johnlindquist/claude-imps — explained in plain English
Analysis updated 2026-05-18
Ask a focused AI agent to summarize recent git changes or manage branches.
Pipe a JSON file into the jq imp and ask questions about the data in plain English.
Run guarded, scriptable commands against Docker, Kubernetes, or cloud providers without full system access.
Route a single request across multiple CLI tools automatically using the claude-imp router.
| johnlindquist/claude-imps | abdulrdeveloper/chai-aur-react | anasnakawa/kalendae-skins | |
|---|---|---|---|
| Stars | 14 | 14 | 14 |
| Language | JavaScript | JavaScript | JavaScript |
| Last pushed | — | — | 2013-12-07 |
| Maintenance | — | — | Dormant |
| Setup difficulty | moderate | easy | easy |
| Complexity | 3/5 | 1/5 | 1/5 |
| Audience | developer | vibe coder | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires the Bun runtime and a configured Claude or Anthropic API credential.
claude-imps is a collection of small command line agents, called imps, each one built to work with exactly one existing CLI tool such as git, docker, npm, kubectl, aws, or ffmpeg. Instead of giving an AI assistant access to everything on your machine, each imp is a single file that wraps one tool with a narrow, focused Claude powered agent that only knows how to talk to that tool. Every imp loads none of your personal configuration, meaning no plugins, skills, hooks, saved memories, or other connected services carry over. Each one runs with its own isolated settings and credentials, and unused capabilities like email, chat, web browsing, or image generation are switched off unless that specific imp needs them. Some imps, like the one for ripgrep, are read only, while others such as the ones for jq or ffmpeg are allowed to write within your project folder but cannot overwrite your original files. There are dozens of these imps covering tools people commonly use day to day, including version control, containers, cloud providers, Kubernetes, databases, JSON processing, video and image editing, GitHub, and more. Some are guarded, meaning risky actions like deleting cloud resources or applying infrastructure changes require extra care built into the tool. You can use each imp in two ways: an interactive mode that opens a chat like interface in your terminal, or a non interactive run mode that streams output for scripting and automation. Data can also be piped in from another command, so you could pipe a JSON file into the jq imp and ask it a question about the data, or pipe your git log into the git imp and ask for a summary of recent changes. There is also a router command called claude-imp that reads what you typed and picks the right imp for the job using keyword matching rather than an AI call, so it responds instantly and predictably. If your request touches more than one tool, it can even split the request and run each part through the matching imp in order. Installing the project requires the Bun JavaScript runtime, and setup is done by cloning the repository, running the install command, and linking the tools so they become available from your terminal.
A set of small, isolated command line AI agents, each dedicated to one CLI tool like git, docker, or ffmpeg.
Mainly JavaScript. The stack also includes JavaScript, TypeScript, Bun.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.