MCP servers and agent workflows
Explore how tool-connected assistants fit into a developer workflow, and where a human approval gate belongs.
AdvancedOPINIONMCPAI AgentsTypeScript
Prompt
Design an MCP server exposing read-only access to our internal service catalogue. List the tools, their input schemas, the authorization model and the audit events emitted. Explicitly mark any tool that mutates state — we will not ship those in v1.
Problem
Assistants are only as useful as the context they can reach, but connecting them to internal systems is a security decision, not a convenience one.
Scenario
A platform team wants assistants to answer service-ownership questions without granting write access.
Architecture
- Read-only tools in v1
- Per-tool authorization
- Full audit log of every call
Why Bob?
Schema and handler scaffolding is mechanical; the authorization model deserves a design review with humans.
Prompt
Adapt the prompt above to your catalogue.
Bob Workflow
- Propose tool surface
- Generate schemas
- Document the audit trail
Generated Result
A tool inventory, JSON schemas and a handler skeleton.
Validation
- Test every tool with malformed input
- Confirm audit events for denied calls
Security
Treat every tool as a new API surface. Write access needs the same review as any production endpoint.
Lessons Learned
- Read-only v1 ships; write-capable v1 stalls in review
- Schemas are documentation — invest in the descriptions
Resources
- Model Context Protocol specification