Claude Code Tooling — Tech Radar

Quadrant	Meaning
Adopt	In production, predictable behavior
Trial	Actively tested, value likely
Assess	To explore, not enough hindsight
Hold	Identified issues — do not generalize

Adopt

MCP — context7

Fetches up-to-date documentation for libraries directly from source.

Claude defaults to context7 over web search for library research
Avoids stale training data on fast-moving libraries (Leptos, Tokio, etc.)
Active by default once installed — for versioned APIs, prompt explicitly:
```
Use context7 to verify the current Leptos 0.8 API before generating code.
```

Trial

Plugin — LSP

Provides language server integration — type information, diagnostics, and cross-file navigation.

Significantly reduces the number of files Claude needs to read manually on large codebases
Most useful on projects with non-trivial type hierarchies

Known limitation: LSP server can fall out of sync after large refactors or incomplete compilations. Claude detects it but recovery has a token cost. If Claude reports inconsistent type information, restart the LSP server before continuing.

Hook — RTK (Rust Token Killer)

CLI proxy that filters and compresses command output to reduce token consumption.

Meaningful token savings on repetitive dev operations (build, test, git)

Known limitation: Over-filtering occurs on some commands — Claude receives incomplete output and reruns the command in a different form, which can negate the savings. Observe cases where Claude reruns a command unexpectedly — likely a filtering artifact.

Assess

MCP — github

GitHub integration — PR management, issue tracking, review comments directly from Claude.

Relevant for multi-repo workflows or when context-switching between code and issues is frequent
Risk: actions visible to others (comments, PR updates) — scope carefully

MCP — postgres / sqlite / mongodb

Direct database queries from Claude. MongoDB is the primary database.

Useful for debugging, data exploration, migration validation
Requires strict scoping — read-only access recommended

Hook — pre-tool-use (destructive commands)

Intercept destructive shell commands (rm, git reset --hard, git push --force, etc.) and prompt for confirmation before execution.

Complements Claude’s built-in caution but enforces it at the tool level

Hook — post-edit (auto-format)

Trigger cargo fmt / gofmt automatically after each file edit.

Removes the need to include formatting in quality gate reminders
Verify it does not conflict with LSP diagnostics mid-edit

Skill — review-pr

Custom PR review skill applying project-specific criteria: architecture rules, quality gates, naming conventions.

Skill — add-feature

Guided feature scaffolding following hexagonal + CQRS structure: domain model → commands → events → adapter → wiring.

MCP — Serena (oraios/serena)

Symbol-aware code intelligence via LSP exposed as MCP tools (find_symbol, find_referencing_symbols, insert_after_symbol, etc.). Supports 40+ languages.

Overlaps directly with the LSP plugin (already in Trial) but operates at the MCP level — potentially more token-efficient and without the desync limitation
If it delivers on the promise, it could replace the LSP plugin entirely
Requires uv + per-language server installation

MCP — Playwright (@playwright/mcp)

Browser automation via MCP — navigation, clicks, form input, screenshots, network interception, multi-browser (Chromium, Firefox, WebKit).

Natural fit for E2E testing on Leptos SSR and React portals
Higher-level than Chrome DevTools — cross-browser, headless-friendly, test-oriented

MCP — Chrome DevTools (ChromeDevTools/chrome-devtools-mcp)

Low-level browser access via Chrome DevTools Protocol — console logs with source-mapped stack traces, network inspection, performance tracing (CrUX), screenshots.

Complements Playwright: where Playwright handles automation, Chrome DevTools handles debugging and profiling
Caution: exposes all browser content to the MCP client — avoid sessions with sensitive data
Collects usage statistics by default (opt-out available)

MCP — markitdown (microsoft/markitdown)

Converts file formats (PDF, Word, Excel, PowerPoint, HTML, images, audio…) to Markdown for LLM consumption. Ships an MCP server package (markitdown-mcp).

Useful for feeding external documentation or specs to Claude without manual copy-paste
Output is optimized for machine consumption, not human-readable fidelity

Hold

MCP — sequential-thinking

Forces explicit step-by-step decomposition before responding.

Sonnet and Opus already do this natively through training — the MCP duplicates a built-in capability and adds token overhead with no observed benefit