Framework adapters

Three adapters share one seam — a GatePort (local or remote) plus a GateSession that accumulates evidence as tools run. Each gates the sensitive tool in the agent's own call path, so BLOCK/REVIEW raise ClearanceDenied before it executes; all three work against an in-process Gate or the remote client unchanged.

LangChain

A callback handler that collects evidence tools and gates the sensitive one:

from evidence_gate import Gate, ManifestBuilder, PolicySet
from evidence_gate.integrations.langchain import EvidenceGateCallbackHandler, LocalGatePort

port = LocalGatePort(Gate(PolicySet.from_dir("policies")), builder)   # or RemoteGatePort(RemoteGate(...))
handler = EvidenceGateCallbackHandler(port, action_mapping={"send_*": "marketing.send_sequence"})
agent.invoke(..., config={"callbacks": [handler]})   # BLOCK/REVIEW raise before the tool runs

CrewAI / LlamaIndex

Both frameworks' event buses are observe-only (they can't stop a call), so these adapters wrap the tool itself and share a GateSession:

from evidence_gate.integrations.base import GateSession, LocalGatePort
from evidence_gate.integrations.crewai import gate_crew_tools        # or llamaindex.gate_llama_tools

session = GateSession(LocalGatePort(gate, builder),
                      action_mapping={"send_*": "marketing.send_sequence"}, now=now)
tools = gate_crew_tools([get_optin, send_marketing], session)        # gated, drop-in replacements

See examples/remote_agent.py, examples/trace_replay.py, and examples/langchain_agent.py for each path end-to-end.