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

    • Host adapters live in the home directory and apply to all projects at once
    • In Auto mode the agent passes confirmations on its own and changes project files, while deletion, push and release stay without permission
    • Task state, reports and memory live in the project and may contain code and log fragments
  • DevSpaceMCP server
    • Gives the connected client arbitrary command execution on your machine
    • Exposes local folders outward through a tunnel
    • vcexecutecode runs arbitrary 1C code in the base
    • vcexecutequery runs arbitrary queries, including potentially heavy ones
    • The agent writes and runs code and uses a browser
    • Stores keys for many services, including Netlify for deploys
  • DevoCLI
    • OS sandboxing and the permission picker are disabled: the agent runs commands and writes files with the user's full rights
    • Project code and context go to the chosen model provider under your key
  • DevopnessMCP server
    • The agent creates and changes servers, networks and apps in your cloud account, and those resources are billed
    • The agent ships changes straight to production servers
    • Permissions follow your Devopness role: a broad role gives the agent broad access to the environment
    • Can sign and send documents on behalf of the organization
    • 7 methods are irreversible, such as the final dispatch of a document to a counterparty
    • Commands read and write organization data: messages, mail, calendar, docs and approvals
    • Sign-in gives the agent access to a DingTalk account and tokens, and the --yes flag removes confirmation prompts
    • The install script downloads and runs a binary, and skills are written into agent directories
    • Can pause campaigns and change budgets, bids and negative keywords in a live account
    • The Yandex token and account data are stored on a third-party commercial service
    • It creates, stops and deletes campaigns, ads and keywords directly in the live Direct account
    • metrika_delete_counter deletes a Metrika counter with its entire data history, with no way back
    • Start with YANDEX_USE_SANDBOX=true and a minimally scoped token until you trust the agent's behavior
    • Holds OAuth access to connected MCP servers and proxies the agent's calls through its gateway
    • The connect command rewrites configuration of Claude Code, Claude Desktop, Cursor and VS Code
    • Starts MCP servers through npx and uvx, and the install script sets up Node, npm and uvx
  • Discord MCPMCP server
    • Sends, edits and deletes messages in channels and DMs
    • Kicks, bans and times out server members
    • Creates and deletes channels, roles and permissions via a bot token
    • The agent posts messages, changes roles and channels, bans members and deletes data on the server as the bot
    • The bot token opens control of the server and must be kept in an environment variable, and the bot's reach should be limited with allowlists
    • Runs third-party MCP servers in containers
    • Stores secrets and OAuth tokens for connected services
    • The agent runs any images and Compose stacks with published ports on your Docker, and access to the Docker daemon is effectively control of the machine
    • Deploying a stack deletes the volumes of a project with the same name along with the data in them
    • Container log output goes to the agent's model and may contain application secrets
    • Full Docker daemon access, which effectively means host control
    • Can remove containers, images and volumes
    • Can connect to remote hosts over SSH
    • In full mode an agent runs commands inside machines and creates sandboxes, and with full network access it reaches the internet
    • A sandbox can be granted host secrets and the SSH agent, which then sit in the environment where the agent works
    • The app installs a network helper, a daemon and virtual machines on the system and keeps container data on the chosen drive
    • The agent drives a real browser: cookies, storage and file uploads are in reach and forms can be submitted from your session
    • The package combines tools for request headers, user agent override and passing checks like Turnstile: use them only on your own or permitted sites
  • DryforgePlugin
    • The `go` command writes code, creates branches and runs checks in your repository
    • After your approval the `migration` command rewrites CLAUDE.md and AGENTS.md, backing them up first
  • DuckleMCP server
    • Runs pipelines that read and write databases, files and cloud storage through saved connections
    • Manages saved connections that may hold access to production systems
    • Downloads and runs the DuckDB engine and extensions on first launch
    • With the flags enabled, the agent can place and cancel trading orders
    • Needs an API key with account access; start with demo and read-only mode
  • ECCPlugin
    • Installs hooks that run on agent events
    • Modifies Claude Code and MCP configuration
    • Runs an npm package with installer code
    • The mutation-applying tools change project forms, templates and extensions
    • Deleting objects is guarded by a separate confirmation token, but the operation is irreversible without a backup
  • EDT MCPPlugin
