7 Claude Code Mods That Actually Fix Real Workflow Problems
Seven Claude Code mods explained with install steps: File Tree, Replay Theater, Blast Radius, Collision Guard, Agent Flow, and Reflect.
What are Claude Code mods and why do they matter?
Claude Code mods are extensions that run inside Anthropic’s coding agent and react to events, add interface panels, or intervene before a tool executes. They solve a specific problem: Claude Code can make large, fast changes to a codebase, but once it’s done you’re left reconstructing what happened, checking whether a risky command actually ran, or noticing too late that two parallel sessions edited the same file. Seven mods, built by independent developers and Anthropic’s own playground repository, address these gaps directly: visibility into files, replay of edits, warnings before destructive commands, conflict detection across sessions, visibility into delegated subagents, and a way to turn repeated corrections into saved instructions.
TL;DR
- File Tree adds a project browser next to the Claude Code conversation so you can see which files Claude is touching without switching to your editor.
- Replay Theater lets you step through the file edits from the latest editing turn, but it records edit requests, so you still need
git diffand your test suite to confirm changes actually landed. - Blast Radius intercepts certain risky bash commands (like
rm -rf) and shows what will be affected before you approve or cancel, though it only recognizes specific patterns and can miss scripts, aliases, or combined shell commands. - Collision Guard flags when a second Claude session tries to edit a file another active session recently touched, and can offer a git worktree as an alternative path.
- Agent Flow gives you a live panel showing which delegated subagents are running, their current tool, and elapsed time, useful when you fan work out across parallel agents.
- Reflect (specifically the “reflect mod” plugin, not the older Claude Reflect plugin) turns a repeated correction into a proposed instruction you review and save into your project’s config file.
- None of these mods make their own model calls just to render their panels, so the display itself doesn’t add to your usage bill. The agents and edits they’re observing still consume your normal Claude allowance.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
How do you install and verify a Claude Code mod?
Mods install through Claude Code’s plugin system, but installing a plugin isn’t the same as installing a mod. A plugin that only contains a skill won’t behave like the interactive mods described here. Before installing anything, check your version with claude --version. These mods were demonstrated on version 2.1.287 or newer, where mods are enabled by default (older experimental environment variable flags aren’t needed anymore). If you’re on a native install, run claude update first.
The general install pattern is:
- Add the plugin’s marketplace repository with
claude plugin marketplace add <repo>. - Install the specific plugin with
claude plugin install <name>@<marketplace>, often with--scope userso it’s available across projects. - Start a fresh Claude session (mods don’t load into a session that’s already running).
- Run
/pluginto confirm it loaded.
Because these extensions run with your permissions, check the publisher and repository before installing anything. Some mods, like Replay Theater and Blast Radius, come from a local marketplace inside Anthropic’s own Claude Code Playground repository, so you clone that repo and register it locally rather than pointing at a remote marketplace.
What does File Tree actually do?
File Tree adds a project browser directly inside the Claude Code terminal interface. Instead of alt-tabbing to your editor every time you want to know what folder Claude is working in, you get a live tree showing activity as Claude reads and edits files. It requires full screen mode (/2e fullscreen then /filetree) and a terminal at least 110 columns wide, or the panel won’t render.
Beyond passive viewing, you can scope the tree to a subfolder (/filetree /src) and select specific files to attach as context for your next prompt, which beats retyping file paths into every message. It’s worth noting that drawing the tree doesn’t cost extra tokens by itself, but if you use it to attach more file content to a prompt, that content still counts toward your context usage. File Tree won’t replace a full IDE, but for anyone who prefers staying inside Claude’s terminal rather than bouncing to an editor, it closes a real gap.
How does Replay Theater let you review Claude’s edits?
Replay Theater steps through the file edit requests from the most recent editing turn, so instead of scrolling back through a long conversation to figure out what changed, you move through the edits with next/previous controls (or N and P keys) via /replay.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
The catch is that Replay Theater shows edit requests, not confirmed outcomes. A denied or failed edit can still appear in the replay, so it doesn’t prove a change actually reached your files. The practical workflow is to use the replay to understand the sequence of changes, then close the panel and run git diff and your test suite to confirm the real state of the code. It’s also a session view, not permanent history: a new editing turn overwrites the previous replay, and restarting the session loses it. For a focused refactor, though, it’s a faster way to follow what happened than reading a raw transcript.
