Claude Code Templates review What installation changes on disk
Real four-component installation from pinned CLI source, with byte checks and MCP merging. Tests exposed overwrite behavior, dry-run writes and a missing skill returning exit zero.
AgentIndex · Published 2026-11-07 · Updated 2026-11-07
Claude Code Templates turns reusable agents, commands, skills and integrations into installable components. For an existing project, the important questions are where it writes, what happens to current configuration and how automation recognizes a failure. We ran the actual CLI and inspected files and exit states.
Four selected component types installed successfully. Downloaded files matched a pinned source snapshot, and the existing MCP entry survived merging. We also reproduced three limitations: agent reinstallation replaced a local edit, component --dry-run still wrote a file, and a missing skill printed a failure while returning exit zero.
Source identity and environment
On November 7, 2026, we pinned commit 177c92592977ae1f94a9d20585cd25de26e1314b. Its CLI package declares 1.32.1. This review executes that checkout with lockfile-installed dependencies; it does not claim byte-for-byte testing of the separately published npm tarball. The host was macOS 26.6.2 with Node.js 26.3.0 ARM64.
The package declares Node 14+, while dependency installation emitted newer engine requirements for some dependencies. We did not establish compatibility on old Node versions. Components went into disposable projects, with supported CCT_NO_TRACKING=true telemetry opt-out. We supplied no --prompt, launched no Claude task and executed no installed hook or database service. Pinned project description
The reproduction fixture retains source fingerprints, CLI output and installed-file hashes. Initial default Node requests returned ECONNRESET. Retrying with Node's --use-env-proxy and existing proxy settings succeeded without disabling TLS validation. We treat the earlier acquisition failure as an environment observation.
Four real installation paths
The command selected a code-review agent, test-generation command, PostgreSQL MCP configuration and frontend-design skill.
| Selection | Actual destination |
|---|---|
development-tools/code-reviewer | .claude/agents/code-reviewer.md |
testing/generate-tests | .claude/commands/generate-tests.md |
database/postgresql-integration | Project .mcp.json |
creative-design/frontend-design | Skill directory containing SKILL.md and LICENSE.txt |
The agent, command and two skill files matched corresponding pinned-source SHA-256 values. For MCP, we checked parsed configuration: the artificial pre-existing service remained alongside the new postgresql entry. This verifies the selected paths, not the entire registry.
Agent and command installation flatten the category away from the destination filename. A team installing overlapping selections should inspect target names instead of assuming catalog categories prevent collisions. Installation implementation
A configured MCP is not a connected service
The new entry supplied a command, arguments and a placeholder connection string. Writing these successfully does not demonstrate a running package, valid database access or useful queries.
We used no database credentials or production data. Users still need to replace placeholders and verify startup and authorization in the intended client. The demonstrated benefit is configuration assembly that preserves the selected existing JSON entry.
Reinstallation and dry-run behavior
After replacing the installed agent with a handwritten marker, we reinstalled it with --yes. The downloaded content replaced the marker. This finding applies to the selected agent path, rather than every component type; settings and plugin helpers have other merge rules.
In a separate empty directory, --agent … --dry-run --yes still created .claude/agents/code-reviewer.md and reported installation success. Although the help describes a preview without copying, this component path performed a write. CLI flags
For this version and path, a disposable target followed by a diff is a more dependable review process. Do not use this flag as a zero-write guarantee for component installs, or assume reinstalling preserves local edits.
Failure text can accompany exit zero
A deliberately nonexistent skill produced a not-found message, zero successful installations and a retry command. It created no requested skill file, yet the process exited zero.
A shell pipeline could therefore continue despite that missing component. Check expected files or the installation summary as well as process status. This single example does not establish that every failure returns zero.
We also ran 31 upstream Node tests for external-skill validation and plugin-install helpers; all passed. Some use simulated GitHub responses, so they are supplementary implementation evidence rather than live plugin installation. We record them separately from the actual downloads and disk writes.
Adoption decision
The CLI is useful for Claude Code users selecting a small, reviewed set of components. It reduces configuration collection work while leaving meaningful review responsibilities with the adopter.
Pinning the CLI alone does not pin all resources: ordinary component URLs in the tested implementation use main. Our live downloads matched the chosen snapshot, but future downloads can differ. Retain installed contents and their provenance when reproducibility matters.
We did not test skill activation, review quality, generated-test correctness, dashboards, plugin enablement or database functionality. For script-level behavior inside a skill collection, see the Anthropic Skills review. For a real local document service, see the OpenKnowledge persistence review.