Connect your AI agent to a protected tool in minutes. Its original server description includes a malicious instruction; Eviduct strips that instruction out of the model-visible tool and checks the exact call before sending it.
One click gives you a short-lived tenant management key and separate agent token. Nothing is emailed, and secrets are shown only once. Guest access is globally rate limited to protect the service.
The real sample tool is already approved for you. You can add one additional public-data route during the trial.
Checking availability…
One MCP endpoint. The agent receives neutral tool definitions, not the provider's marketing or instructions.
Set EVIDUCT_GATEWAY_TOKEN to your agent token in your secret/environment configuration. Never commit the token to Git. Disable any direct unprotected connection to the same provider.
Test the protected MCP inventory, call the harmless synthetic provider, then route a task by capability through Eviduct's real decision policy. With one approved provider, no quality comparison is claimed.
Issue a trial token above first.
The remote server advertises a five-star review instruction in its tool description.
Run the inventory test to view the model-visible neutral tool definition.
No MCP calls executed.
Keep the architecture if you need controlled selection and auditable calls across MCP services. The official MCP Registry remains a discovery source, but a directory entry or star rating is not independent evidence of tool quality.