DEV Community: Deduped

How to set up Crush, the multi-model coding agent for your terminal

Deduped — Tue, 10 Nov 2026 02:00:33 +0000

Crush is Charm's open source coding agent for the terminal. It works with many model providers, lets you switch models mid-session, and runs on macOS, Linux and Windows. Here is a practical setup based on the Crush README.

Install it

Pick the package manager you already use:

# Homebrew
brew install charmbracelet/tap/crush

# npm (macOS, Linux, Windows)
npm install -g @charmland/crush

On Windows, winget install charmbracelet.crush works, and the README also lists Scoop, Arch, Nix, FreeBSD, and apt and yum repos for Debian, Ubuntu, Fedora and RHEL.

Connect a model

Run crush inside your project folder. The quickest start is the built-in model picker, which offers Charm's own Hyper provider. If you would rather use a provider you already pay for, press ctrl+l to open the picker, choose the provider and paste your API key.

You can also set an environment variable before launching. The README maps these to providers:

export ANTHROPIC_API_KEY=your-key
export OPENAI_API_KEY=your-key
export OPENROUTER_API_KEY=your-key
export GEMINI_API_KEY=your-key
crush

Use your own keys, and keep them out of shell history files you sync or share.

Why it is different from other terminal agents

Multi-model: switch LLMs mid-session and keep the context.
Session-based: keep separate sessions and contexts per project.
LSP-enhanced: it uses language servers for extra context, like an editor does.
Extensible: add MCP servers over http, stdio or sse.

Configure it with a crushrc

Crush runs fine with no configuration. When you want to customize it, the README describes a crushrc, which is Bash with Crush specific builtins, so it behaves the same on every platform. For example, to add a local Ollama provider:

provider add ollama --type ollama --base-url "http://localhost:11434/v1"

The README also says Crush ships with a built-in skill for configuring itself, so you can simply tell it what you want changed.

A sensible first session

Open a small repo and ask Crush to explain the layout. Reading is the safest first task.
Ask for one scoped change, such as a failing test or a missing type.
Review the proposed edit and any command before approving it.
Switch to a cheaper model for routine edits, keeping the same session.

Things to check

Model cost comes from your own provider account, so set spending limits there.
Treat MCP servers like any other code you run: only add ones you trust.
Release details change quickly, so check the README for current flags and providers.

More on Deduped

There is a step by step page for this: start coding with Crush in your terminal. To compare it with other agents, see best AI repos for coding agents.

Want the setup commands for 50 AI repos people actually ship with? Get the free AI Repos That Actually Ship pack: https://dedupedai.com/ship-pack/?utm_source=devto&utm_medium=article&utm_campaign=shippack

Get one email every Monday with the week's hottest AI repos and exactly how to use them: subscribe free at dedupedai.com

How to let your coding agent message your teammates' agents with Reason Network

Deduped — Mon, 09 Nov 2026 20:04:51 +0000

When you work on a team, your coding agent usually has no idea what your teammates' agents are doing. Reason Network, from the reason-machines/reason-network repo, connects them. It runs one small daemon per computer, peer to peer with no server, and messages land in the session you already have open.

What it supports

According to the README, every coding agent session on a teammate's machine shows up: who, which repo and branch, busy or idle, and what it is doing now. Any of them can message the right one. Supported agents include Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI, Grok CLI and opencode.

Install

The easiest way is to paste this line to your coding agent:

Join the Reason Network: https://reasonmachines.com/network

Or run it yourself in a terminal:

curl -fsSL https://reasonmachines.com/install.sh | bash -s -- network

On Windows, in PowerShell:

irm https://reasonmachines.com/network/install.ps1 | iex

Piping a remote script into bash is a trust decision. Open the URL in a browser and read the script first if you are on a work machine.

Setup flow

Sign in with Reason (free, with Google, GitHub, Apple or email). Your browser opens, and the agents on that computer are set up.
Your other computers join by signing in with the same account.
Friends join with your invite line. Whoever pastes it to their agent signs in and joins.