    • Can change configuration metadata and code and update the infobase
    • Reads infobase data including the event log
  • EDT MCP ServerMCP server
    • Changes project code and metadata and can delete objects and projects
    • Updates the infobase and launches 1C applications
  • EDT-BridgeMCP server
    • Changes project metadata and code through the EDT engine
    • Updates infobase configurations and terminates cluster sessions
    • The debugger can run arbitrary BSL against a live base
    • The hook sees every Claude Code tool call and may delay or reject it
    • In EMILIA mode details of money-related and external actions go to the emiliaprotocol.ai server, and the EP_API_KEY key gives access to your organization
    • The manifest attaches an MCP server that starts through npx -y and receives EP_API_KEY in its environment
    • The agent signs in to social accounts and posts on your behalf, which may violate platform rules
    • Skills run local scripts, a browser and FFmpeg, and model keys and sign-ins are stored on the machine
    • Creates real products, payment links, invoices and payouts through the third-party EasyPay payment service
    • The partner API key carries money-moving rights; use a limited personal employee key rather than the shared partner-wide one
    • Skills run queries and changes on clusters and Elastic Cloud with API keys
    • Security skills process alert data that may contain attacker-controlled content
    • Deploys builds, restarts and deletes apps in the 1C:Enterprise.Element cloud
    • Stores a Console API secret
  • Elementor MCPMCP server
    • The agent creates and changes pages of a production site
    • Some features are paid in the Pro version
    • A third-party plugin and not from Elementor
    • It can send, forward and delete email on the user's behalf
    • Mail credentials are stored in a Docker volume; server access equals mailbox access
    • The agent reads mail and attachments, sends email in your name and moves messages between folders
    • Mailbox passwords sit in plain text in the .env file or the client config
    • The agent flashes and debugs real hardware
    • Version 0.1.0 with some skills untested on hardware
    • The agent edits files and runs shell commands, tests and a linter in your project
    • Provider keys are stored locally and model requests go to external APIs
    • The installer is downloaded and run as a script from the network
  • EmuloPlugin
    • The plugin reads full session logs, which may hold secrets, client names and third-party code
    • With mining on a hosted model, selected fragments of your messages go to the provider
    • A profile with verbatim quotes installed in AGENTS.md easily ends up in a shared repository
    • After approval it can create campaigns and change budgets, bids, strategies and goals
    • Works with archiving, deletion and creative publication in a live account
  • EvoPlugin
    • Runs experiments that execute code
    • Can operate in remote sandboxes via third-party providers
    • The write tools change the content and formatting of Excel files at the given absolute path
    • The server reads any workbook it can reach, so point the agent only at the files you need
  • ExcelMcpMCP server
    • Drives the real Excel application and changes workbook files
    • Runs VBA and Python in Excel, that is arbitrary code
  • ExpectSkill
    • By default the tool extracts your browser cookies and opens pages under your accounts; the --no-cookies flag turns that off
    • The skill launches subagents and drives a browser: it clicks, fills forms and runs code on pages
    • Installation uses an init script that the agent downloads from a website
    • Can place and cancel real exchange orders immediately on a tool call
    • The API token grants access to the user's account, trades and transactions
  • FIRERPA (lamda)MCP server
    • Full device control, including input and remote desktop
    • Traffic capture and reverse engineering require permission to analyze apps
    • It can read local app data on the device
  • FXHoudini MCPMCP server
    • The execute_python tool runs arbitrary code inside the Houdini session, with no authentication or per-tool permissions
    • The agent can delete nodes, change the scene and write files and caches to disk
    • Runs a headless browser that scrapes Facebook pages, and the service's terms may restrict automated collection
    • A script on your machine goes to the network, and other advertisers' ad texts enter the agent's context
    • With frequent requests the results go empty for a while
    • The server reads the cookies of your Chrome with an open Facebook session
    • The author warns about violating Facebook's terms and about a CAPTCHA or an account flag
    • Writes and updates project files and runs scripts in the open Godot editor
    • Installs an addon with a native library and a local daemon that keep the agent connected to the editor
    • The built-in chat sends requests to cloud providers with your keys
    • Creates, edits and deletes variables, components and styles in Figma files
    • The Figma token opens your files
    • Writing needs a plugin in Figma
    • Places orders and takes part in payment
    • Sends requests to the external Filtmall service

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.