Kobel
БесплатноНе проверенDocumentation for Kobel — a desktop permission gateway between AI assistants and your computer (Windows/macOS, MCP)
Описание
Documentation for Kobel — a desktop permission gateway between AI assistants and your computer (Windows/macOS, MCP)
README
Kobel is a desktop application for Windows and macOS that runs a local Model Context Protocol server. It sits between your AI assistant and your machine, and it decides — per file — what the assistant is allowed to do.
Everything runs locally on your computer. There is no Kobel cloud, and files are not uploaded to one.
Website: https://kobel.app/en/
About this repository. Kobel is a commercial, closed-source desktop app. This repository holds its public documentation: the permission model, the MCP tools it exposes, and the connection guides for each AI client — so that they can be read, linked and checked by anyone without installing anything.
The problem
Giving an AI assistant access to your files is usually all-or-nothing. You either paste content into a chat window by hand, or you hand a filesystem MCP server one or more directory paths and hope it stays inside them. There is no middle ground where this spreadsheet is read-only, that project folder is editable, and the folder next to it does not exist as far as the assistant is concerned.
Kobel is that middle ground, with a graphical interface instead of a JSON config
file — aimed at people who use AI daily but do not want to edit mcp.json.
The five permission levels
Every file carries exactly one level:
| Level | What the AI may do |
|---|---|
| Edit | The AI may edit this file directly. |
| Edit + backup | The AI may edit; a backup is created automatically before every change. |
| Edit copy | The AI works on a copy, the original stays safe. |
| Read only | The AI can see the file but cannot change it. |
| Blocked | The AI cannot see this file. |
Two rules make this predictable in practice:
- Levels apply per file. A whole folder can be set at once, and everything placed inside it inherits that level — including files added later. A file never ends up unprotected because it arrived after the rule was made.
- A folder always shows what is actually inside it. If its files carry different levels, it is shown as mixed — never as a single colour its contents do not have. What you see is the real state, not an inherited label.
Blocked files are not filtered out of an answer after the fact. They are not visible to the tool layer at all: a listing does not contain them, and a read request for one fails.
Details: docs/PERMISSIONS.md
What the assistant can actually call
Kobel exposes a fixed set of MCP tools. Every tool is prefixed kobel_, every
description starts with [Kobel], and every title reads Kobel · <name> — so
that in an assistant's approval dialog it is always visible which application is
asking.
Files — kobel_list_root_folders, kobel_list_files, kobel_read_file,
kobel_read_file_lines, kobel_read_file_base64_chunk, kobel_get_file_info,
kobel_search_in_files, kobel_diff_files
Writing — kobel_save_file, kobel_save_file_chunked, kobel_write_file,
kobel_patch_text, kobel_prepare_for_editing, kobel_create_directory,
kobel_copy_file, kobel_move_file, kobel_rename_file, kobel_delete_file
Images — kobel_bild_info, kobel_bild_skalieren, kobel_bild_zuschneiden,
kobel_bild_ecken_abrunden, kobel_bild_konvertieren
Other — kobel_list_projects, kobel_store_knowledge, kobel_hilfe
There is deliberately no general shell or script execution tool. Narrow, named operations were chosen over a generic "run this" escape hatch.
Beyond the file tools, connected accounts and third-party MCP servers appear as additional namespaced tools — see App proxy.
Full list with parameters: docs/TOOLS.md
Supported clients
Kobel speaks plain MCP, so any MCP client works. Guided setup exists for:
- Claude Desktop — local stdio connection
- ChatGPT (the normal chat window) — over an HTTPS tunnel, because OpenAI connects to MCP servers from its own cloud, not from your device
- OpenAI Codex
- Cursor
- GitHub Copilot / VS Code
- Gemini CLI
- Ollama and other local runtimes
- Templates for Qwen Code, Kimi, CodeBuddy, Trae, Cherry Studio, Chatbox, Roo Code, Warp and others
Step-by-step per client: docs/CONNECT.md
Activity log
Every tool call is written to a local log with a timestamp: which assistant, which tool, which file, and what came back. Permissions can be withdrawn while a session is running; the change takes effect on the next call.
