Built-In Tools vs Custom Tools vs Skills vs MCP.

Architects extending what Claude can do choose among four distinct mechanisms, each with a different trade-off of setup cost, reusability, and scope. Built-in tools are platform-provided (Read, Write, Edit, Bash, Grep, Glob, web search) — zero setup, zero maintenance. Custom tools are hand-written functions or API calls you define, typically served through an in-process MCP server in the Agent SDK. Agent Skills package instructions plus optional scripts and resources that Claude loads automatically when relevant. MCP servers expose external systems as tools over a standard protocol, callable locally or remotely.

Anthropic explicitly positions these as complementary layers in the same stack rather than competing options. The architect's job is fit-for-purpose selection: the same agent often uses built-in tools for file work, a custom tool for one private API, a Skill for a repeatable procedure, and MCP for org-wide integrations — all at once.

Built-in first. Reach for platform tools for straightforward, native, well-defined tasks — they are the fastest path with no external setup and no maintenance overhead. A file-editing agent that reinvents Read/Write/Edit is over-building.

Applying built-ins effectively. Within the built-in set, match the tool to the operation: `Read` for a known file path, `Grep` to search file *contents* by pattern, `Glob` to find files by *path* pattern, `Bash` for shell commands and anything without a first-class tool, and `Edit` for a targeted in-place change versus `Write` for a full-file replace. The exam rewards NOT dropping to Bash (e.g. grep/cat/ls) when a dedicated tool exists — first-class tools are more reliable, cheaper to reason about, and easier to permission.

Custom tools for private, one-off access. When you need a proprietary API or internal service no built-in covers, define a custom tool via the Agent SDK's in-process MCP server so Claude can call your functions directly. Running in-process avoids the IPC/round-trip of an external server, so latency is typically lower for a single private call.

Skills for reusable, multi-step procedures. A Skill is the right unit when a capability is a stateful, orchestrated workflow you want Claude to invoke automatically and reuse across sessions — not a single function call. A Skill can itself call custom tools or MCP servers for its steps, so it composes with the other layers rather than replacing them. Constraint: API Agent Skills run via code execution, which is excluded from Zero Data Retention — a ZDR-required flow can't rely on them (this is about API Agent Skills; Claude Code skill files are a separate surface with their own ZDR rules).

MCP when you connect external systems. Prefer an MCP server when the job is integrating outside services and you want one shared, governable integration point across a team rather than a bespoke per-app function. Note the setup asymmetry: consuming a pre-built MCP server is a one-line `.mcp.json` entry; authoring your own is the higher-effort path. MCP is a protocol — the centralized auth, audit, and versioning are governance *you build on it*, not automatic.

Anthropic teaches this as a live refactoring skill. A Code w/ Claude 2026 session frames it as "tool, skill, or subagent?" — decomposing an agent whose logic has outgrown its prompt, categorizing each capability, and running evals after each change to confirm the split didn't regress behavior. The decision is iterative, not a one-time pick.