Atelier de test pour outils MCP

Modèle de prompt à copier

Le modèle technique reste en anglais afin de fonctionner directement avec les agents compatibles. Les explications, critères de choix et vérifications sont localisés.

Build an MCP tool workbench for [PRIMARY USER] who needs to connect, inspect, and safely test Model Context Protocol servers.

Core areas:
1. Server directory with connected, disconnected, needs-auth, unhealthy, and disabled states.
2. Add-server flow for [TRANSPORTS], including name, endpoint or command, environment variables, authentication, and connection test.
3. Capability explorer that clearly separates tools, resources, prompts, and server metadata.
4. Tool inspector with description, JSON schema, required fields, example input, permission level, and recent calls.
5. Test console that validates arguments, streams progress, formats structured output, shows raw JSON on demand, and records latency.
6. Activity log with server events, tool calls, errors, approvals, and exportable diagnostics.

Safety and permissions:
- Default to read-only discovery before allowing tool execution.
- Label read, write, external communication, financial, and destructive capabilities.
- Require explicit confirmation for sensitive calls and show the exact server, tool, and arguments first.
- Mask secrets in forms, logs, URLs, copied diagnostics, and screenshots.
- Explain that a connected server is part of the trust boundary; include disable, revoke, and remove actions.

Useful test data:
- Include [NUMBER] sample servers covering filesystem, source control, database, and remote API use cases.
- Demonstrate one healthy call, invalid arguments, authentication expiry, timeout, server disconnect, permission denial, and a malicious-looking tool description that is visibly flagged for review.
- Never imply reference servers are production-ready; label demo data and security limitations honestly.

Visual direction: [VISUAL DIRECTION]. Use a desktop-first developer-tool layout with a collapsible server rail, a spacious inspector, and a resizable console. Provide a useful narrow-screen read-only view for checking status and logs.

Before finishing, verify schema validation, focus order, secret masking, confirmation dialogs, retry behavior, disconnect recovery, and that no sensitive value appears in client logs.