With an invite link from a terminal:

curl -fsSL https://reasonmachines.com/install.sh | bash -s -- network join '<invite>'

Commands you will use

reason-network who
reason-network send alice/api-claude "did the token format change?"
reason-network send alice/api-claude --notify "approve the deploy?"
reason-network send alice/api-claude --file schema.json "the schema I meant"
reason-network listen --as terminal
reason-network link
reason-network setup

who lists who is around and what their sessions do. send messages one agent, and --notify also pops a heads-up on that person's screen. --file attaches a small text file. setup shows which agents are on the network and how messages are handled.

Tools your agent gets

Agents get three tools under the MCP server reason-network: list_team, send_message and ask_user, plus check_inbox where messages cannot be pushed. Hosted assistants such as ChatGPT can join by adding the MCP server https://mcp.reasonmachines.com/mcp.

The approval pop-up

My favorite detail: when a teammate's agent asks yours for something only you can approve, ask_user shows a pop-up with Approve or Deny and a note. The answer goes back as yours, not as a message anyone could forge. That keeps a human in the loop for things like deploys.

Files and safety notes

A message can carry up to 10 text files, 256 KB each and 512 KB in total. The README says recipients save copies under ~/.reason/network/inbox, never executable and never opened for you. Traffic goes directly between computers, or through a relay that only sees encrypted bytes. Only team members can connect.

Where it fits

Good for two developers on one repo who want agents to ask each other about API changes without a Slack detour. Less good if your policy forbids a third party relay: the README says a team can run its own relay.

More on real usage, who it suits and what to watch for: Reason Network use cases on Deduped. If you run Claude Code in CI too, see Claude Code Action.

How to install Claude Code the recommended way (curl, Homebrew, WinGet)

Deduped — Mon, 09 Nov 2026 20:04:50 +0000

Claude Code is Anthropic's terminal coding agent. The install path changed: the npm route is now marked deprecated in the anthropics/claude-code README, and native installers are recommended. Here is a clean setup, straight from the README.

Install it the recommended way

macOS or Linux:

curl -fsSL https://claude.ai/install.sh | bash

Homebrew on macOS or Linux:

brew install --cask claude-code

Windows PowerShell:

irm https://claude.ai/install.ps1 | iex

WinGet on Windows:

winget install Anthropic.ClaudeCode

The old route still exists, but the README marks it deprecated:

npm install -g @anthropic-ai/claude-code

If you installed with npm a while ago, move to one of the native installers and follow the uninstall steps in the setup docs so you do not end up with two copies on your PATH.

Start a session

Go to your project directory and run:

cd your-project
claude

That is the whole start. Claude Code reads your codebase from the folder you launch it in, so launching from the repo root matters. It can explain unfamiliar code, run routine tasks, and handle git workflows from plain language prompts. The README also says you can use it in an IDE or tag @claude on GitHub.

First prompts worth trying

"Explain how authentication works in this repo and list the files involved."
"Run the tests, and fix the first failing one. Show me the diff before you apply it."
"Write a commit message for my staged changes."

Ask for diffs before edits on a codebase you do not know yet. You stay in control, and you learn what the agent assumes about your project.

Plugins

The repository ships several Claude Code plugins that add custom commands and agents. They live in the plugins directory of the repo, with their own README. Read what a plugin does before enabling it, since commands and agents can run with your permissions.

Report bugs from inside the session

Use the /bug command in Claude Code to file an issue without leaving the terminal. You can also open a GitHub issue.

Data you should know about

The README states that Anthropic collects feedback, including usage data such as code acceptance or rejection, conversation data, and anything you submit with /bug. It also lists limited retention periods and a policy against using feedback for model training. Read the data usage policies before using it on private code.

How people actually use it

Setup is the easy part. The interesting question is what people do with it afterward. We collected real use cases for the repo, with who they fit and what they solve: Claude Code use cases on Deduped. If you want to run it in CI or on pull requests, the sibling project is covered here: Claude Code Action use cases.

Quick checklist

Install with the curl, brew, irm or winget command for your OS.
Run claude from the project root.
Ask for a plan or a diff before any edit.
Review plugins before enabling them.
Skim the data usage policy once.

How to set up Cline in VS Code, JetBrains or the CLI

Deduped — Mon, 09 Nov 2026 02:00:25 +0000

Cline is an open source coding agent that runs in VS Code, JetBrains, a terminal CLI and a desktop app, with a human approval step before it acts. Here is a practical setup from the Cline README.

