Skip to main content

8.3 Agentic customization: built-in vs custom vs Skills vs MCP

The four options​

OptionWhat it isYou maintainReuse across apps
Built-in / server-side toolsAnthropic-provided capabilities: web search, web fetch, code execution, bash, text editor, memory; in Claude Code also Read/Write/Edit/Grep/GlobNothingAutomatic
Custom toolsYour function definitions plus handlers in your processSchema, handler, testsOnly by copying code
SkillsA packaged procedure — instructions, optionally scripts and restricted tools — loaded on demandA markdown file (plus assets)Via the repo or a plugin
MCP serversA separate process exposing tools, resources, and prompts over a protocolA service with its own lifecycleAcross every MCP client

Choosing​

RequirementChoose
Public web informationBuilt-in web search / web fetch — do not write your own scraper tool
Run code, compute, transform dataBuilt-in code execution
Read and edit files in a repoBuilt-in Claude Code tools
Reach your database, internal API, or SaaS accountCustom tool or MCP server
One application needs it, nobody elseCustom tool — least infrastructure
Several applications need it, maintained independentlyMCP server
Encode how your team performs a recurring taskSkill — it is procedure, not capability
A capability plus a procedureBoth: a tool for the capability, a skill that describes the workflow
Large tool inventory, most irrelevant per requestTool search with defer_loading, or split by subagent role
The distinction items hinge on

A tool is a capability — something the agent can do that it otherwise cannot. A skill is a procedure — instructions for doing something with capabilities it already has. "The team keeps doing this five-step review differently" is a skill. "Claude cannot see our inventory levels" is a tool or an MCP server. Answering "write a skill" to a capability gap, or "write a tool" to a consistency gap, is the mistake being tested.

What each option costs​

OptionToken costLatencyOperational costRisk
Built-inDefinition in the prefix; results can be largeServer-side executionNoneOutput is untrusted external data
Custom toolDefinition in the prefixYour handler's latencyYour code, your testsYou own correctness and authorization
SkillZero until invokedIts own turnsA file to keep currentStale procedure silently followed
MCP serverDefinitions in the prefix, per connected serverProcess or network hopA service to run, version, and secureUntrusted output; a shared blast radius

Two consequences worth memorising: connecting many MCP servers is not free — every tool definition sits in the prefix of every request — and a skill costs nothing when unused, which is why procedure belongs in a skill rather than in CLAUDE.md.

Combining them​

A realistic support agent:

LayerChoice
Look up customers and ordersMCP server in front of the internal APIs — reused by the support agent, the ops dashboard agent, and Claude Code
Issue a refundCustom tool in the application process, so the approval gate and the caller's authorization stay in one place
Check public carrier statusBuilt-in web fetch
Summarise the shift's escalationsSkill — the procedure the team agreed on
Enforce the refund capHook — not a tool at all

The last row is the point: the customization question and the enforcement question are separate. Adding a tool never enforces a rule about that tool.

Quiz​

  1. Question 1

    Reviewers apply the team’s five-step code-review procedure inconsistently. Claude already has every tool the review needs, and the procedure should run identically whether a reviewer asks for it by name or simply asks Claude to review a pull request. What should be built?

  2. Question 2

    An agent needs current public information from the web, must run Python to compute a projection, and must read internal sales figures from a private warehouse. Which combination is correct?