CLI Tools

Since 1.7.5, hatch3r ships a first-class CLI-tools surface area as the token-efficient default for agentic coding. MCP is demoted to opt-in.

Why the pivot

External research and runtime measurement converged on the same conclusion: piping output through CLI binaries is consistently more token-efficient than tool-calling an MCP server for the same operation. Three primary signals drove the pivot:

• Anthropic engineering — Code execution with MCP (Nov 4 2025): code that orchestrates CLI binaries beats one-shot MCP tool calls when the operation can be expressed as a pipeline.
• GitHub Blog — Improving token efficiency in agentic workflows (May 7 2026): agent runs that prefer CLI tooling over MCP shave 30-60% off token consumption on file-walk / search / format / diff workloads.
• Cloudflare — Code Mode (Feb 20 2026) + ThoughtWorks Tech Radar Vol 34 (Apr 2026): industry consensus is pushing agents toward subprocess-tool composition over network-tool federation.

hatch3r still supports MCP — the pivot is about defaults, not deprecation. npx hatch3r init gates MCP behind a Yes/No prompt; npx hatch3r mcp setup reopens it on demand.

The 29-tool catalog

Three tiers. Tier-1 is default-on for every project. Tier-2 is pre-checked per detected project signals. Tier-3 is opt-in advanced.

Tier 1 — default-on (10 tools)

Tool	Probe	Description
ripgrep	rg	Fast recursive grep with sane defaults and gitignore awareness
fd	fd	User-friendly find replacement, gitignore-aware
jq	jq	JSON processor and query language
yq	yq	YAML processor (mikefarah Go implementation)
gh	gh	GitHub CLI — repos, issues, PRs, releases, gists
delta	delta	Syntax-highlighting git diff pager
bat	bat	cat clone with syntax highlighting and git integration
sd	sd	Intuitive sed replacement with literal string patterns
ast-grep	sg	Structural search and rewrite for code via AST patterns
zstd	zstd	Fast lossless compression with high ratio

Tier 2 — conditional (11 tools)

Pre-checked when the matching trigger holds.

Tool	Probe	Trigger	Description
Playwright	playwright	web-project	Browser automation, web testing, and UI interaction
duckdb	duckdb	data-project	Embedded analytical database with first-class CSV/Parquet
qsv	qsv	data-project	Fast CSV toolkit (slice, search, join, stats, 80+ commands) — actively-maintained xsv successor
taplo	taplo	rust-project / python-project	TOML toolkit (format, lint, query)
glab	glab	gitlab-remote	GitLab CLI — merge requests, issues, pipelines
az-devops	az	azure-remote	Azure DevOps work items, repos, pipelines via az CLI extension
Docker	docker	docker-detected	Container runtime and CLI
llm	llm	ci-llm-project	simonw/llm — invoke LLMs from the command line
fzf	fzf	interactive-tty	Interactive fuzzy finder for TTY pickers
lazygit	lazygit	interactive-tty	Terminal UI for git with keyboard-driven workflows
difftastic	difft	interactive-tty	Structural diff that understands syntax

Tier 3 — opt-in advanced (8 tools)

Never pre-checked. Opt in informed.

Tool	Probe	Description
RTK	rtk	CLI output-compression proxy — ⚠ see pipe-output corruption caveat in the canonical hatch3r-cli-rtk skill
Stagehand	stagehand	Browserbase Stagehand — AI-driven browser automation
aichat	aichat	Multi-provider LLM chat CLI with RAG and session memory
mods	mods	Charm mods — Unix-friendly LLM pipeline tool
Comby	comby	Structural search and replace across languages with declarative patterns
miller	mlr	awk/sed/cut/join for CSV/TSV/JSON/Parquet streams
csvkit	csvlook	Python CSV toolkit (csvlook, csvsql, csvjoin, csvstat)
Podman	podman	Daemonless container engine, rootless by default (Docker alternative)

Install commands per OS

The installer prints copy-paste commands grouped by package manager. It never executes on your behalf. Examples for the tier-1 default-on tools:

