Quick start

pip install "inter-agent-guard[all,otel]"
python scripts/download_release_model.py   # ~164 MB INT8 ONNX from GitHub Releases
agentguard status

from agentguard import AgentGuard, CapabilityManifest

guard = AgentGuard(
    risk_threshold=0.85,
    task_objective="Analyse Q3 competitor pricing",
    audit_log_path="./audit.jsonl",
    require_ml_model=True,  # after download_release_model.py
)
guard.register_agent(
    "research-agent",
    CapabilityManifest.from_yaml("manifests/research_agent.yaml"),
)
secured = guard.wrap(my_langgraph_graph)

Without the ONNX model, rule filtering, trust attestation, and capability enforcement still run. Set require_ml_model=True only after the model is installed.

Inter-agent messages (trust attestation)

Trust attestation is on by default. Inter-agent payloads must be signed with a recipient-bound envelope before inspect_message will forward them:

payload = b"Research summary ready for internal report."
text = payload.decode()
sig = guard.sign_payload("researcher", payload, recipient_id="writer")
decision = guard.inspect_message(
    "researcher",
    "writer",
    text,
    payload,
    signature=sig,
)

sign_payload requires recipient_id matching the recipient passed to inspect_message. Unsigned messages are BLOCKed; legacy raw 64-byte signatures are rejected.

For unsigned boundaries (user input, framework hooks), use inspect_content(sender_id, recipient_id, message) instead — it runs the rule filter, ML scorer, and consistency check without trust verification.

LangChain agents

For LangChain 1.0 create_agent, use the official middleware integration:

from langchain.agents import create_agent
from agentguard import AgentGuard
from agentguard.adapters.langchain import AgentGuardMiddleware

guard = AgentGuard(task_objective="Analyse Q3 competitor pricing")
agent = create_agent(
    model="gpt-5.5",
    tools=[...],
    middleware=[AgentGuardMiddleware(guard, agent_id="researcher")],
)

User input is scanned before the first model call, every tool output is inspected for indirect injection before the model sees it, and capability manifests registered under agent_id are enforced before tools execute. Requires pip install "inter-agent-guard[langchain]".