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

    • A PostToolUse hook makes a git commit after every file write in the vault
    • The permissions in settings.json let the agent write and edit vault folders and run git, cp and mv
  • Obsidian MCPMCP server
    • Reads and changes the Obsidian vault files
    • Clients can call destructive tools
    • Deletion goes to the bin by default
    • Launches many agents that change code in parallel
    • Runs a local API and installs Claude Code hooks
    • Works with your Apple Developer account: the App Store Connect key, certificates and signing profiles
    • Publishes builds to TestFlight and the App Store and can update the app's store listing
    • Project code is built on the service's remote machines, and after the skill install the odevio commands are pre-approved in Claude Code
  • Odoo MCPMCP server
    • Connects to an Odoo database with credentials
    • Can write changes to ERP data
    • Works with production systems
    • Can create paid resources and charge the account balance
    • A full API token opens access to every service on the Timeweb Cloud account
    • It covers write operations like card uploads, prices and promotions with no separate read/write gate
    • The helper only guards against a wrong host and rate limits, not against the act of writing data itself
    • By default the server only reads, but with the --ydb-allow-write flag it executes arbitrary SQL, including data changes and deletions
    • Token or service account authentication grants broad database access unless the key's rights are restricted
    • Through MCP the agent creates, edits, publishes and runs workflows in your n8n instance, and workflows reach your services
    • The plugin installs SessionStart, PreToolUse and PostToolUse hooks, which are shell scripts that run on agent calls
    • The agent deals with n8n credentials and keys, so tokens must stay out of node text fields
    • Uses a paid ZenMux key kept in the system store and sends your images and prompts to an outside generation service
    • Runs Python and TypeScript scripts, ffmpeg and a local key-entry page, and writes asset files and code into the project
    • Every generation request costs money, so a batch without your approval of the pilot clip is ruled out by the skill's instructions
    • The wallet private key and Circle keys open real accounts, and an agent with the `pay` command moves money on its own
    • A mistake in the policy file or an oversized budget lets the agent spend more than you planned
    • The agent and owner tokens live in the `.env` file and the policy file: a leak of either opens payment control
  • OmnigraphMCP server
    • The server listens on the network (`0.0.0.0:8080` in the example) and works with your storage, including S3 through `AWS_*` keys
    • After the `--mcp-allow-writes` flag the agent changes the graph and creates, deletes and merges branches
    • The install script runs through `curl | bash` from the main branch and puts binaries in `~/.local/bin`
  • OneBridgeMCP server
    • execute_code runs arbitrary 1C server code as the current user
    • execute_query gives direct access to database data, including catalogs and documents
    • The skills handle sensitive money and data: statements, invoices, contracts, CRM records, and with connected mail they read emails and create drafts
    • The `agent-army` skill launches dozens of parallel agents that edit files in bulk and use up model limits
    • Some skills run Python scripts on the user's files and reach the network, for example for SEC EDGAR filings
    • An agent with a full-access key sends mail from your address and domain on its own, and human confirmation works only while `hitl_enabled` is on
    • The agent reads mail from strangers, and the scanner checks text and HTML but gives no protection against attachments until ClamAV is on
    • The server keeps relay passwords, webhook secrets and DKIM keys in its database, encrypted with a single key
    • Drives the real desktop session: clicks, types and can change data in apps
    • Requires Accessibility and Screen Recording permissions
    • Runs the agent CLI with no iteration limit and, by default, no confirmations, so the agent can change and delete project files and run commands
    • Commits to git after every iteration and writes state to `.ralph/`, while repeated requests spend the agent's paid plan
    • Servers operate on real AWS resources and can modify them
    • Cloud account credentials are required
    • The agent controls the phone and can act inside apps
    • Screenshots are sent to the model, possibly a third-party service
  • Open-ReverseLabMCP server
    • It runs reverse-engineering tools and arbitrary binary analysis
    • It handles malware and exploit samples
    • The catalog covers write operations on orders, prices and content, reachable through an explicit allow_write flag
    • wb_raw_request allows calling arbitrary endpoints outside the catalog
    • The agent executes arbitrary Python code and controls a browser: it can press buttons, submit forms and download files
    • Saved logins give access to your accounts: cookies and localStorage are stored in a profile on disk
    • The allowed-domain list is empty by default, so sites are not restricted
    • Acts through the user's real logged-in browser and can take actions under their accounts
    • Writes and runs adapters that decide themselves how to interact with a site
  • OpenChatXMCP server
    • ChatGPT gets tools that run commands, edit files and control applications with your user's permissions
    • An OpenAI platform API key is stored on your machine and runs the tunnel
    • Other MCP servers are attached through this one, so their permissions become available to ChatGPT too
    • Integration skills send emails and publish posts on your behalf