Tool	macOS	Linux	Windows
ripgrep	brew install ripgrep	sudo apt install ripgrep	scoop install ripgrep
fd	brew install fd	sudo apt install fd-find	scoop install fd
jq	brew install jq	sudo apt install jq	scoop install jq
yq	brew install yq	sudo snap install yq	scoop install yq
gh	brew install gh	sudo apt install gh	winget install GitHub.cli
delta	brew install git-delta	sudo apt install git-delta	scoop install delta
bat	brew install bat	sudo apt install bat	winget install sharkdp.bat
sd	brew install sd	cargo install sd	scoop install sd
ast-grep	brew install ast-grep	cargo install ast-grep	scoop install ast-grep
zstd	brew install zstd	sudo apt install zstd	winget install Facebook.Zstandard

Run npx hatch3r cli-tools install to print install commands for everything currently selected but not detected on PATH.

Decision tree — which tool for which job?

The hatch3r-cli-overview skill that ships in every project includes the full decision tree. Summary:

• Text search → rg (fast literal/regex grep). For structural patterns (AST-aware) use ast-grep instead. For cross-language declarative rewrites, escalate to comby.
• Find files → fd (sane defaults, gitignore-aware).
• JSON queries → jq. For tabular CSV/TSV pipelines, prefer qsv or miller.
• YAML / TOML → yq for YAML, taplo for TOML.
• View a file → bat (syntax-highlighted cat). For diff display use delta.
• Edit-in-place → sd for literal-string substitution. For structural rewrites use ast-grep or comby.
• GitHub / GitLab / Azure DevOps APIs → gh / glab / az (with the azure-devops extension).
• Compression / archive → zstd.
• Browser automation → playwright (deterministic) or stagehand (AI-driven).
• Data analytics on CSV/Parquet → duckdb.
• LLM in a pipeline → llm (simple), mods (Unix-friendly streaming), aichat (RAG + session memory).
• Containers → docker (default) or podman (rootless / daemonless).

The full umbrella skill (skills/hatch3r-cli-overview/SKILL.md) ships into every project and surfaces this decision tree directly to your agent at runtime.

Choosing between CLI tools and MCP

Dimension	CLI tools	MCP servers
Token cost per call	Low — output goes back as text via subprocess stdout	Higher — each tool description + each call carries protocol overhead
Latency	Local subprocess (ms)	Stdio / HTTP roundtrip (+startup if not warm)
Composability	Native shell pipelines (pipe, redirect, substitution)	One operation per tool call; chaining requires repeated tool-calling
Discoverability	command -v probe + skill recipes	Tool list announced by server at connect time
Provenance	Per-tool install command from registry (vendor-verified)	Server package + version pin
Secret handling	Env vars set in shell (.env.mcp source pattern reusable)	Server config injects env vars per call
Best for	File walks, grep, JSON/YAML/CSV processing, GitHub/GitLab API surfaces, diff/format/lint, container ops, structural rewrites	Authenticated remote services without a stable CLI (Linear, Sentry, Context7), tool surfaces tied to a vendor protocol

The pivot makes CLI tools the default, not the only option. npx hatch3r mcp setup reopens the MCP picker; both surfaces coexist.

Manage CLI tools

npx hatch3r cli-tools           # open picker (default action)
npx hatch3r cli-tools list      # selection + install status
npx hatch3r cli-tools install   # print install commands for missing tools
npx hatch3r cli-tools detect    # read-only detection report

Selection is recorded in hatch.json under cliTools.selected and survives npx hatch3r clean → npx hatch3r init via PreservedManifestFields. The umbrella hatch3r-cli-overview skill plus the per-tool skills are emitted to the 13 skill-capable adapters during sync.

Upgrading from 1.7.1 or earlier

The CLI tooling pivot lands in 1.7.5. Two upgrade considerations:

• --yes non-interactive MCP default flipped: npx hatch3r init --yes no longer auto-configures MCP. CI scripts must pass --mcp to opt back in.
• Opt-in flow on existing projects: Run npx hatch3r cli-tools to open the picker. Tier-1 plus matching tier-2 are pre-checked; detection runs against PATH and the installer offers per-OS commands.

See the Workflow guide for the full lifecycle context.