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
- ComPDF MCPMCP server
- Whole files go to an external cloud API, and contracts and invoices can hold personal and financial data
- The tools encrypt, remove passwords and watermarks, merge and split files, and the result is written to disk
- Comfy MCPMCP server
- The tools run workflows with arbitrary nodes, manage the ComfyUI process and write files to disk
- Partner models and paid nodes spend Comfy credits on your account
- CompartmentMCP server
- By its instructions the /compartmentalize skill saves passwords, API keys, tokens and contacts from a conversation into the vault
- Installation registers the server in agent configurations, adds a hook to Claude Code settings and a block to CLAUDE.md, and a copy of the settings is kept
- The write tools change the memory, and `forget --shred` destroys a record's key with no way back
- ComposioCLI
- The agent can act inside connected external services
- Account tokens are stored in a third-party cloud
- Compound EngineeringPlugin
- Commits, pushes and replies on pull requests
- Reruns CI and may update branches
- ConcatMCP server
- Runs the whole application and writes projects, frames and exports to the home folder; every `edit_apply` call changes the project
- Reads your video and audio through `media_probe` and `media_import` but does not upload them to a cloud
- Confluent MCP ServerMCP server
- Can delete topics, schemas and connectors
- Produces messages to clusters, including production
- shell_exec runs arbitrary commands in the box, and the platform MCP manages containers, networking and secrets
- The installer runs as root and the admin token is stored on the host
- The project is mounted read-write, so the agent can change or delete files in it
- Enabled configurations and optional credentials are available inside the container
- The container keeps network access
- Contentful MCP ServerMCP server
- Publishes and deletes content on a live site
- Changes the content model and environments
- A local proxy sits between the client and the API and sees whole requests, images and auth headers included
- The installer edits ~/.codex/config.toml and a proxy block in ~/.codex/.env
- A bundled MCP server and hooks run inside every session where the plugin is enabled
- The background bridge process indexes the workspace and tracks changes, and an outside service processes the data
- The sign-in data sits in ~/.ctxce/auth.json and is shared by the bridge process and the MCP server
- The memory tool keeps knowledge between sessions and returns it to the agent as context
- ContextForgeMCP server
- Stores credentials and tokens for connected services
- Proxies tool calls that write to external systems
- Continuous ClaudeWorkflow
- Creates and merges pull requests without your involvement
- The install script runs through the shell
- Continuous ClaudeWorkflow
- Installs hooks that run commands on agent actions and changes the global config in ~/.claude
- Deploys a PostgreSQL database and a daemon, and handles API keys
- Coolify MCPMCP server
- Deploying, stopping and changes on your server
- A Coolify token is needed
- Copilot MCPPlugin
- The extension installs third-party MCP servers and skills into several agents' settings, and the servers then run code and use the network
- CoreCoderCLI
- The agent writes and edits files and runs shell commands, and an "always allow" decision persists between runs
- `--yes` in `-p` mode approves all mutating calls up front, and hooks run your commands around every call
- CotalCLI
- Starts agents and a background broker that read and write files in the project
- An agent can spawn other agents on the mesh with cotal_spawn
- The installer is downloaded from the internet and run in a shell
- CozempicCLI
- Writes hooks into Claude Code settings globally and starts a background process in every session
- Rewrites session files and terminates the Claude Code process so it can resume after cleaning
- Updates itself from PyPI, so new code arrives without your review, and sends anonymous usage counters
- Crew44CLI
- Launches installed console agents that write files and run commands in the projects you choose
- The built-in browser downloads Playwright and Chromium and visits addresses the agent names
- Mobile pairing passes a stream through an external relay until you host your own
- CrewshipCLI
- Runs coding agents in Linux containers that execute arbitrary code
- Stores and proxies keys and secrets, and manages the host's container runtime
- Open beta: some safety layers are early or off by default
- CrushCLI
- The agent runs shell commands and edits project files
- Tools can be allowed to run without confirmation
- Cursor pluginsPlugin
- Integrations write to external systems: email, CRM, repositories
- Some plugins access financial accounts
- Custom AI AgentMCP server
- The content of intercepted traffic, including cookies and tokens, goes to the chosen model when privacy mode is off
