Guard
Runtime policy for every AI agent.
Conduct Guard enforces runtime policy across any MCP-compatible AI agent, model gateway, and MCP tool — before consequential actions execute.
Allow. Approve. Block. Prove.
Install in 10 minutes. Evidence for the CISO from day one.
One policy across your AI agent stack.
Write the rule once. Guard applies it wherever your agents work — through the CLI hook, the HTTP proxy, and the MCP layer. No separate configuration per tool.
AGENT TOOLS
- Claude Code
- Cursor
- Copilot
- Codex
CLIENTS
- Claude DesktopMCP
- ChatGPTMCP
- CursorMCP
- Custom agentsproxy / MCP
BYO GATEWAYS
- Azure OpenAI
- OpenRouter
- Portkey
- Helicone
- LiteLLMPreview
- ConductAI native
Three decisions. One engine.
Guard evaluates every action before it executes and returns one of three decisions. The decision, the rule that fired, and the reason are all recorded.
Policy that can be inspected.
Policies are YAML files checked into your repository. No black-box rule engine. Any engineer can read, modify, and audit the rules that govern your agents.
policy: refund-cap version: "1.0" rules: - id: block-high-value-refund match: action: process_refund amount_gt: 500 decision: BLOCK reason: > Refunds over $500 require human approval per FIN-07.
Consequential actions
Guard was built for actions that cannot be undone. Refunds, production deployments, secret reads, network changes — a policy decision before they execute, not after.
Lens
“Ask Lens: ‘what did Guard block last week?’”
Every decision leaves a receipt.
Guard records the agent, action, resource, decision, rule, reason, and user for every evaluation. The receipt is hash-chained — altered entries break the chain.
Deploy Guard where your controls need to live.
Same policy engine, same CLI, same audit trail — regardless of where Guard runs.
SaaS
SHIPPED
conductai.ai · US-based
Docker
SHIPPED
Self-hosted Compose
Kubernetes
PREVIEW
Reference templates
Air-gapped
PLANNED
On-prem, no external calls
One policy where your stack isn't one vendor's.
Native platform controls stay in place. Guard applies one policy and evidence model across the mix of agent tools your team actually runs. Every route through Guard: explicit about what remains partial or planned.
Consequential actions guarded by policy
Refunds, deployments, network changes, secret reads — any action routed through Guard is evaluated before execution.
SHA-256 audit trail integrity
Hash chain on every decision. Altered entries break verification.
Pre-call prompt injection
Guard evaluates the tool call after the model has chosen it. Attacks that alter the model's intent before tool selection are upstream of Guard.
Cross-agent correlation — planned
Guard evaluates each action in isolation. Context from prior actions by other agents is not yet available. This is the Operations gap.