Compare
One policy. Every agent.
Many AI governance tools cover one or two surfaces: the prompt, the gateway, the MCP server, or one framework. Conduct enforces a single policy wherever an agent acts, and runs alongside the tools you already use.
Coverage by approach
Typical coverage for each category. Individual products vary.
| Capability | Content guardrails | AI gateways | MCP gateways | Framework callbacks | Conduct |
|---|---|---|---|---|---|
| Inspects prompts and responses | Yes | Partial | No | Partial | Yes |
| Enforces policy on tool and MCP calls | No | No | Yes | Partial | Yes |
| Enforces policy on local CLI and shell actions | No | No | No | No | Yes |
| One policy across frameworks and vendors | No | Partial | Partial | No | Yes |
| Human approves a specific action before it runs | No | No | Partial | Partial | Yes |
| Spend caps that stop the call | No | Yes | No | No | Yes |
| Tamper-evident audit trail | No | Partial | Partial | No | Yes |
| Mapped to compliance frameworks | No | No | No | No | Yes |
● Yes · ◐ Partial · ○ No
Named tools
From each vendor's public documentation as of October 2026; CrowdStrike added November 2026. A dash means we found no documentation either way, not that the feature is missing. Spot an error? Email hello@conductai.ai and we'll correct it.
| Capability | Conduct | Lakera Guard | Zenity | Lasso | Prompt Security | Kong AI Gateway | Cloudflare AI Gateway | CrowdStrike Falcon AIDR |
|---|---|---|---|---|---|---|---|---|
| Inspects prompts and responses | Yes | Yes | Not documented | Yes | Yes | Yes | Yes | Yes |
| Enforces policy on tool and MCP calls | Yes | Partial | Yes | Yes | Yes | Yes | Partial | Yes |
| Enforces policy on coding-agent CLI actions | Yes | Not documented | Yes | Yes | Partial | No | No | Partial |
| One policy across LLM, MCP, and CLI | Yes | Partial | Partial | Partial | Partial | Partial | Partial | Partial |
| Human approves a specific action before it runs | Yes | No | Not documented | Not documented | Not documented | No | No | Not documented |
| Spend caps that stop the call | Yes | Not documented | Not documented | Not documented | Not documented | Yes | Yes | Not documented |
| Tamper-evident (hash-chained) audit trail | Yes | Not documented | Partial | Partial | Partial | Partial | Partial | Partial |
| Mapped to compliance frameworks | Yes | Partial | Partial | Yes | Not documented | Not documented | Not documented | Not documented |
● Yes · ◐ Partial · ○ No · – Not documented
Sources
- Lakera Guard: Part of Check Point. Docs list human approval gates as the customer's responsibility. [1]
- Zenity: Native hooks for coding agents and an MCP gateway. Audit logs; immutability not documented. [1][2]
- Lasso: Claude Code lifecycle hooks and an MCP gateway. Maps to NIST AI RMF, EU AI Act, ISO 42001, OWASP. [1][2]
- Prompt Security: Part of SentinelOne. Coding-assistant coverage is for data leakage. Searchable audit log. [1]
- Kong AI Gateway: Prompt guards, MCP tool ACLs, and spend limits via AI rate limiting. Network gateway only. [1][2]
- Cloudflare AI Gateway: Guardrails, DLP, spend limits that return 429. MCP portals allowlist tools. [1][2][3]
- CrowdStrike Falcon AIDR: Includes Falcon Guardian. Claude Code hooks check and block prompts and tool calls; an MCP proxy blocks or redacts tool calls; gateway and SDK collectors cover LLM traffic. Guardian allows or blocks whole agents on managed endpoints. Events go to AIDR logs and Next-Gen SIEM; immutability not documented. Acquisition of Fiddler (agent observability, to ship as Falcon Observe) announced November 2026, not yet closed. [1][2][3][4]
Where each approach stops
What each category is built for, and what it can't see.
Content guardrails
Filter what goes into and comes out of a model: injection, PII, toxicity.
Stop at the response. The tool call, shell command, or deploy the agent makes next is out of view.
AI gateways
Route, meter, and rate-limit LLM traffic across providers.
See model requests only. MCP tool calls and actions on a developer's machine never pass through.
MCP gateways
Allow or deny tool calls made over MCP.
Cover one transport. Model calls and CLI actions need a second and third policy.
Framework callbacks
Hooks inside one agent framework.
Policy lives in each app's code, so five frameworks means five copies of the rules.
Works with what you run
Conduct is the policy layer underneath, not a replacement. Keep your gateway, guardrails, and frameworks.
- Guardrails: NVIDIA NeMo Guardrails via
conduct-nemo-guard. - Gateways: LiteLLM via
conduct-litellm-guard, plus bring-your-own Azure, OpenRouter, Portkey, and Helicone. - Agent frameworks: Claude Agent SDK, OpenAI Agents, LangChain, Google ADK, and CrewAI via
conduct-agent-guard. - Coding agents: Claude Code, Codex, GitHub Copilot CLI, Cursor, and Windsurf via CLI hooks and MCP.
