Directory

Claude Code tools

Six searches against the GitHub API turned up the fifty-two tools below, in six categories, with the licence, the star count and the last commit date on every row. Every result was read before it went in, and the ones that only matched the words were dropped.

Built 17 October 2026, with the orchestrator rows re-read on 23 October. Those rows now carry the last commit on each default branch rather than the date the API hands over, and the rest of the tables still carry the older reading.

Why the last commit column is the one to read first

Star counts measure popularity and nothing else. The column worth reading first is the last commit, because a tool in this ecosystem can go stale in a couple of months while its stars keep climbing.

Four widely recommended projects in this space are dead, deprecated or have quietly become something else, and all four still turn up in articles published this year.

$ gh api repos/OWNER/REPO --jq '"\(.stargazers_count) stars, pushed \(.pushed_at[0:10])"'

Running many agents at once

The orchestrators are the most crowded of the six categories. Every agent gets its own git worktree so that two of them cannot overwrite the same file. Where they part company is what the screen shows when an agent stops, taken apart in the longer comparison and again in the same list sorted by interface shape.

Status lines

A status line sits under the prompt and shows the context left, the model, the branch and the cost so far. Of the six categories on this page, this is the cheapest to install.

Usage and cost trackers

On a subscription the thing to avoid is a limit arriving in the middle of a task. On API billing the question is what a session actually cost. Both answers are drawn from the same local files, so the two kinds of tool share one table.

Watching what agents do

Dashboards, menu bar apps and observability built on hooks all land in this category, and it overlaps heavily with the orchestrators in the first table. The split is that a monitor watches agents something else started, while an orchestrator starts them itself.

Hooks and configuration

These change how Claude Code itself behaves rather than what sits around it. Hooks fire on tool calls and on prompts. Most of the monitoring in the table above is built that way.

Skills, commands and guides

The entries here are collections rather than programs. Most of what people need has already been written down by somebody.

What is missing from this list

Anything behind a sign-up. Every entry needed a public repository with a readable licence, which ruled out several hosted products that other lists carry. Conductor is the one exception, because it comes up in every conversation about orchestrators.

Editor extensions. VS Code and JetBrains plugins are a category of their own, left to their own marketplaces.

Anything that could not be verified. The searches turned up repositories whose descriptions did not match what their names claimed. Dropping them rather than padding the tables is why the page carries 52 tools and not 80.

How this list was built

Six searches went out to the GitHub API, one per category, and every result was read before the off-topic ones were dropped. A draw.io demo does not become a Claude Code tool because a query matched it.

$ gh api -X GET search/repositories \
    -f q='claude code statusline' -f sort=stars -f per_page=8 \
    --jq '.items[] | "\(.full_name) \(.stargazers_count) \(.pushed_at[0:10])"'

What the licence column means

MIT, Apache-2.0 and GPL-3.0 are open source, so anybody can fork a repository that carries one of them. "Source-available" means the code is readable but the licence restricts what else may be done with it. "No licence file" means exactly that, and it is legally the most restrictive of the three, because the default without a licence is all rights reserved.

Nothing here says whether a tool is good

Judging that would mean running fifty-two of them for a week each. The tables report only what a morning of reading can verify, which is what each tool says it does, what licence it carries, and whether anybody is still working on it.

Questions about picking a tool

Which Claude Code tool should I install first?

A status line. It costs a minute, it shows how much context is left, and it answers the question people ask most often in the first week. Everything else can wait until there is a problem it solves.

Do I need an orchestrator?

Not until more than two or three agents are running and getting hard to keep track of. Below that the CLI's own worktree support is enough and an extra app is overhead.

How do I know a tool is still maintained?

The last commit column. Anything quiet for more than two months in this ecosystem is worth a second look, because the agent CLIs change and somebody has to keep up with them.

Are these safe to install?

They run with the permissions of whoever installed them and most read the Claude Code configuration. The licence, the source and the number of contributors are all worth a look first, which is the same care any dependency deserves.

Do any of them cost money?

Almost none. Conductor is $50 a month for Pro. Everything else on this page is free, and most is open source under MIT, Apache or GPL.

Will they work with Codex too?

Some will. The orchestrators mostly do, because driving two CLIs is barely harder than driving one. The status lines and hooks are usually Claude Code specific.

Questions about the ecosystem

Why are there so many of these?

Because Claude Code ships as a terminal program with hooks, and that is an invitation. The barrier to building something useful on top is a weekend, which is also why so many of them go quiet.

Which ones are actually used?

Stars are the only public proxy and they are a poor one. A 1,000-star tool with a commit this week is usually a safer bet than a 20,000-star one that stopped in February.

Is there an official list?

Anthropic publishes three plugin marketplaces, an official one plus a community and a demo, which Claude Code adds on first run, and a web catalogue that marks entries Anthropic verified. Those cover plugins, so they overlap the hooks, skills and status line categories here and not the orchestrators. The community lists are large and useful, and they share the problem this page tries to fix, which is that entries go stale and nobody rechecks them.

What happened to the tools that disappeared?

Most were side projects whose authors moved on. One, Vibe Kanban, was a company that could not find a business model and said so plainly. The repository is unarchived and Apache-2.0, so keeping it is still an option.

What a good setup looks like, week by week

Most people end up with three or four of these rather than twenty.

Week one is a status line

$ claude
# then add a statusLine entry to ~/.claude/settings.json

Context remaining is the one number worth keeping on screen, because running out of it halfway through a task is the most common frustration people hit early.

Week two is a cost tracker

Even on a subscription, learning which kinds of task burn through a limit changes how the next one gets written. The usage tools above read the same local session files Claude Code already writes, so none of them need an API key.

An orchestrator, past three agents

$ git worktree add ../app-task -b task && claude

The command above is worth running by hand first. The point where it starts to annoy is the point where somebody can say which orchestrator they want, and twenty-three of them are compared here.

Hooks, only when something needs them

Hooks reach deepest into Claude Code, and they are the easiest way on this page to break a setup that was working. They pay for themselves when every tool call needs a check run against it.

Where Agentbox sits in the list

Agentbox is the project behind this page and sits near the bottom of the first table by stars. It runs dozens of Claude Code and Codex agents in parallel, each in its own git worktree, and hands the finished ones back as rows to answer like email or to work through as a board.

Agentbox is four weeks old, carries 94 stars against Orca's 102,844, has no native Windows build and has no company behind it. If any of that matters more than being free and readable, the tables above name the alternatives.