Pick where to run it

VS Code extension: install it from the VS Code Marketplace.
JetBrains plugin: install from the JetBrains Marketplace (IntelliJ IDEA, PyCharm, WebStorm, GoLand and others).
CLI, interactive or headless for CI and scripts:

npm i -g cline

Desktop app for macOS and Windows, which can run sessions in any folder and schedule routines.
SDK, if you want to build your own agent on the same engine:

npm install @cline/sdk

What it does in your project

Cline reads your project structure and makes coordinated edits across files. It watches linter and compiler errors while it works and fixes things like missing imports and type mismatches. In VS Code and JetBrains each edit appears as a diff you can review, change or revert, and checkpoints let you roll back the agent's work.

It also runs bash commands and watches the output live: installs, builds, tests. For long running processes like a dev server it keeps working in the background and reacts to new output, such as a crash or compile error.

Use Plan and Act deliberately

Cline has a Plan mode and an Act mode. A good habit:

Start in Plan mode and describe the task. Let it read files and propose steps, without changing anything.
Correct the plan if it misunderstood the codebase.
Switch to Act mode and approve each edit and command as it appears.

Keep approvals on

The human-in-the-loop prompts are the point. Approve file edits one by one at first, and read every shell command before allowing it. Only loosen approvals for commands you trust, like running your test suite.

Headless in CI

The CLI can run non-interactively for scripting. Start with read-only tasks, such as summarizing a failing build, before letting it change code in a pipeline. Check the CLI docs for the exact flags, since they change between releases.

Things to check

Model costs come from your own provider key, so cap usage on big repos.
Review diffs for secrets and for changes outside the files you intended.
The JetBrains plugin is not open source, per the README.

More on Deduped

Compare Cline with other tools in best AI repos for coding agents. If you add MCP servers to it, see best AI repos for MCP servers.

How to use Aider to pair program in your terminal with git undo

Deduped — Mon, 09 Nov 2026 02:00:25 +0000

Aider is a terminal pair programmer: it maps your repo, edits files, and commits each change to git so you can undo anything. Here is how to get it running against your own codebase, based on the Aider README, and how to keep its edits safe.

Install

python -m pip install aider-install
aider-install

Then go to your project and start it with a model of your choice. The README shows these three:

cd /to/your/project

# DeepSeek
aider --model deepseek --api-key deepseek=<key>

# Claude 3.7 Sonnet
aider --model sonnet --api-key anthropic=<key>

# o3-mini
aider --model o3-mini --api-key openai=<key>

Aider works with most hosted models and local ones too, so pick whatever you already pay for.

Why the repo map matters

Aider builds a map of your whole codebase and sends the relevant part with each request. That is what lets it work on larger projects instead of only single files. You add the files you want edited to the chat, and the map gives the model context on everything around them.

Git as your safety net

Every change Aider makes is committed with a sensible message. That gives you a simple loop:

Ask for one small change.
Look at the commit with git diff HEAD~1 or your usual git tool.
Keep it, or run git reset --hard HEAD~1 (or Aider's /undo) to drop it.

Work on a branch, and commit your own changes first so Aider's commits stay separate from yours.

Let tests and linters check its work

Aider can lint and test after each edit and try to fix what fails. Point it at your test command so a bad edit is caught before you read the diff. Review the flags in the README's lint and test docs, since the exact options depend on your stack.

Other ways to drive it

Add a comment in your editor and let Aider pick it up (IDE watch mode).
Add screenshots or web pages to the chat for visual context.
Speak requests with voice-to-code.
Copy and paste to a web chat if you only have a browser subscription.

Things to check

You pay your model provider for every request, so watch token use on large repos.
Never paste real API keys into shared shell history or commit them.
Review diffs on anything touching auth, payments or infrastructure.

More on Deduped

Aider sits in a crowded category. See the best AI repos for coding agents and, if you want to run several agents at once, best AI repos for running multiple coding agents.

