Built for engineering teams.
One policy across your AI agent stack.
Conduct Guard enforces runtime policy across any MCP-compatible AI agent, model gateway, and MCP tool — before consequential actions execute.
Install in 10 minutes. Evidence for the CISO from day one.
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
Five agent tools shouldn't require five policy models.
Every new agent surface creates a new enforcement gap. The problem compounds every time a new tool lands in your stack.
One policy model per tool
Claude Code has hooks. Cursor has settings. Copilot has org controls. Each enforces differently, stores differently, audits differently. You can't write a rule once and trust it runs everywhere.
Consequential actions run without review
A refund processes. A production deploy lands. A secret is read. By the time you know, the action is done. Policies in documents don't stop actions at runtime.
No tamper-evident trail
Log files change. Agent activity disappears when the session ends. When security or compliance asks what happened, the answer shouldn't be 'we think.'
Allow. Approve. Block. Prove.
Four outcomes. Every agent action gets one. Runtime, not retrospective.
Allow.
Actions within policy proceed immediately. No friction for compliant work.
Approve.
Consequential actions pause for human review before they execute.
Block.
Actions outside policy are stopped before they reach a model or tool.
Prove.
Every decision lands in a hash-chained audit trail. Integrity is verifiable.
Control the action before it becomes an outcome.
Guard intercepts at the point of intent — not after a refund processes, a deploy lands, or a secret is read.
Write the rule once. Apply it where agents work.
One policy definition — one set of rules for which actions require approval, which are blocked, and which are audited. Guard applies it across your entire agent fleet: CLI hooks, LLM proxy (prompt + response gates), MCP layer, and Lens chat.
- →5 enforcement surfaces: CLI hook, LLM proxy (prompt), LLM proxy (response), MCP layer, Lens
- →6 BYO gateway adapters: Azure, OpenRouter, Portkey, Helicone, LiteLLM (Preview), ConductAI
- →35 pre-built playbooks with Guard enforcement built in
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
One policy across your agent stack.
Native platform controls stay in place. Conduct applies one policy and evidence model across the mix of agent tools your team actually runs.
CLI hook · HTTP proxy · MCP layer
SDKDrop-in base URL replacement. No SDK changes.
CLIPost-tool-use hook on Claude Code, Cursor, Codex, Copilot.
MCPGuard wraps MCP tool invocations before they reach the server.
Know exactly what happened — and why.
Every Guard decision is recorded with agent, action, resource, matched rule, reason, user, and timestamp — in a SHA-256 hash-chained audit trail. Altered entries break the chain. Export-ready for SOC 2, HIPAA, and PCI DSS.
- ✓Hash-chained integrity — not just logged, cryptographically ordered
- ✓Approval decisions captured with actor, timestamp, and rationale
- ✓Compliance report generation: SOC 2, HIPAA, PCI DSS
- ✓Export and verification API — machine-readable proof
We publish where Guard stops.
Every capability on this list maps to code in the repo. SHIPPED means it is in production. PREVIEW means it is working but not GA. PLANNED means it is on the roadmap, not in the codebase.
Last audit: 2026-09-01 · Source: automated codebase scan
- SHIPPEDPre-call policy evaluation (allow / block / approve / audit)
- SHIPPEDSHA-256 hash-chained audit trail
- SHIPPEDCLI hook — Claude Code, Cursor, Codex, Copilot
- SHIPPEDHTTP proxy enforcement
- SHIPPEDMCP tool interception
- SHIPPEDHuman approval gates (Slack + UI)
- SHIPPED15 compliance packs (OWASP, SOC 2, HIPAA, PCI, EU AI Act, NIST, ISO 42001...)
- PREVIEWKubernetes deployment templates
- PREVIEWLiteLLM Guard integration
- PLANNEDAir-gapped / on-prem deployment
- PLANNEDCross-agent workflow correlation (Operations)
Open where trust matters.
The enforcement engine, CLI, compiler, and developer tools are Apache-2.0. You can read, audit, fork, and self-host them. The hosted product adds workspace management, multi-user access, and the managed SaaS layer.
packages/conduct-cli
Agent lifecycle management, Guard sync, testing
apps/api/app/guard
Core enforcement engine — evaluation, scoring, audit
apps/api/app/compiler
YAML playbook definition and execution graph
tools/booster
Developer productivity tools for Claude and Cursor
Apache-2.0 includes an explicit patent grant. View on GitHub →
Deploy Guard where your controls need to live.
Start on SaaS in minutes. Move to self-hosted Docker when you need data residency. Kubernetes templates are in preview.
Air-gapped
- PLANNED
On-prem deployment with no external connectivity. On the roadmap.
Put runtime policy in front of your agents.
Agent Discovery mode runs for 14 days, read-only. See every agent action across your team before you enforce anything.