An MCP Approval Server for AI Agents

说明与用途

Impri 的审批服务器为 AI 代理提供安全机制，允许其提出操作、等待人工批准，并仅在批准后执行，确保可审计性。

问题

AI agents can perform destructive actions without a reliable human approval mechanism, leading to potential mistakes or malicious behavior.

结果

Agents can safely propose destructive actions, receive human approval, and execute only upon approval, ensuring auditability and preventing accidental or malicious changes.

工作流

• Configure Impri MCP server in client config (set IMPRI_BASE_URL and IMPRI_API_KEY).
• Add Impri tools (impri_push_action, impri_await_decision, impri_report_result) to agent’s tool list.
• Wrap destructive actions (e.g., GitHub issue close) in a function that: 1. Calls impri_push_action to create pending action and notify reviewer. 2. Calls impri_await_decision to block until approved, rejected, or timeout. 3. If approved, perform the actual destructive action. 4. Call impri_report_result to log success or failure. 5. Return status to caller.

证据

Impri's MCP approval server (npx @impri/mcp) gives Claude Code and other MCP clients a real tool call to pause an agent and wait on a human's decision.

来源

impri_push_action, impri_await_decision, and impri_report_result show up as ordinary tools in the client's tool list.

来源

If the reviewer rejects it, or 30 minutes pass without a decision, closeIssueWithApproval returns without ever calling github.issues.update.

来源