Build from award winning sites with Awwwards-mcp in Claude Code

Deduped — Sun, 08 Nov 2026 20:02:16 +0000

Asking an AI agent to "make it look award winning" gives you generic output. Awwwards-mcp is an MCP server that lets Claude Code or Codex search Awwwards sites, turn a few of them into a cited design brief, and check your build against a reference. Here is the setup and three workflows from the README.

Add the server

For Claude Code:

claude mcp add awwwards -- npx -y awwwards-mcp

For Codex CLI:

codex mcp add awwwards -- npx -y awwwards-mcp

Some tools (page structure analysis, build verification, scroll maps) need Playwright. The README links the setup, so install that before using them.

Workflow 1: a brief from references

Ask the agent to call search_sites with a color, tag or free text such as "soft pastel e-commerce".
Shortlist 2 to 5 site slugs and write one clear project goal.
Call create_design_brief with those slugs and the goal. You get a Markdown direction plus JSON.
Build the page using only the cues the brief marks as evidence-backed.

The brief is only as good as the references you pick. Empty Interaction Plan fields mean no evidence was found, not that you should invent an effect.

Workflow 2: score your build against a reference

Run analyze_page_structure on the reference URL and on my local
file:// build, then run verify_build and fix the lowest scoring bands.

verify_build returns a score, a verdict and evidence lines. Fix what the lines name and rerun until the verdict is pass. It compares structure and motion fingerprints only, so a pass does not prove visual polish.

Workflow 3: scroll animation to a GSAP stub

Run analyze_scroll_map on a target URL for a per-element census, optionally get_motion_dna to confirm the animation libraries, then pass the census to build_gsap_stub for a single file page. Treat it as a starting point: trigger timing and easing usually need hand tuning, and keep the reduced-motion guard.

Watch for new winners

Call new_winners once to seed a baseline, add watches with watch_site for studios or tags, and call new_winners again later to see only what changed. It reports from a cached listing, so it never fetches on its own.

Things to check

Copying a site's motion is fine for learning. Do not ship a clone of someone else's design.
Results depend on the live Awwwards site layout, which can change.

More on Deduped

See the Awwwards-mcp overview and use cases. For more tools in this space, see best AI repos for coding agents.

Explain an unfamiliar codebase with Claude Code in your terminal

Deduped — Sun, 08 Nov 2026 20:02:15 +0000

You just joined a team and the repo has hundreds of files. Reading it file by file takes days. Claude Code runs in your terminal and can explain a module in plain language, name the files involved, and answer follow-ups. Here is the setup and the questions that work.

Install Claude Code

The README lists several install methods. The recommended one on macOS or Linux:

curl -fsSL https://claude.ai/install.sh | bash

Start in the project

cd path/to/your/project
claude

Claude Code reads the project from the directory you start it in, so launch it at the repo root.

Ask about one thing at a time

Broad questions give vague answers. Name a module or feature and ask for the flow plus the dependencies:

Explain how the billing module works. Walk through the main flow
and list the files it depends on.

A good answer names the entry point, the order things happen in, and the files to open. Ask follow-ups on whatever is still unclear:

Where does the retry logic live, and what calls it?
What would break if I changed the invoice model?

Verify before you trust it

Open one of the files it named and check that the function or flow it described is really there. A model can describe code confidently and still be wrong, so spot check before you start editing.

A prompt you can paste

Open the project at <project-path>, then explain how <module-or-feature>
works, including the files it depends on. Keep it short and name each file.

Things to check

The README says feedback sent through the /bug command can include conversation data. Avoid putting secrets or private code in a session you plan to report.
Claude Code is an agentic tool: it can also edit files and run commands, so for pure exploration tell it to only read and explain.

More on Deduped

The step by step card for this is Explain complex code in an unfamiliar project, and the repo overview is here. To compare other tools, see best AI repos for coding agents.

Dictate prompts into any app with OpenWhispr, privately

Deduped — Sun, 08 Nov 2026 02:04:56 +0000

Typing prompts to a coding agent is slow, and cloud dictation sends your voice somewhere you may not want. OpenWhispr is an open source desktop app: press a hotkey, speak, and the words appear at your cursor in any app. Here is how to install it and what the README says it can do.

