Client best practices

Patterns for scaling MCP host applications across many servers and tools.

Adapted from Client Best Practices.

Scale by adding clients, not by multiplexing unrelated servers on one connection.

Executive summary

Production hosts need namespacing, lazy discovery, approval policy, elicitation support, resilient transports and observability — not just a single demo connection.

Recommended patterns

1. One client per server

Keep isolation clear: crash, auth failure or version mismatch on one server should not take down others.

2. Namespace tools

Prefix or qualify tool names with server identity when aggregating for the LLM (filesystem.read_file vs bare read_file).

3. Lazy load and cache discovery

• Call server/discover / tools/list on connect.
• Honour TTL / cache headers when present.
• Refresh on listChanged notifications instead of polling blindly.

4. Human control policy

Tool class	Default policy
Read-only	May allow auto-approve after user opt-in
Writes / sends	Always confirm
Destructive / irreversible	Confirm + typed acknowledgement

5. Elicitation readiness

Implement form and URL elicitation so modern servers can request confirmation without failing the tool call.

6. Timeouts, cancellation, progress

Bound every remote call. Surface progress when servers emit it. Allow users to cancel hung tools.

7. Observability

Log: server id, tool name, latency, success/error class, protocol version. Never log secrets or full PII payloads by default.

Negative cases

• Dumping 200 tools into every prompt → context bloat; filter by relevance or user selection.
• Global auto-approve → unacceptable for enterprise.
• Shared mutable connection pool across unrelated tenants → isolation bug.
• Ignoring deprecation of sampling/roots → future breakage.

Give users visibility into what is connected. Source: modelcontextprotocol/docs.

Key takeaways

• Isolate clients; namespace tools; lazy-discover.
• Encode approval policy by risk class.
• Observe latency and errors without leaking secrets.

What to do next

Implement using an official SDK, then harden with Security best practices.