    • They need keys to ad, analytics and CRM services
    • The assistant sends messages in messengers on your behalf
    • Runs tools and commands on the gateway host
    • Stores provider keys and channel tokens
    • In build mode it edits files and runs shell commands
    • Sends project code to the chosen model provider
    • The code-review skill posts comments on a GitHub pull request when asked
    • The SageMaker and Hugging Face Jobs skills create cloud resources and can spend money
    • The skills run Semgrep, CodeQL and commands in the repository working copy
    • The installer changes OpenCode settings and disables the built-in explore and general agents in opencode.json
    • Agents write files and run commands
    • Optional features work with git worktrees, the GitHub CLI and GitHub Actions
    • Prompts sent through the Telegram chat reach an agent that edits files and runs commands on your computer
    • The bot token and user ID sit in a .env file, and local JSON commands run in a shell with the bot's permissions
    • Project files can be downloaded from the chat with /ls, and attachments go to the agent and the speech-to-text service
    • Launches terminals and agent sessions with access to your projects and git worktrees
    • A Worker web UI open to the local network gives terminal control to anyone with the password
    • The opencove command controls the canvas from scripts, and the installer downloads and unpacks a runtime bundle
  • OpenDesignMCP server
    • Install via curl piped to a shell script
    • The app launches local agent CLIs with filesystem access
    • Handles model API keys
  • OpenDiaMCP server
    • Acts in your browser and in all open sessions
    • Can read pages and submit forms in your name
    • The tunnel mode makes the server address reachable from the network
  • OpenGUISkill
    • The agent taps, types and opens apps on a real phone that holds personal messages, banking and other accounts
    • Screen screenshots go to the chosen model, and the client gets the AccessibilityService and USB debugging
    • The plugin installer runs through curl piped into bash and prepares an environment on your machine
    • Agents run arbitrary commands, with full filesystem access when not sandboxed
    • Automations receive tokens for Slack, GitHub and other services
    • Sends code to the chosen LLM provider
    • Sends messages and changes data in external services on your behalf
    • Handles tokens and runs SQL against connected databases
  • OpenPost MCPMCP server
    • With the mcp:full scope the agent publishes and schedules posts in connected social networks on your behalf, replies to comments and hides them
    • It uploads media and reads the workspace's accounts, comments and analytics
    • A developer token stays valid until you revoke it and opens workspaces without signing in again
    • Executes contracts with the host's tools: the agent starts helper agents, reads and writes files and runs commands from the sources
    • The delivery library can send email, Slack messages and webhooks, and `prose serve` brings up local schedules and HTTP triggers
    • State, receipts and dependencies are written to the OpenProse root, where `.env` may hold keys
    • init writes global agent instructions, MCP settings, a git pre-push hook and Cursor and Cline rules
    • Project changes go to the model of your Claude Code or Codex login, and scanners download executables on first run
    • A scanner you connect runs its own command with your permissions after trust
    • Several agents edit project files and run commands in parallel
    • rig setup changes ~/.tmux.conf and the npm install runs a post-install script
    • The agent asks permission to run rig without repeated confirmations
  • OpenTabsMCP server
    • Actions go in your name through a signed-in session
    • Service terms may forbid automation
    • Telemetry is on by default
    • Installs hooks into agents and adds instructions to their context
    • Writes to the project's working folder
    • The AGPL-3.0 license has copyleft conditions
    • The contents of the chosen files go to external providers, and files can hold secrets
    • Browser mode drives Chrome and uses your signed-in ChatGPT or Gemini session
    • API mode spends paid tokens on provider keys
  • OrcaCLI
    • The agent runs shell commands and edits project files
    • On macOS, computer use lets the agent operate other applications with your permission
    • Requests to DeepSeek go out with your paid key, and every task spends money from the account
  • OrcaCLI
    • Runs several agents that execute commands and change code in parallel
    • Enables remote control of agents from a phone
    • The Git token grants full read and write access to the whole Overleaf project
    • The write_file and write_section tools change project files and create commits
    • The skills reach the network through the vendor's paid services and pass requests and target URLs to them
    • They operate with credentials in environment variables
    • Some capabilities rely on anti-bot bypass and CAPTCHA solving, which carries legal responsibility
  • Oya BrowserMCP server
    • The agent acts in a browser with your imported logins and can perform actions as you on websites
    • Pages the agent reads may contain instructions aimed at it, so keep sign-in, payment and email-sending steps for a person
    • The API key opens access to your Oya account and its cloud browsers
    • Changes product prices and stock and can cancel a posting
    • Can send messages to buyers and reply to reviews on behalf of the shop
    • Start with a read-only key if Ozon offers one
  • Ozon MCPMCP server
    • Can change prices, products and stocks in the seller cabinet
    • Requires secret Seller API keys
    • Performance API ad methods involve spending budget

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.