Access levels of skills and MCP servers

The access level of every catalog entry and the reasons behind it. The level shows what a tool adds to what the agent can already do. It does not rate code quality.

Broad access

    • create_encashment writes a cash movement entry with no separate confirmation
    • raw_request grants access to any iikoServer endpoint, including potentially data-changing ones
    • Reads private correspondence and local messenger databases, handling sensitive data
    • Uses platform API tokens and runs Python collectors that access the network
  • iphone-useMCP server
    • The agent taps and types in any phone app, including banking apps and messengers
    • Through the screen the agent reads personal data such as messages and Health records
    • The source is under PolyForm Shield: commercial use is limited by the author's terms
  • iterm-mcpMCP server
    • Runs arbitrary commands in your terminal
    • No built-in command safety checks
    • Can start and interrupt any process
    • Handles personal tax documents
    • Runs extraction and decryption scripts
    • Each skill's instructions are fetched from npm at run time, and the agent itself runs the commands
    • Some skills sign in to accounts (bookings, an employer portal, tax invoices) and take secrets from a password manager or a file
    • Requests go through the developer's shared proxy, and a usage counter goes to PostHog
    • The CLI covers cards, automations and other write operations of the Kaiten API
    • Stores the API token in a config file on disk
    • Creates, edits and archives cards in the team's live tracker
    • Stores the Kaiten API token in a config file
    • Uploaded files get a public URL
    • The kaiten-cli-mutations skill performs write operations on Kaiten cards and data
    • The CLI runs under the user credentials set in its configuration
  • kaiten-mcpMCP server
    • Tools create, move, archive and delete cards in a live Kaiten workspace
    • The remote mode encrypts and stores teammates' personal tokens on a third-party Vercel deployment
    • Tools cover write operations across nearly the whole Kaiten API, including roles, groups and automations
    • A shared HTTP server with OAuth grants access to multiple users of one instance, so the config needs careful review
    • The agent receives the router admin password and can block devices, bring interfaces down and create networks
    • A careless interface shutdown or segment deletion can cut off your own access to the router
  • kiwi-memMCP server
    • Stores the content of your conversations, personal details included, in a database on your server
    • There is no built-in authentication, and a backup can contain API keys
    • Passes all requests to the model provider through its own gateway
  • kmcpCLI
    • The controller manages cluster resources
    • The install script runs through the shell
  • kmsgCLI
    • Reads your personal and work chats and sends messages and images in your name
    • Requires macOS Accessibility access, meaning control of other apps' windows
    • Automation may lead to an account restriction under the service's rules, and the author places responsibility on the user
  • kommo-mcpMCP server
    • The token grants Kommo account administrator rights
    • Tools create and bulk-update leads and can trigger a salesbot on a lead
    • Runs kubectl and bash commands against the cluster
    • Can modify resources, production included
    • Runs real kubectl operations against the cluster, including mutating ones
    • Its scope equals the kubeconfig permissions
    • An agent mistake can affect a live cluster
    • Can extract cookies from Chromium-family browsers to access accounts
    • Runs Python scripts and installs additional CLIs
    • Uses third-party API keys
    • Keeps login cookies for Weibo, Zhihu, Bilibili and Douyin and API keys on disk and then uses them in requests
    • Reaches eight platforms and the paid services TikHub, jisuapi and Baidu Qianfan
    • Runs Python scripts and opens a Chromium browser during login
    • `ljg-blind` reads yesterday's AI conversations from `~/.codex/sessions`, which may hold personal and work data
    • `ljg-push` runs `git push` over ssh to the author's repository and switches branches of the local clone
    • `ljg-card` installs dependencies and downloads Chromium, and the skills create files and run Bun scripts
  • llmwikiPlugin
    • The tool reads the full session histories of all your agents in the home directory, and they can hold keys, passwords and personal data
    • The installer can add a SessionStart hook to ~/.claude/settings.json, and the built site can be published to GitHub Pages
  • lsbotMCP server
    • The agent gets shell command execution, file writes and batch deletion and process termination on your machine
    • Scheduled tasks call tools without your involvement, and the install runs a script from a third-party domain
    • Writes agent instruction files into the project: AGENTS.md, Cursor rules, skills
    • The secrets vault injects macOS keychain values into commands you approved
    • The .mancode folder accumulates plans and verification results and may capture sensitive context, so review it before committing
    • Can change prices, stock and orders in seller accounts
    • Stores account API keys locally
    • Methods outside the verified core are imported from specs and not hand-checked
  • mcp-1cMCP server
    • The --install command adds an extension and updates the base configuration
    • The agent reads live data and the event log
    • exec runs arbitrary BSL code with no restriction at the extension level
    • The author explicitly warns not to publish this on a production base
  • mcp-bitrix24MCP server
    • Creates, closes and reopens tasks in a live Bitrix24 portal
    • The webhook carries whatever rights were set when it was created; issue a separate webhook scoped to just the needed methods
