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

    • Includes offensive techniques: exploit development, EDR bypass and attack chains
    • Runs external reverse-engineering and network tools on hosts and operates against test targets
    • By default Codex runs with the --dangerously-bypass-approvals-and-sandbox flag: no approvals and no sandbox
    • The stop hook blocks the Claude Code session from finishing and runs a generated script
    • Project code and the diff are passed to an external model, and the command edits ~/.codex/config.toml
  • roasr-mcpMCP server
    • Can pause and resume Yandex Direct ad campaigns
    • Reads spend, KPIs and competitor data through the third-party RoASr service
    • The pg_live key grants access to the whole workspace according to its scopes
    • The before: block and the shell and shell_tools tools run shell commands on your machine
    • The write_file tool and Python functions can change files, and confirmation is turned off by the --safe-yes flag
    • Provider keys sit in the environment or config.yml, and prompt and file text goes to the chosen provider
  • rustunnelMCP server
    • Exposes a local port at a public address: anyone with the link can reach the service
    • The agent itself launches the rustunnel process and picks the port, and the token is stored in the MCP config
    • The agent changes and deletes elements of a live Revit model, and the send_code tool runs arbitrary C# code
    • The installer edits client configs and installs add-ins into Revit with the Windows user's permissions
  • sbis-mcpMCP server
    • It logs in with the login and password of the whole Saby account
    • The agent gets access to legally significant documents and counterparty data
  • screenpipeMCP server
    • Continuously records screen, audio and keystrokes, including potentially sensitive data
    • Needs careful filter setup to avoid capturing passwords and personal data
    • Spends money on paid image and video generation
    • Runs external CLIs and ffmpeg
  • selectel-mcpMCP server
    • Creates paid servers and public IPs that spend account balance
    • Deletes servers, volumes, buckets and objects
    • Deploys open SSH to the whole internet by default
    • Uses cloud account credentials
    • Can manage billable cloud resources
    • Stores a real Yandex account login and password in a .env file for automated sign-in to fetch Wordstat stats
    • Automated requests via a logged-in browser instead of the official API can get the Yandex account blocked
    • The index keeps keys and passwords from your sessions in plaintext, and get_credential hands them to the agent
    • The SessionStart hook runs the index refresh script on every Claude Code start
    • The server reads every Claude Code transcript and the source files from the chosen directories
    • The commands erase and delete devices, uninstall apps and clear their data, and a collection can do the same on several devices at once
    • The tool runs arbitrary SQL queries against app databases, uploads and downloads files in an app sandbox and adds root certificates to a device keychain
    • The tool starts a local server with a web dashboard and is not designed for untrusted networks
    • executes commands on a production server via ssh and changes production configuration
    • configures YooKassa payment acceptance and SMTP email credentials
    • Runs Codex CLI, which can change files and execute commands
    • In full-access mode Codex reaches the network
    • A generated wrapper calls the API with your tokens from environment variables and can send requests that change service data
    • For MCP over stdio the server start command is passed as a string, and wrappers run the package through npx on every call
    • The generator writes files into agent directories and keeps a `lock.json` file in your home folder, and the improvement skill may edit skills listed there
  • snipCLI
    • The PreToolUse hook rewrites the shell commands the agent runs and replaces their output
    • Filters from a project directory can change output, so they must be approved with `snip trust`
    • Installation runs a `curl | sh` script, and savings statistics are stored in SQLite
    • Publishes and replies on social networks on your behalf
    • Drives a logged-in browser with working account sessions
    • Runs coding agents in parallel, and they edit files and create branches, worktrees and commits
    • A local server manages agent execution
    • Agents over MCP get read and write access to the project's issues and specs
  • sxCLI
    • Writes skills, rules, hooks and MCP servers into the settings of several AI clients
    • Installed hooks and MCP entries are run by the client later and need the sx binary
    • The cloud relay exposes the vault through an MCP address for claude.ai and chatgpt.com
    • Accesses a brokerage account and can place real orders
    • Installs via curl | bash
    • Stores T-Invest API tokens in a local file
    • The agent interacts with open browsers and web pages
    • It can perform actions on sites in open sessions
    • It is an agent that runs tasks and commands in your environment
    • It connects model providers and tools with your keys
  • tinvest-mcpMCP server
    • Works with a real brokerage account and can place orders with real money
    • Stores broker access tokens
    • A model error or prompt injection could cause an unwanted trade if trading is enabled
    • Lets you approve agent commands remotely
    • Serves a web terminal with shell access
    • Forwards session excerpts and diffs to a messenger
    • The send command types text straight into another pane's prompt, and the screen it reads may hold secrets
