rusyander/claude-control

**Every Claude Code setting in one place.**
A local web panel for Claude Code: it reads and edits the real files on disk — CLAUDE.md, settings.json, skills, hooks, MCP servers — through a UI instead of by hand in JSON and Markdown. Plus a sandbox, where a change is tested before it becomes your daily setup.
Runs entirely on your machine: no account, no server, no telemetry.
By default the panel configures Claude Code, where everything is available; it also edits the configuration of Codex, Gemini, Qwen Code, Continue, Goose, Kimi Code, Cursor, OpenCode and Aider — each in its native format, see CLIs other than Claude.
🇷🇺 Русская версия · 🔧 Setup · 🛠 Troubleshooting · 🏗 How it works · 🔒 Security · 📱 Access from a phone · 💬 Chat and parallel agents · 👥 Groups · 📡 Contour · 🧑‍💻 Development · 🚫 What the panel does not do
> [!TIP]
> **Won't start?** pnpm doctor explains every finding. Still stuck — open Claude Code in this
> folder and say "it won't start, figure it out": it reads AGENTS.md, the map of this
> project written for the agent, and starts with diagnostics rather than with the code.
---
Why
Claude Code's configuration is plain files, which is a good decision while there are few of them. Then comes a thousand-line CLAUDE.md, dozens of skill folders, hooks three levels deep in settings.json, MCP servers in a different file altogether, permissions nobody remembers adding. Three simple questions get hard: **what is actually active right now, what does this hook fire on, what breaks if I add this rule.**
The panel answers them: a visible shape, a switch that deletes nothing, and a sandbox where a change is tested against real Claude Code before it becomes permanent.
What it does
**Configuration**
- **Rules** — the ## ПРАВИЛО: sections of CLAUDE.md, each with search and its own switch; a separate page edits the whole file
- **Skills** — the skills/ folder: file tree, editor, rename, a library of SKILL.md templates
- **Commands** — everything invoked through / in one list: skills, command files, plugins, built-in CLI commands — with a description, an owner and a jump to editing
- **Hooks** — grouped by event, matcher shown explicitly, order within an event rearranged by hand
- **Scripts** — files from hooks/, nested folders included, with a flag on the ones nothing calls
- **MCP servers** — a real protocol-level connection check (stdio, HTTP, SSE), interactive OAuth login, a tool listing, and mcp__server__tool permissions generated from it
- **Permissions** — allow / ask / deny, mcp__* patterns included; entries move between settings.json and settings.local.json
- **Environment** — from settings.json and .mcp-secrets.env, secrets masked
- **Plugins** — installed ones and the marketplace catalog, source management, a scaffolder for your own
- **Projects** — the project level of a chosen folder (CLAUDE.md, .mcp.json, .claude/settings.json) on top of the user level; the «From the project» tab shows the project's own set — skills, hooks and rules from its .claude — read-only, it is edited in the project's git
**Working with it**
- **Sandbox** — test a rule, skill, hook or MCP server in isolation; a hook against a prepared or custom JSON event
- **Chat** — a conversation with Claude Code inside the panel: streaming, attachments, voice, artifact preview, model and thinking depth, branching by editing a message, search, export to md/json
- **Several projects at once** — tabs, parallel agents, dev servers launched from the tab: one target per package in a monorepo, each on its own port, autostarted when the panel boots
- **Parallel branches** — one button makes a working copy of the repository for a branch (git worktree) next to the project and opens it as an ordinary tab: three agents on three branches never collide, a copy with a live agent cannot be removed, and the panel never merges anything
- **Task fan-out** — hand the agent a list of independent tasks and it proposes a split as a card: "do them here in turn" or "split into N chats", where every group gets its own branch, its own copy and its own chat
- **Continuing in a clean session** — once a task is closed the agent tidies its working files and offers to carry on in a new conversation: the expensive context stays behind and the new session reads the checkpoint file. By button — or on its own: auto-continue is on out of the box and held by safeties (a failed run is never continued, an unchanged checkpoint stops the chain, at most eight in a row). Other CLIs do it too: they have no session, so a continuation is a new conversation carrying the checkpoint and the original task