We put an MCP endpoint in 49 business apps. Here is what a read-only key can and cannot do to an invoice register.
说明与用途
通过 MCP 端点实现对 49 个自托管业务应用的统一、受限读取访问,防止写入并确保审计记录。
问题
Integrating a Model Context Protocol (MCP) endpoint into multiple self-hosted business tools while ensuring secure, read-only access for language models.
结果
Uniform, secure read-only AI access to 49 self-hosted business apps without exposing write capabilities or sensitive operations.
工作流
- 1. Deploy the MCP endpoint on the self-hosted app (POST /mcp).
- 2. Generate a read-only API key (apk_…) via the app’s settings.
- 3. Configure the AI client (Claude, OpenAI, n8n, curl) with the endpoint URL, Authorization header, and optional require_approval flag.
- 4. The client sends a tools/list request; only read tools are returned.
- 5. The model can query invoices but cannot invoke write, delete, or email tools.
- 6. All actions are logged with the key’s user for audit.
- 7. If write access is needed, generate a separate key with appropriate permissions and set require_approval to "never" if desired.
证据
When you mint a key you choose what it may do. Pick read only, and the endpoint stops advertising the tools that change anything.
Everything a read-and-write key does lands in the audit log with the key's user beside it.
- 记录日期: 2026-09-05T18:08:12.051Z