AI Agent Governance · Averon product
AgentShield Know exactly what your AI agents did, and prove it.
AgentShield instruments AI agents and turns every tool call, model call, success and failure into a tamper-evident audit record, with risk classification, anomaly alerts, a human-approval trail and signed evidence packages.
- Status
- Python SDK on PyPI
- Surfaces
- Python SDK · Web dashboard · Self-hosted API
- Owner
- Averon Technologies
from agentshield_control_plane import AgentShield
shield = AgentShield(api_key="ags_live_xxxx",
endpoint="http://localhost:8000")
with shield.session(agent="my-agent") as session:
with session.tool_call("search", query="weather in SF"):
... # your real tool-call code
with session.llm_call("gpt-5", tokens=120):
... # your real LLM call- SDK
- Ingestion
- Risk classification
- Anomaly detection
- Hash-chained ledger
- Signed evidence
The problem
Why AgentShield exists
AI agents increasingly read customer data, trigger financial operations and change production systems. When something goes wrong, or an auditor, customer or investor asks, most teams cannot reconstruct what an agent actually did, let alone prove the record has not been altered. Application logs were never designed to be evidence.
Built for
Engineering and platform teams running AI agents in production, particularly those already being asked how agent behaviour is overseen and audited.
What it does
Key capabilities
Instrument any agent
A Python SDK with three manual calls, plus adapters for LangChain, CrewAI and the OpenAI Agents SDK that hook each framework's own extension points instead of patching internals.
Tamper-evident audit ledger
Every action is stored as a hash-chained event. Editing a row, even directly in the database and bypassing the API, breaks the chain and is detectable on the next read.
Risk classification & anomaly alerts
Actions are classified by risk, from financial transactions and data exports to PII access and configuration changes, and deterministic anomaly rules raise alerts with a tracked acknowledge-and-resolve lifecycle.
Human approval trail
Approve and deny decisions on flagged actions are recorded with the reviewer's reason and timestamp, as a permanent part of the same audit trail.
Signed evidence packages
Recorded events are exported as Ed25519-signed evidence packages that can be verified independently and offline. Evidence packages are not a legal compliance determination.
Session replay dashboard
A Next.js dashboard with agent, session and event explorers, alerts, approvals and evidence views, every screen reading the same tenant-scoped API.
Engineering highlight
Designed around failure, not perfection
AgentShield assumes prompt injection will succeed, models will make mistakes and instrumentation can be bypassed. So the record itself has to be trustworthy: events are immutable and hash-chained, and every capability (classification, anomaly detection, approvals, evidence) is built as a subscriber to that one event stream rather than as a separate system. New capabilities are additive, never rewrites.
Where it stands
Python SDK on PyPI
The Python SDK is publicly available on PyPI as agentshield-control-plane. The backend is self-hosted; there is no public hosted service.
AgentShield records and flags agent actions. It does not currently block actions before they occur.
Built to last
- Public package agentshield-control-plane: MIT licensed, Python 3.10+
- Tenant isolation on every table and query, with hashed per-tenant API keys and rate limiting
- SDK releases published through GitHub Actions with OIDC trusted publishing, no stored tokens
Technology
- Python
- FastAPI
- SQLAlchemy
- PostgreSQL
- Next.js
- TypeScript
- Ed25519
- Docker
More products
Discuss your project
Tell us what you need to build and we will talk through the scope, architecture and how an engagement could be structured.
Start a conversationRelated services