Install

Download the build for your platform from the releases page: a .dmg for macOS, an .exe for Windows, or .AppImage, .deb, .rpm and .tar.gz for Linux. On macOS or x86_64 Linux there is also a community Homebrew cask:

brew install --cask openwhispr

To run it from source (Node.js 24+ required):

git clone https://github.com/OpenWhispr/openwhispr.git
cd openwhispr
npm install
npm run dev

Dictate into any app

The core feature is a global hotkey with automatic pasting. Click into your editor, terminal or chat box, hold the hotkey, talk, and the text lands at the cursor. That works for long prompts in Claude Code or Cursor, where speaking is faster than typing.

Keep it private

The README says all core features work with local models or cloud providers. For offline use, pick a local speech-to-text model such as Whisper or NVIDIA Parakeet. Local Whisper can be GPU accelerated on Metal, CUDA and Vulkan (AMD and Intel). With a local model, audio stays on your machine.

More than dictation

A voice assistant hotkey that sends what you say straight to an AI model as a command, and edits highlighted text in place
Meeting transcription that detects Zoom, Teams and FaceTime calls, with speaker labels
Notes with folders and semantic search
Audio import for existing files or a pasted YouTube or audio URL
A public API and an MCP server, so an AI assistant can read your notes and transcriptions

Things to check

Intel Macs lose live speaker identification and voice fingerprinting, because ONNX Runtime stopped shipping macOS x86_64 binaries in 1.24. Meetings still record.
Cloud providers are optional. If you connect one, your audio or text goes to it, so read its policy.
The mobile app is a separate Expo project in openwhispr-mobile with its own build process.

What we did not verify

We worked from the README and did not install the app for this post. Hotkeys and model names change between releases, so confirm in the app settings.

Our walkthrough is on the dictate into any app privately page, with the repo overview here.

Give Claude Code a browser with the agent-browser CLI

Deduped — Sun, 08 Nov 2026 02:03:13 +0000

Browser automation for coding agents usually means a heavy Playwright setup and a lot of tokens spent on page dumps. vercel-labs/agent-browser is a native Rust CLI that gives an agent short commands instead. Here is the install and the loop from its README.

Install

The README recommends a global install, which pulls the native binary:

npm install -g agent-browser
agent-browser install  # Download Chrome from Chrome for Testing (first time only)

On macOS you can use brew install agent-browser, and there is a Cargo option too. On Linux, add system libraries with agent-browser install --with-deps. To pin a version per project, use npm install agent-browser and then agent-browser install. Upgrade later with agent-browser upgrade.

The snapshot and ref loop

The core idea is that the agent reads an accessibility tree with short refs, then acts on a ref instead of writing a selector:

agent-browser open example.com
agent-browser snapshot                    # Get accessibility tree with refs
agent-browser click @e2                   # Click by ref from snapshot
agent-browser fill @e3 "test@example.com" # Fill by ref
agent-browser get text @e1                # Get text by ref
agent-browser screenshot page.png
agent-browser close

Traditional selectors still work when you want them, for example agent-browser click "#submit" or agent-browser find role button click --name "Submit".

Why it suits Claude Code, Codex and similar agents

Each command is one shell call, so any agent that can run a terminal can drive a browser with no MCP server to register. The README also notes that clicks fail early when another element covers the target, such as a consent banner. Dismiss the covering element, then take a fresh snapshot before retrying.

To keep screenshots cheap, the README recommends:

agent-browser screenshot --if-changed

That skips an image when the page has not visibly changed, so an agent polling a page does not pay for the same picture twice.

Skills that match your installed version

The CLI serves its own instructions for agents:

agent-browser skills list
agent-browser skills get <name>

The README says this content always matches the installed CLI version, which avoids stale cached copies in your agent setup.

Things to check

The browser is Chrome for Testing, downloaded on first install. Existing Chrome, Brave, Playwright and Puppeteer installs are detected automatically.
Building from source needs Node.js 24+, pnpm 11+ and Rust. A normal install does not.
WebMCP support is experimental, and page-provided tool descriptions are untrusted data. Do not let an agent follow instructions found in page content.

What we did not verify

We followed the README and did not run these commands on a fresh machine for this post. Flags change between releases, so check agent-browser --help after installing.