Is Blast Radius worth using before destructive commands?
Blast Radius is built for the moment an agent decides the cleanest fix is to delete a folder or run a reset command. It watches for certain risky bash patterns, holds the command, and shows a report of what will be affected, with proceed and cancel controls, before anything executes.
In practice this means you can cancel an rm -rf call, verify nothing was touched, then approve it separately and confirm only the intended files are gone. It’s genuinely useful as a second check against an expensive mistake. But it has real limits: it recognizes specific bash patterns, so a script, a shell alias, a different tool, or a command chained together with other operations can fall outside its coverage. It should be treated as an added layer, not a replacement for reading the actual command Claude is about to run or for keeping normal permission prompts in place.
What problem does Collision Guard solve?
Collision Guard targets a problem that shows up once you start running multiple Claude Code sessions against the same project: two sessions editing the same file without either one knowing about the other. Both sessions need the mod loaded for it to work, since tracking relies on shared information between active chats.
When a second session tries to edit a file another session recently touched (the default look back window is 30 minutes), Collision Guard surfaces a decision before the edit proceeds. For files tracked in git, it can also offer to move the second task into an isolated git worktree, letting that session continue without colliding, though any changes already made in the original checkout still need to be resolved manually. It tracks Claude’s own editing tools on the same machine, so edits made through a bash-invoked formatter, a separate editor, or another machine won’t be visible to it. It’s best suited for running parallel Claude sessions inside one checkout, with separate worktrees still being the safer default for genuinely independent tasks.
How does Agent Flow help with delegated subagents?
When Claude Code delegates work to subagents, the parent conversation can go quiet even while a lot is happening underneath. Agent Flow adds a panel (/flow, or /flow text for a simpler version) showing each agent’s row, its current state, which tool it’s using, and elapsed time, so you can expand a row and watch what a specific subagent is doing.
Agent Flow doesn’t create agents or make the underlying work run faster. It’s a visibility layer. The value scales with how often you actually delegate substantial, independent work across multiple agents. For a single short conversation, File Tree and Replay Theater are likely more useful day to day.
What does Reflect do with repeated corrections?
Reflect is meant for the annoyance of repeating the same correction to Claude over and over, whether that’s a package manager preference, a testing convention, or where certain code should live. It converts a correction into a proposed instruction that you review before saving.
There are two different plugins with similar names in circulation, and they behave differently, so it matters which one you install: the “reflect mod” plugin (not the older Claude Reflect plugin) is the one that implements this review-and-save workflow. Using an explicit remember: prefix sends a candidate instruction straight to a review interface, skipping any separate detection step. You choose a scope (such as repository-wide), edit the proposed text if needed, and save it, after which it’s written into your project’s Claude configuration file for future sessions to pick up automatically.
Frequently Asked Questions
Do Claude Code mods cost extra in API usage?
The mods themselves don’t make additional model calls just to render their panels or reports. File Tree, Replay Theater, Blast Radius, Collision Guard, and Agent Flow all rely on local tracking or passive display. The underlying coding tasks and delegated agents they’re observing still consume your normal Claude usage.
Do I need a special Claude Code version to use these mods?
Yes. The walkthroughs were done on Claude Code version 2.1.287 or newer, where mods are enabled by default. Older setups may require manual update via claude update, and you shouldn’t need legacy experimental environment variables on current versions.
Can Blast Radius or Collision Guard fully prevent mistakes?
No. Blast Radius only recognizes certain bash command patterns and can miss scripts, aliases, or chained commands. Collision Guard only tracks edits made through Claude’s own tools on the same machine, so edits from other tools or machines are invisible to it. Both are meant as an additional safety layer, not a replacement for normal permissions and attention.
What’s the difference between Reflect mod and Claude Reflect?
They’re two separate plugins from the same repository area, and only one, reflect mod, implements the review-and-save workflow for turning corrections into saved instructions. Running both for the same capture workflow isn’t recommended.
Which mod should I try first?
For a single developer working mostly in the terminal, File Tree and Replay Theater address the most common daily friction: finding what Claude is working on and understanding what it changed. Blast Radius is worth adding if you ever let an agent run destructive bash commands unattended. Collision Guard and Agent Flow become relevant once you’re running multiple parallel Claude sessions.
