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

    • Changes prices, promotions, ad bids and budgets in live seller accounts
    • Handles financial data and keys for several stores
    • The unauthenticated dashboard lets anyone who can reach the port add stores
    • Changes prices, stock and product cards in the live account with no sandbox
    • Can confirm or cancel order shipment
    • Requires Seller API key with Admin role
    • Tools modify prices, stocks and product cards in a live store
    • In HTTP mode, key is passed in URL
    • With OZON_ALLOW_WRITES enabled the server can change a real shop's prices, stock, promotions and postings
    • By default access is read-only; enabling writes should be a deliberate choice
    • 240 catalog methods change data in the live account
    • 11 methods are marked irreversible, including order and return operations
    • With OZON_ENABLE_ORDERS on it can spend money and place orders
    • With OZON_ENABLE_WRITES on it changes the cart, favorites, lists and cancels orders
    • The MCP endpoint does not check who connects to it
    • Changes real product stock in the Ozon seller account
    • By default sets a fixed 9999 stock without checking actual availability
  • PAI-ProWorkflow
    • Paid generations spend the balance of your PAI Pro platform account
    • The app runs an unattended agent, and Docker mode keeps it inside a container
    • Setup scripts put skills into the agent's directory, and the app opens a local web interface
    • Enabling TBANK_ALLOW_TRADING allows the server to place and cancel orders on a real account
    • Universal calls provide access to any T-Invest API method, not only a safe subset
    • The panel has a built-in terminal and session control, and a zrok tunnel opens them beyond the local network
    • The macOS installers are not notarized, and installation sets up pi and dependencies in the home directory automatically
    • The API key lets the agent write to the tracker, including deleting projects and tasks
    • A role with broad rights opens every project on the instance to the agent
    • Sends and deletes messages on behalf of a user or bot
    • Can manage workspace users, chats and webhooks
  • Pachca MCPMCP server
    • Sends messages as the user
    • Reads all work conversations
    • Stores the account session on disk
    • Starts coding agents in panels, sends them text and keystrokes and reads their terminal screens
    • Wider toolsets add commit, push, rebase, stash and other writing git operations
    • At launch the app writes its own pane server into the Claude Code, Codex and Cursor configs
    • The installer downloads and runs a full suite with its own Python and a local web server, and the skill updates it itself on every call
    • The skill reads the folder with research materials, which can hold private data and unpublished results
    • The agent writes the paper files and the check scripts run on your machine
    • Works with a company's financial and tax data and keeps the Qonto and Stripe keys in environment variables
    • Runs Node scripts and generates official documents: FEC, tax bundles, invoices
    • Launches agents that change code and run commands unattended
    • Stores agent keys and issues them access tokens
  • PayCLI
    • Signs stablecoin payments from a local wallet, which means spending real money, although each signature waits for approval
    • Keeps wallet keys in the system store and works with external 1Password and Openfort wallets
    • Calls third-party paid APIs, and provider responses may carry harmful instructions the agent must not follow
    • Processes order payments and refunds
    • Sends invoices to customers and changes subscriptions
  • PeekabooMCP server
    • Sees the whole screen and controls input in any app
    • Requires system Screen Recording and Accessibility permissions
    • Subagents actively test targets with exploits and payloads, and the user sets scope themselves with no authorization check
    • The hook gets full shell access for recon and hunt commands, limited only by the scope.yaml list
    • The bounty-platforms MCP server holds HackerOne, Bugcrowd, Intigriti and YesWeHack tokens and can submit reports on your behalf
  • Personal ModelMCP server
    • Continuously reads what is on the Mac screen: interface text, window titles and, when enabled, screenshots with text recognition
    • Passes derived personal context to the chosen language model provider or through a Codex, Claude Code or Cursor Agent subscription
    • Any connected MCP client gets access to a model with personal data, and the author asks you to connect only trusted clients
    • Sends requests and selected material to the external Pexo service over the network
    • Operates with a paid API key and spends credits on generation
    • Runs bundled shell scripts on the user's machine
    • The photoshop_execute_jsx tool runs any code inside your Photoshop
    • An agent creates and saves documents, so it can overwrite working files
  • PiCLI
    • Runs arbitrary bash commands without built-in restrictions
    • Has access to all of the user's files and credentials
    • In BUILD mode the agent changes source files, builds artifacts and can load the configuration into a test infobase via /update1cbase