- Tools in Unsafe mode send requests to targets and start Burp scans
- API keys and the MCP token are stored in Burp's settings together with the encryption master key
- DB MCP ServerMCP server
- Without an API key the transport is open
- With read_only off the agent writes to the databases
- DBHubMCP server
- Runs SQL against the connected database, including writes unless read-only
- Handles a connection string containing a password
- DSH iOS SimulatorPlugin
- The agent taps and types on the simulator and on a real iPhone, including real accounts
- The plugin builds and launches WebDriverAgent on the phone and compiles helper code on your Mac
- The screen stream is served by a local web server
- DaJet MCP ServerMCP server
- execute_query runs agent-composed queries against a real database
- Database connection credentials are stored in an unencrypted config file
- DaVinci Resolve MCPMCP server
- Modifies DaVinci Resolve projects and files on disk
- Runs code and reaches the running application on the host
- DaguMCP server
- Runs arbitrary commands, containers and SSH operations
- The MCP server lets clients control runs in your environment
- Damn Vulnerable MCP ServerMCP server
- A deliberately vulnerable server with no safeguards, dangerous outside isolation
- Challenges include arbitrary code execution and remote system access
- DashClawPlugin
- Intercepts the agent's tool calls and decides whether shell commands and file writes run
- Issues an API key and keeps a decision ledger that all agents connect to
- Connects to Telegram, Discord and a mobile PWA for remote approvals
- DataLens Dev MCPMCP server
- Creates, publishes and deletes objects in a production DataLens organization
- Uses a Yandex Cloud user IAM token
- DataLens: a Java database MCP serverMCP server
- The server runs queries against a live database, though QueryGuard is meant to block dangerous operations
- DataNewton MCPMCP server
- The agent's requests, including INNs and names of the companies checked, go to the DataNewton server, and replies with data on sole proprietors, managers and founders reach your agent's model and its provider
- The connection is paid: the plan token is spent on requests, and an agent left unchecked can use up the limit fast, especially when listing companies
- Databricks AI Dev KitPlugin
- The installer runs a remote script via curl and bash
- The agent gets access to data and resources in a cloud workspace
- David Ondrej SkillsSkill
- Skills run shell commands, start processes and configure servers and databases
- Orchestration launches and delegates work to other agents
- Some skills call external services with keys
- DebugMCPMCP server
- The debugger runs your code
- Expression evaluation executes code in the process
- The install script downloads and installs a binary
- Commands need cluster access and can change configuration, modules and users
- Runs shell commands and edits project files as a coding agent
- Sends project code to an external model API under the user's key
- DeepSeek Bridge for Claude CodeMCP server
- Launches the local DeepSeek Harness agent, which can write files and run commands
- What it may do is set by DSH permissions, not by the bridge
- Spends the DeepSeek API key from the bridge environment
- Runs code and commands proposed by the model
- Loads third-party plugins with access to files, network and credentials
- No security audit, still a developer preview
- DeepTideCLI
- The agent reads and writes files, runs shell commands and connects MCP servers
- The DeepSeek key is kept in settings and requests containing project code go to the provider's API
- Installation uses a script from the network and packages that run code on install
- DeerFlowCLI
- The agent runs code and bash commands in a sandbox or on the host depending on settings
- Stores model and search provider keys in .env
- The autonomous pipelines execute the target's code and require a gVisor sandbox
- The customize skill edits pipeline code and runs validation commands
- The skills handle code and vulnerability findings, which are sensitive project data
- Delight MCP module for 1C-BitrixMCP server
- Can run arbitrary PHP code, shell commands and SQL queries on the site server
- Gives full access to the site's database and file system through MCP tools
- Delphi IDE Remote MCP ServerMCP server
- The server edits files, runs MSBuild, git and deploy to real targets
- It can drive the desktop and run programs remotely
- The remote mode exposes the machine over the network behind a token
- Bulk-changes bids by rule and creates campaigns and ads
- The universal direct_call tool opens access to every changing API v5 method
- Desktop Commander MCPMCP server
- Runs arbitrary shell commands on your machine
- Reads, edits and deletes files outside the project
How we set the access level
MethodologyThe 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.