Our step-by-step is on the let your agent browse with short CLI commands page, and the repo overview is here.

Review every pull request with Claude Code Action on GitHub

Deduped — Sat, 07 Nov 2026 20:03:57 +0000

You can have Claude review every pull request and answer @claude mentions in issues without running a server. anthropics/claude-code-action is a GitHub Action that runs on your own runner. Here is the setup from its README, plus the review pattern it points to.

What it does

The README calls it a general-purpose Claude Code action for GitHub PRs and issues. It picks its mode from your workflow context, so you do not configure it: it responds to @claude mentions, to issue assignments, or to an explicit prompt in a workflow. It can answer questions about code, review a PR, and implement simple fixes. It runs on your GitHub runner, and API calls go to the provider you choose: Anthropic direct, Amazon Bedrock, Google Vertex AI or Microsoft Foundry.

The quickest setup

Open Claude Code in your repo terminal and run:

claude
/install-github-app

The command walks you through installing the GitHub app and adding the required secrets. Two catches from the README: you must be a repository admin, and this quickstart is only for direct Anthropic API users. For Bedrock, Vertex or Foundry, follow docs/cloud-providers.md in the repo.

Make it review every PR

The README links a Solutions Guide (docs/solutions.md) with complete examples. The review patterns it lists:

Automatic PR code review
Path-specific reviews that trigger on critical file changes
Special handling for external contributors
Custom review checklists that enforce team standards
Security-focused, OWASP-aligned reviews
Issue triage and labeling
Scheduled repository health checks

Start with the automatic review example from that guide, copy it into .github/workflows/, and open a test PR. Then narrow it: limit it to the folders where a mistake is expensive, and add your own checklist to the prompt.

Tip: the v1.0 inputs are prompt and claude_args. If you copied a workflow from an older blog post, the README's Migration Guide covers upgrading from v0.x.

Things to check before you trust it

Tool access is limited by default. Extra tools are enabled through configuration, so give it only what the review needs.
Read docs/security.md for access control, permissions and commit signing before you let it push code.
Treat its review as a first pass. It does not replace a human approving the merge.

What we did not verify

We followed the README and did not run the action on a live repository for this post.
Workflow syntax changes between versions, so confirm against the repo's current docs.

Our walkthrough of the install is on the install the GitHub app from Claude Code page, and the repo overview is here. For alternatives, see best AI repos for code review.

Install Addy Osmani agent-skills in Claude Code: spec, plan, build, review

Deduped — Sat, 07 Nov 2026 20:02:14 +0000

Most "AI coding" advice stops at the prompt. addyosmani/agent-skills goes after the process: it packages the habits senior engineers use (spec first, small tasks, tests as proof, review before merge) as skills your agent follows every time. Here is how to install it in Claude Code and what you get.

What it is

The README describes production-grade engineering skills for AI coding agents, organized around a lifecycle: define, plan, build, verify, review, ship. Nine slash commands map to those phases:

/spec: spec before code
/plan: small, atomic tasks
/build: one slice at a time
/test: tests are proof
/constraints: decide the quality bar once
/review: improve code health
/webperf: measure before you optimize
/code-simplify: clarity over cleverness
/ship: faster is safer

Skills also activate on their own. Per the README, designing an API triggers api-and-interface-design and building UI triggers frontend-ui-engineering.

Install in Claude Code

The marketplace route, straight from the README:

/plugin marketplace add addyosmani/agent-skills
/plugin install agent-skills@addy-agent-skills

If the install fails with git@github.com: Permission denied (publickey), the README gives two fixes. Add the full HTTPS URL when you add the marketplace, or rewrite GitHub SSH URLs to HTTPS once:

git config --global url."https://github.com/".insteadOf git@github.com:

To try it from a local clone instead:

git clone https://github.com/addyosmani/agent-skills.git
claude --plugin-dir /path/to/agent-skills

One naming catch: Claude Code already has a built-in /review, so use the namespaced form /agent-skills:review.

Not on Claude Code?

