Copy-paste setups that put Vulnify in front of an agent tool call. Vulnify is a runtime authorization layer: each action comes back ALLOW, REVIEW, or BLOCK, and the decision is written to the audit log.
REVIEW holds the action. A person approves or denies it in the Vulnify app or in Slack. BLOCK stops it. ALLOW runs it.
Each folder has a README, a gitignored .env for a TEST key from https://app.vulnify.io, pinned dependencies, and a mock mode that runs with no key. The first seven rows use one scenario: reading a file is ALLOW, exporting 800 records to an external email is REVIEW, and deleting a production database is BLOCK. The Claude Code hook maps that export to EXTERNAL_API because the tool call is a file upload, and it only checks the file read when VULNIFY_HOOK_CHECK_READS=true.
| Example | Stack | What it shows |
|---|---|---|
| langgraph/ | Python, LangChain tools, LangGraph | langgraph_tool_guard on a ToolNode. REVIEW and BLOCK become an error tool message. |
| google-adk/ | Python, Google ADK | before_tool_callback on an LlmAgent. REVIEW and BLOCK are returned to the model and the tool does not run. |
| crewai/ | Python, CrewAI | guard_crewai_tool on three BaseTools attached to an Agent and a Crew. |
| vercel-ai-sdk/ | Node, Vercel AI SDK | guardAiSdkTools before execute. A blocked call returns { blocked: true } to the model. |
| openai-agents-sdk/ | Python, OpenAI Agents SDK | guard_openai_agents_tool on each function_tool. The Runner gets the reason text instead of a result. |
| openai-agents-ts/ | Node, OpenAI Agents SDK | The same guard, with tool() and a scripted model from @openai/agents/testing. |
| claude-code-hook/ | Node, Claude Code | PreToolUse hook config plus vulnify-hook.mjs. Also npx @vulnify/cli audit, which stays on this machine. |
| claude-agent-sdk/ | Node, Claude Agent SDK | canUseTool asks Vulnify before sending email or rm. |
| mcp-proxy/ | Node, MCP SDK | A stdio proxy. Calls are forwarded only on ALLOW. |
| slack-review/ | Python | An 800-record external export gets REVIEW, then waits for Approve or Deny in Slack. |
| tinyfish/ | Python, TinyFish | A web agent fills out a form only after a Vulnify check. |
Mock mode does not call the API. It applies a local demo rule, so the production delete is BLOCK. A TEST key does not do that. The risk engine ALLOWs the read and sends the 800-record external export to REVIEW. The delete is DELETE_DATA with records_affected 1, and that score is often ALLOW until a policy BLOCKs DELETE_DATA.
The SDK clients have no method that creates a policy. In the app at https://app.vulnify.io, open Policies and create block-deletes: action family DELETE, resource type ANY, condition action DELETE_DATA, decision BLOCK, mode ENFORCE. The file is policies/block-deletes.yaml. It matches every DELETE_DATA, not only the resource name Production Database. A policy file has no field for that name. A rule that only blocks large deletes does not match these demos.
apply needs an org-wide LIVE key. The TEST key used to run an example is not enough.
npx -p @vulnify/sdk@0.3.3 vulnify policies validate policies/block-deletes.yaml
npx -p @vulnify/sdk@0.3.3 vulnify policies apply policies/block-deletes.yamlPython: pip install "vulnify[cli]==0.4.2", then vulnify policies validate policies/block-deletes.yaml and vulnify policies apply policies/block-deletes.yaml. Run the example again with the TEST key. The delete comes back BLOCK.
Keys stay in .env. Do not commit them.
skills/vulnify-guard-actions/SKILL.md tells Claude to call Vulnify decide_action before a consequential action and to act on ALLOW, REVIEW, or BLOCK. Copy the folder into ~/.claude/skills/ or upload it as a skill. It is guidance for the model. The Claude Code hook above is the enforced path.