    • Reads private mail and attachments, and the mailbox password is stored on your machine
    • Can send and forward mail in your name, as well as delete and move messages
    • The email_command tool runs arbitrary IMAP and SMTP commands when the owner enables it for an account
    • Writes and changes data in your Google Sheets and Google Drive
    • Can grant spreadsheet access to other users by email
    • Handles service account keys or OAuth credentials
  • mcp-installerMCP server
    • Installs and runs third-party packages from npm and PyPI
    • Edits the client config and launches arbitrary MCP servers
  • mcp-klingMCP server
    • Local files go to a third-party Supabase storage whose address is hardcoded in the server, and get a public link
    • Kling API keys give access to your account's resources
    • Results are downloaded to disk by default
    • The agent can create, change and delete pages and blocks of your graph unless a write namespace list is set
    • Notes are read in full, and with hosted embeddings their text is sent to a third-party provider
  • mcp-maigretMCP server
    • The server launches Docker containers on your machine at the agent's command
    • Requests to many third-party sites come from your IP, and some sites may rate-limit or block it
    • Reports collect information about people's accounts, so a lawful basis and the account owner's consent matter
    • The server clicks and types in any app, so the agent can send an email, delete a file or press a pay button
    • Accessibility and screen recording permissions allow reading window contents, including passwords on screen
    • Every response saves a screenshot and the full interface tree to disk
    • WEREAD_COOKIE is your WeRead account session: while the cookie is valid, the server reads the account's data with your rights
    • With CookieCloud the cookies are synced through a server; the documentation names cc.chenge.ink, which you can replace with your own
    • Can change product prices and stock on a live store
    • Can perform actions on orders, returns and shipments
    • The entire API is reachable through the generic ym_execute, including write operations
    • Can delete counters, goals, filters, segments and access grants
    • A token with metrika:write scope grants full control over the account's counters
    • Can upload offline conversions and expenses, affecting attribution data
    • The server writes and deletes messages in your name in private chats, groups and channels
    • The session and API hash open the account's whole correspondence, and downloaded media land on disk
    • The session opens the account's private conversations: the assistant sees every dialog and message the server returns
    • The API hash and sign-in session give access to the account, and a leaked hash cannot be reissued
    • Can call write operations on external APIs, including POST requests and GraphQL mutations
    • Via --mcp-stdio it spawns arbitrary server commands and stores OAuth tokens and secrets in a cache
  • mcpcCLI
    • Calls arbitrary tools on connected MCP servers
    • Stores OAuth tokens and bearer keys
    • Supports experimental x402 crypto payments
  • mcpoCLI
    • Exposes MCP server tools, which may run commands, over the network
    • Unsafe without an API key and network restrictions
    • The sync command writes other agents' configs, and importing servers reads their entries
    • Through call and exec the agent calls any connected server and runs code in a sandbox
    • It stores connections and OAuth data: secrets are saved by reference to an environment variable or a file
    • Creates drafts and uploads images into a WeChat account, that is, writes to an external system
    • Handles WeChat credentials and external service keys
    • Calls a paid API and image generation providers over the network
    • The server reads personal notes with details about you and your relatives and runs memory code on your computer
    • In HTTP mode through a tunnel, ChatGPT connects without authentication: the tunnel address opens your memory to anyone who knows it
    • The GitHub and Google Docs connectors need access tokens that must be handled with care
    • Hooks record every dialog turn, and the content goes to the chosen model provider to build the memory
    • The npx install pulls fresh code straight from GitHub without pinning a version, and the provider key goes into the config as plain text by default
    • A local service runs in the background on 127.0.0.1:3878, and the memory holds working facts about the project
    • Auto-capture reads conversation content and saves it to disk, and entries can include passwords and personal data
    • Entry text is sent to the embedding provider, reranker and extraction model under your key
    • The `delete`, `delete-bulk`, `upgrade` and `migrate` commands remove and rewrite the store
  • mirroir-mcpMCP server
    • The agent controls a real phone: taps, text input, launching and closing apps
    • The screen can show private messages, mail and banking apps, and screenshots go to the model
    • Without a permissions file only 11 non-mutating tools are available, and action rights are opened by hand
    • The MLX Core app's agent mode runs shell commands on the Mac unless the sandbox with an isolated virtual machine is on
    • The server opens port `11234`, and models can be shared across the local network through Bonjour
    • Launchers write agent settings: Codex runs against your own home folder, and the fx entry lands in `~/.fx/settings.json`

How we set the access level

Methodology

The editors set the level and the reasons from the README, the way the tool runs and the permissions it needs. Reading project files the agent can already see does not raise the level. There is no automatic code audit yet.