87 commands executed and verified on macOS + Linux

Zero trial-and-error.
Every command, ready to run.

Describe the task. clihub returns the tool and the exact command that works, with the failures other agents hit and how to avoid them.

kill process on port replace text in files convert pdf to image run command with timeout extract tgz extract text from pdf screenshot a page query sqlite generate secret

Give it to your agent

Install as a skill

Agent reads it and stops guessing

https://clihub.com/.well-known/agent-skills/clihub/SKILL.md

Connect over MCP

No key, no install

{ "mcpServers": { "clihub": { "type": "http", "url": "https://clihub.com/mcp" } } }

Or just curl

Plain text, one call

curl -s 'https://clihub.com/a?q=kill+process+on+port&platform=linux'

84: tools
291: proven recipes
87: verified by execution
2026-10-03: last verification run

Verified recipes

Each command below was executed by our harness on both macOS and Linux. The failures listed are the ones agents actually hit.

All tools →

jq →

Extract a field from a JSON file

Verified macOS · Linux · 2026-09-17

jq -r '.name' package.json

Common failures (1)

output is wrapped in quotes-r prints raw strings without JSON quoting

yq →

Read a YAML value

Verified macOS · Linux · 2026-09-17

yq '.services.web.ports[0]' compose.yml

Common failures (1)

Error: unknown command / different syntaxthese recipes target mikefarah yq v4 (brew/apt yq). The Python yq (pip) uses jq syntax instead.

sed →

Replace text in a file in place (macOS/BSD sed)

macOS only

Verified macOS · 2026-09-17

sed -i '' 's/old/new/g' file.txt

Common failures (1)

on Linux: sed: can't read : No such file or directoryGNU sed takes no argument after -i: use sed -i 's/old/new/g' file.txt

Built for agents first

Every page here has a machine-readable twin. Nothing needs a key or an account.

Always send platform. BSD and GNU tools take different flags, so the right command on macOS is the wrong one on Linux. Say which you are and clihub withholds the form that cannot run, instead of returning both and leaving your agent to pick.

Endpoint	Returns
/a?q=<task>&platform=<p>	The commands as plain text. The cheapest call an agent can make.
/api/tools?q=<task>&platform=<p>&format=compact	The same answers as JSON, about a quarter of the bytes.
/api/tools?q=<task>	Full records; each hit carries matchedRecipes.
/tools/<slug>/md	One tool as Markdown: install, every recipe, known failures.
/mcp	Streamable HTTP MCP: get_command, search_tools, get_tool, list_categories.
/.well-known/agent-skills/clihub/SKILL.md	Installable skill, so an agent consults clihub on its own.
/llms.txt	Every surface above, plus the whole catalog as links.

Commands, not a terminal

Recipes that open a terminal UI are withheld from agent responses. Handing an agent htop produces a hung tool call, not an answer.

Verified means executed

87 commands were run by a harness on macOS and Linux, and a test fails the build if a claim has no passing result behind it. The rest are labelled as curated estimates.

curl -s 'https://clihub.com/api/tools?q=extract+text+from+pdf&platform=macos&format=compact'