Searches public GitHub repositories to discover and add plugins or skills from within the app.
Claude Code, Codex and Cursor each keep their own settings. Kitden puts their plugins, skills, MCP servers and instruction files on one map, so what one agent has and another lacks shows at a glance. Add what an agent is missing, turn an item off or remove it, on the same page.
Search the official MCP Registry, skills.sh, the Claude plugins marketplace and any public GitHub repository without leaving the page. Tick the agents a plugin or skill goes to. An MCP server goes to the gateway, and says first whether it needs a sign-in or a token.
Move a server into the Kitden gateway and every agent you use gets it: set it up once, not once per agent. Every tool has a switch, for all projects or one, and a tool that is off is refused, not just hidden. A token or sign-in is given once and checked before Apply.
Nothing is written until you press Apply. Every change shows a preview first, with the exact edit under Show changes.
Apply keeps a backup, and Undo puts it back. Changes keeps the history of what you did. If a file changed since, Kitden stops instead of overwriting it.
Everything stays on your computer. The page runs at 127.0.0.1 behind a private link. No account, no telemetry. Kitden goes online only for the catalogs you open and the remote servers you add.
Secrets live in the macOS Keychain or your Linux keyring only. The page shows no more than their first and last characters.
Kitden's agent tools only read. Asking an agent to show your setup changes nothing.
Add only servers and plugins you trust: they run with your permissions.
Feature | What it does |
Per project | Turn a server or a tool on or off for one project only. |
See what runs | How often each tool is called, and which tools only read. |
Instructions in view | Which CLAUDE.md, AGENTS.md, rules, hooks, commands and agents each client loads, and where they come from. |
Ask your agent | "Show my setup" works from any agent, through Kitden's own read-only MCP server. |
Claude Code plugin | With Kitden installed, |
Related MCP server: whats-inherited-mcp
Needs macOS or Linux, uv and at
least one of Claude Code, Codex or Cursor. On Linux, install libsecret-tools (Debian,
Ubuntu) or libsecret (Fedora, Arch) so Kitden can keep server secrets in your keyring.
Try it on a sample first. Your own settings are not read or changed:
uvx --python 3.13 kitden demoInstall it and open it for your own setup:
uv tool install --python 3.13 kitden && kitdenOr with Homebrew: brew install konstantinstarovoytov/tap/kitden, then kitden.
Update later with uv tool upgrade kitden or brew upgrade kitden.
Guide · Roadmap · Releases · Issues · Contributing · Apache-2.0
Explain one name: every definition across clients and scopes, which one loads in the project, and why. Use it after show_setup or list_findings names an item; show_setup lists everything. kind is optional (mcp_server, skill, plugin, hook, marketplace, instructions). Read-only: this tool cannot change settings. Do not edit client settings files to work around this; tell the developer to make the change on the local page (run kitden open), which previews, backs up and can undo it.
Inputs: name (required), kind, project_path. Returns: object
List problems for a project: duplicates, items loaded from another client's folder, the same name from another source, overrides, unknown states, untested versions, and MCP servers that bypass the gateway. Each names its clients. Use it to check a setup for problems; show_setup lists everything that loads, and explain_item follows one name. Read-only: this tool cannot change settings. Do not edit client settings files to work around this; tell the developer to make the change on the local page (run kitden open), which previews, backs up and can undo it.
Inputs: project_path. Returns: object
List the projects the clients know, which clients know each one, and whether the folder still exists. Use it to find a project_path for show_setup, list_findings and explain_item. Read-only: this tool cannot change settings. Do not edit client settings files to work around this; tell the developer to make the change on the local page (run kitden open), which previews, backs up and can undo it.
Inputs: no named parameters. Returns: object
Show what each AI coding client loads for a project: MCP servers, skills, plugins, hooks, marketplaces and instruction files, with source, scope and on/off state. Start here for the full picture; use list_findings for problems only and explain_item for one name. client (claude_code, claude_desktop_code_tab, codex, cursor) defaults to all. Read-only: this tool cannot change settings. Do not edit client settings files to work around this; tell the developer to make the change on the local page (run kitden open), which previews, backs up and can undo it.
Inputs: client, project_path. Returns: object
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 4 tools
Each tool targets a distinct granularity: list_projects finds a path, show_setup gives the full inventory, list_findings filters to problems only, and explain_item drills into one name. Descriptions explicitly cross-reference one another ("Start here... use list_findings for problems only and explain_item for one name"), leaving no realistic chance of misselection.
All four names follow a clean snake_case verb_noun pattern (show_setup, list_findings, explain_item, list_projects) with semantically motivated verb choices. No camelCase mixing or vague bare verbs.
Four tools is well-scoped for a read-only setup-inspection server; each covers a necessary step of the discover-inspect-explain workflow with no filler. The intentionally read-only surface is enforced consistently across all tools.
The read-only lifecycle is fully covered: find projects, list everything loaded, surface problems, and explain a single item, with mutation deliberately delegated to the local page. Minor gaps exist (no export/summary or way to inspect the gateway itself beyond the bypass finding), but agents can work around them.
Read-only catalog of prompt packs and rule files: Cursor, CLAUDE.md, AGENTS.md, SKILL.md.
Path-scoped team memories, rules and skills for Claude Code, Cursor, Codex and other MCP clients.
Scan AI agent skills, MCP server packages and MCP configs before you install them. Static analysis.
Free, source-checked directory of AI agent tools, MCP servers and skills, with read-only search.