    • `tmux-ide integration install claude` writes hooks into `~/.claude/settings.json`, and `tmux-ide adopt` adds a status bar, a dock and keys to tmux
    • A background daemon watches panes in every session and connects to remote machines over SSH, with a local control socket under ~/.tmux-ide
  • tty7Skill
    • The skill lets an agent send text and keystrokes to any pane, so it can run commands in other sessions
    • The client reads the screen of other panes, including ssh sessions
    • A background server keeps shells and remote connections alive after the window closes
    • Modifies Claude Code executables
    • Some patches loosen permission checks and can run scripts from URLs
    • The tool reads your account cookies from the browser and acts as you: posts, deletes, likes and follows
    • Unofficial access can get an account blocked, especially with frequent requests
    • The tool handles account authorization tokens through the environment and files
    • Commands create and change networks, firewall rules, VPN and device settings on a real controller
    • A controller API key or password opens control of the whole network and can be stored as plain text in the config
    • The install script runs through curl and itself puts the skill into ~`/.agents/skills`
    • The installer writes agents, skills, rules and hooks into `~/.claude`, and hooks fire on every tool call
    • Agent chains in the full profile run commands and change code without pausing at every step
    • Self-learning stores rules from past errors and applies them in other projects
    • The agent edits `package.json`, renames configs to `.cjs`, removes `next` and installs `vinext` and `vite`
    • Publishing with `npx @vinext/cloudflare deploy` builds and ships the app to your Cloudflare account through wrangler
    • Other platforms connect through the Nitro plugin and the `NITRO_PRESET` variable, so the build executes code from npm
  • vk-mcp-ruMCP server
    • Publishes posts, messages customers and changes ad campaigns
    • Some methods are irreversible: deleting products or posts, bans
    • The community token grants access to customer conversations
    • Through a walkerOS account the agent gets the right to create and change tracking projects and deployments
    • The skills steer the agent to bundle scripts and run npx walkeros, that is, to execute code on your machine
    • The raw command sends a request with any HTTP method and body to the Wildberries API using the seller token
    • It can change prices, cards and orders; there is no separate write flag yet
  • wcgwMCP server
    • Unfiltered access to the shell and files of the machine
    • The server restricts nothing itself
  • werfCLI
    • Deploys to Kubernetes and deletes images from the registry
    • Requires cluster and registry credentials
  • x64dbg MCPMCP server
    • Controls the debugger: it launches programs, stops processes and reads memory and registers
    • With rights enabled in config.json, the agent writes to memory and registers and runs debugger scripts
    • The server can be opened to a network, and token checking is off by default
    • Grants full control over the debugged process: memory reads and writes, patches, arbitrary debugger commands
    • The server listens on all interfaces by default and access is guarded only by a token
  • xiaohongshu-mcpMCP server
    • Publishes content and acts on behalf of a real social network account
    • Unofficial browser automation with saved cookies, with a risk of account restrictions or a ban
    • The skill drives a signed-in account in the browser: it fills the publishing page and can post comments in your name
    • The account session is kept in the built-in browser profile, and covers need a Gemini API key
    • The agent reads other people's notes and accounts and saves analyses to local files
    • Can send emails and create calendar events on the user's behalf
    • Writes and deletes files on Yandex Disk through MCP write tools
  • yagames-mcpMCP server
    • Can publish and update games on behalf of the developer account
    • The Session_id and CSRF token grant full console access rather than scoped rights
    • Authentication relies on unofficial cookie access rather than a platform-issued key
  • yahoo-finance2MCP server
    • The data feeds financial decisions and is not investment advice
    • An unofficial source, with no guaranteed accuracy or availability
    • It reaches the external Yahoo Finance service
    • Creates campaigns, ads, bids and Metrika goals in a live ad account
    • Sends the click.ru token and account data through third-party hosted MCP servers
    • Can suspend, resume, archive and delete ad campaigns
    • Can change bids and add keywords, which directly affects budget spend
    • Without the sandbox variable it acts on the live ad account
    • Write tools change budget, bids, campaign status and ad text
    • The direct_rpc escape tool calls any API method with no extra restrictions
    • One OAuth token grants access to the entire ad account across all brands
  • yandex-eda-mcpMCP server
    • Can place and pay for a real order with a real charge when confirm: true is passed
    • Stores the Yandex account login session in a local browser profile
    • Browser automation depends on a third-party site's layout and can break when it changes
  • yandex-iot-mcpMCP server
    • The agent can switch real devices and sockets on and off
    • Running scenarios changes the home state without the user present
    • The token grants full control of the smart home

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.