App proxy and connected accounts
Kobel is not limited to files. The same permission layer can be put in front of other things you want an assistant to reach:
- Third-party MCP servers, proxied through Kobel so they inherit its approval and logging behaviour
- Accounts connected over OAuth — LinkedIn, Instagram, Facebook, Telegram, TikTok, e-mail
- Cloud storage — Google Drive, Dropbox, OneDrive
Credentials are stored encrypted on the machine.
Teams
An administrator can publish a policy over a URL or a UNC path. Clients read it and apply it, which allows folder rules and blocked paths to be set centrally rather than per workstation.
Prompt injection
The realistic threat with a file-connected assistant is not that the model turns malicious — it is that a document, an e-mail or a web page it reads contains text that instructs it to do something the user never asked for.
Kobel's answer is that the instruction and the permission are separate. Text inside a file cannot raise its own level. A file marked Read only stays read-only however convincingly a document argues otherwise, a Blocked file is not reachable by the tool layer, and anything the assistant did attempt is in the log.
Longer write-up: docs/SECURITY.md
Install
- Windows — Microsoft Store or a direct signed installer from https://kobel.app
- macOS — Mac App Store
Built with Tauri v2 (Rust core, web front end). The interface is available in 15 languages.
Pricing
| Basis | Free — 1 AI, 2 folders / 50 files, 500 MB |
| Pro | €49 one-time |
| Pro | €3.90 / month |
| Team | Several sizes, see the website |
Security contact
Vulnerability reports: see https://kobel.app/.well-known/security.txt (PGP key published, ed25519). A CycloneDX SBOM is maintained for both the Rust and the JavaScript components.
Questions
Common questions — how to give ChatGPT access to local files, whether it is safe to let an assistant read your disk, how this differs from pointing a filesystem MCP server at a folder — are answered in docs/FAQ.md.
Links
- Website — https://kobel.app/en/
- Help — https://kobel.app/hilfe.html
- Connecting ChatGPT to local files — https://kobel.app/chatgpt-local-files.html
- Security — https://kobel.app/security.html
License
The documentation in this repository is published under CC BY 4.0. The Kobel application itself is proprietary.
Установка Kobel
У этого сервера нет опубликованного пакета — он собирается из исходников. Открой репозиторий и следуй инструкции в README.
▸ github.com/Kobel123/kobel-mcpFAQ
Kobel MCP бесплатный?
Да, Kobel MCP бесплатный — установка в пару кликов через Unyly без оплаты.
Нужен ли API-ключ для Kobel?
Нет, Kobel работает без API-ключей и переменных окружения.
Kobel — hosted или self-hosted?
Self-hosted: сервер запускается локально на твоей машине командой из раздела установки.
Как установить Kobel в Claude Desktop, Claude Code или Cursor?
Открой Kobel на unyly.org, выбери вкладку своего клиента (Claude Desktop, Claude Code, Cursor) и нажми Install — конфиг сгенерируется автоматически, без правки JSON.
Похожие MCP
Fetch
Web content fetching and conversion for efficient LLM usage.
Roblox Studio
Enables AI coding tools to control Roblox Studio for workspace exploration, instance manipulation, and script management. It provides tools for playtesting, sce
автор: paralovAWS KB Retrieval
Retrieval from AWS Knowledge Base using Bedrock Agent Runtime.
автор: modelcontextprotocolSpring AI MCP Server
Provides auto-configuration for setting up an MCP server in Spring Boot applications.
llm-analysis-assistant
A very streamlined mcp client that supports calling and monitoring stdio/sse/streamableHttp, and can also view request responses through the /logs page. It also
автор: xuzexin-hzMCP-Agent
A simple, composable framework to build agents using Model Context Protocol by [LastMile AI](https://www.lastmileai.dev)
автор: lastmile-aiSpring AI MCP Client
Provides auto-configuration for MCP client functionality in Spring Boot applications.
mcp.natoma.ai
A Hosted MCP Platform to discover, install, manage and deploy MCP servers by [Natoma Labs](https://www.natoma.ai)
MCPHub
Website to list high quality MCP servers and reviews by real users. Also provide online chatbot for popular LLM models with MCP server support.
MCP Servers Rating and User Reviews
Website to rate MCP servers, write authentic user reviews, and [search engine for agent & mcp](http://www.deepnlp.org/search/agent)
Compare Kobel with
Не уверен что выбрать?
Найди свой стек за 60 секунд
Автор?
Embed-бейдж для README
Похожее
Все в категории ai
