECORE_BEARER_TOKEN=...
GATECORE_AGENT_ID=agent:example
GATECORE_ACCESS_GRANT_ID=grant:example
GATECORE_KEY_ID=key:example
GATECORE_PRIVATE_KEY_PATH=/secure/path/agent-key.pem

Set exactly one of GATECORE_PRIVATE_KEY_PATH or GATECORE_PRIVATE_KEY_PEM.

Usage

from gatecore_crewai import GateCoreToolkit, execute_procured

tools = GateCoreToolkit().get_tools()
procure_result = tools[2].run(listing_id="listing:example")
decision = execute_procured(procure_result, {"query": "example"})

CrewAI tools are sync-native. Their synchronous implementation uses asyncio.run() for the MCP transport and therefore must not be called from a thread that already owns a running event loop; use arun() in async applications.