The README lists a one-command path for 70+ agents (Claude Code, Cursor, Codex, Copilot, Cline) through the open skills CLI:

npx skills add addyosmani/agent-skills            # install all 25 skills
npx skills add addyosmani/agent-skills --list     # browse before installing

Or pick single skills:

npx skills add addyosmani/agent-skills --skill code-review-and-quality
npx skills add addyosmani/agent-skills --skill interview-me
npx skills add addyosmani/agent-skills --skill test-driven-development

The README warns that a per-skill install copies only the skill folder, not the shared references/ directory, so supplementary checklists will be missing. Install the whole repo if you want them. For Gemini CLI the command is gemini skills install https://github.com/addyosmani/agent-skills.git --path skills.

A first session

Try the lifecycle on a small feature rather than a rewrite:

Run /spec and describe the feature. Let the agent ask questions before it writes anything.
Run /plan and check that the tasks are small enough to review one at a time.
Run /build for the first slice, then /test.
Run /agent-skills:review before you merge.

The README also mentions /build auto, which plans once, asks for your approval, then implements every task in one pass. It says each task is still test-driven and committed individually, and that it pauses on failures or risky steps.

Who it is for

Anyone whose agent jumps straight to code and leaves you the cleanup. If you already have a strict team process, treat the skills as a starting point and edit them.

What we did not verify

We followed the README and did not run this end to end for this post.
The skill count and commands may change, so check the repo before you rely on them.

For a step-by-step version, see our agent-skills page and the spec, plan, build walkthrough. More like it are in best AI repos for coding agent skills.

Build a feature with one /autopilot command in Claude Code

Deduped — Fri, 06 Nov 2026 20:04:01 +0000

Driving Claude Code through plan, code and QA by hand takes a lot of prompts. oh-my-claudecode (OMC) is a plugin that adds multi-agent orchestration on top: describe a feature once and let staged agents work through it. This is how the README sets it up, and what to check before you trust the result.

Install it as a Claude Code plugin

These are Claude Code slash commands. Enter them one at a time, because pasting both lines at once fails, as the README warns:

/plugin marketplace add https://github.com/Yeachan-Heo/oh-my-claudecode

Then:

/plugin install oh-my-claudecode

Run the setup command inside the session:

/omc-setup

If you prefer the npm CLI path instead of the marketplace flow, the README gives:

npm i -g oh-my-claude-sisyphus@latest

npm may print a prebuild-install deprecation warning. The README says it comes from an upstream dependency and has no repo-side fix yet.

Build a feature with /autopilot

Give it a feature in your own words:

/autopilot "build a REST API for managing tasks"

Autopilot runs in stages: plan, build, check. If you want a fixed sequence, the README describes a plan-build-qa profile configured under autopilot.workflows in .claude/omc.jsonc. Whatever profile you pick, review the diff and run the tests before merging. An agent that checks its own work is still an agent checking its own work.

A safer first run is to ask it to stop after the plan:

Install Yeachan-Heo/oh-my-claudecode in Claude Code, run /omc-setup, then /autopilot to build <feature> in this repo. Stop after the plan so I can approve it before execution.

Clarify the idea first with /deep-interview

Agents that start from a vague prompt bake in assumptions you never agreed to. OMC includes a Socratic interview that asks questions and measures clarity across weighted dimensions:

/deep-interview "I want to build a task management app"

Answer the questions, then hand the clarified spec to /autopilot or /team.

Run several agents in parallel

The CLI can spawn tmux worker panes, including workers from other providers:

omc team 2:codex "review auth module for security issues"
omc team 1:claude "implement the payment flow"
omc team status
omc team shutdown

Inside Claude Code the native version is /team 3:executor "fix all TypeScript errors". Native teams need CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS set to 1 in ~/.claude/settings.json. Without it, OMC falls back to non-team execution.

Is it too much?

The README itself says some people find OMC overkill and points to a lighter project by the same author, gajae-code. That is a useful signal. If you only need one agent on one small task, plain Claude Code is fine. OMC pays off when you repeatedly run the same plan, build and review loop, or want parallel reviewers.

What we did not verify

We followed the README and have not run this end to end for this post.
You need Claude Code installed and a plan that covers the extra usage multi-agent runs consume.
Check the repo for newer instructions before you rely on these steps.

The step-by-step version is on our use-case page: build a feature end to end with one /autopilot command. The repo overview is here.