    • The optional memory and browser MCPs send project data to external or local services once separately enabled
    • Any participant with steering rights makes the agent run commands on your machine as you
    • The cloudflare mode opens the server to anyone with the link, and the sign-in link works as the owner's key
    • The built-in browser and schedules let the agent reach the network and work while you are not watching
    • The app copies skills into agent directories and runs scripts, ffmpeg and yt-dlp
    • The agent downloads third-party content from the internet, and it may contain instructions that must not be followed
    • Deploys and can delete server-side resources: the Worker, domain binding and D1 database
    • Publishes files to IPFS, where content becomes public and effectively permanent
    • Requires account credentials or an app key
    • Creates indexes and writes data to an external service
    • Uses a secret API key
    • The installer downloads and runs a script
    • In autonomous mode it runs commands and edits files
    • Code is sent to model providers
    • The API key lets the agent create, change and delete projects and tasks in the workspace
    • Requests to Plane are logged together with their content until you set LOG_PAYLOADS=false
    • Changes schema and data in production databases
    • Handles passwords and service tokens
  • Planfix MCPMCP server
    • Creates and edits tasks and contacts in a live account
    • Requires an API key with access to company data
    • The agent generates and executes arbitrary Node.js code
    • Setup downloads Playwright and Chromium binaries
  • PlaywriterMCP server
    • Gives the agent access to your browser with real logins and sessions on every site
    • Can perform arbitrary actions on websites on your behalf
    • Plugins send emails and messages, edit the calendar and upload videos to YouTube
    • They need access to external accounts and keys, and some features act on your behalf
    • The API key spends a paid balance, and a new key has no limits until you set them in the console
    • Prompts, photos, video and audio go to the PopNow server, and a local file is uploaded automatically when passed as an argument
    • The assembly skills run Python and Node.js scripts, Chrome and FFmpeg on your machine and write files into the project
    • A local service becomes reachable from the internet at a public address: publish only what you are ready to show
    • Traffic goes through someone else's relay, and a relay holding the certificate key could in theory impersonate your endpoint
    • Installing with a script piped from the internet through curl and bash runs downloaded code
    • The write tool runs DDL and DML with the rights of the profile's role, and the server itself does not check what the agent may do
    • Loading and describing CSV reads local files from approved folders
    • Works with a database and can change data in unrestricted mode
    • Keeps database credentials in the config
  • Postmypost MCPMCP server
    • The server publishes and deletes posts in connected social accounts on behalf of your account
    • delete_publication can remove a publication everywhere, including the social network
    • The token gives access to all projects and accounts within the account permissions
    • The agent creates, changes and deletes semantic model objects with your permissions, including in a Fabric workspace
    • Model data and metadata reach the client and may go on to the LLM provider, and access tokens are sent to allowed hosts
    • The agent edits and saves open presentations, and powerpoint_evaluate executes arbitrary Python on your computer
    • The server reads slide content, notes and comments of open files
    • Resets databases, applies migrations and runs SQL
    • The remote server can delete databases and connection strings
    • Bulk-updates product prices and stock in the live account
    • Stores credentials and a log of every operation in its own PostgreSQL database
    • The plugin reads your resume, work history and LinkedIn contact list and keeps them in the ~/.proficiently/ folder
    • Through the Chrome browser it fills out application forms on employer sites on your behalf
    • The Telegram bot token is written to a file in plain text, and the bot takes commands from a chat
    • Writes, overwrites and deletes files, including directories, from the model's reply and without asking
    • Sends the contents of the listed files and of files whose paths appear in the task text to OpenAI
    • Uses the text fetched from an http or https address as the task when the task is passed as a link
    • The agent changes a cloud account through Terraform, command-line utilities or the web console, and some changes cannot be rolled back
    • The Prowler API key goes to a remote MCP server, and data about your account and its findings passes through it
    • The report in the `.prowler/` folder holds account and finding identifiers and is easy to commit by mistake
  • ProxmoxMCP-PlusMCP server
    • Manages live virtualization: VMs, LXC, snapshots, backups
    • Requires a Proxmox API token with mutating permissions
    • Native MCP over HTTP can be exposed without auth if misconfigured

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.