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

    • Skills run external tools and scripts and read session history
    • GitHub and git skills modify repositories
    • Auto-install runs a script downloaded from the network
  • dbskillSkill
    • The skills write files into `~/.dbs/` and `~/.config/dbs/` and keep paid keys for the video services
    • They run scripts (update check, key setup, video extraction) and reach the network: GitHub, TikHub, 轻抖
    • Instructions for three-digit codes are pulled from a public GitHub catalog at use time
    • The agent acts in Studio in your name: it reads the organization's chats and files and calls tools that can delete and overwrite data
    • Sharing a file by link opens it to anyone with the link, and when the CLI is not on PATH the latest package version runs through bunx
    • Launches third-party CLI agents that edit the working tree and run commands
    • Some CLIs do not enforce a read-only mode
    • Runs arbitrary JavaScript in the browser
    • Can attach to the user's browser with logged-in sessions
    • Performs legally binding actions: signing, accepting, rejecting and cancelling documents
    • Stores the Didox partner token and account password
    • Works with the local signing key via E-IMZO
  • docx-mcpMCP server
    • The server creates, edits, merges and saves docx files on disk, including over the original
    • Documents hold personal and commercial data, and the personal-data removal feature is marked experimental
  • dotaiSkill
    • `cleanup` removes git worktrees and cache folders with git worktree remove --force, and `test-audit` deletes and rewrites tests
    • `passwords` works with passwords and one-time codes through 1Password and a Keychain token
    • `video-transcripts` uploads video to Gemini and, for private attachments, uses GitHub tokens and Linear cookies
    • Changes the EasyEDA Pro project: creates and edits schematics, boards and libraries and saves files, with no programmatic undo in the editor API
    • Installs the easyeda binary and a background daemon that starts by itself, and installation runs a script from the internet through curl and bash
    • Through the connector extension the agent's commands reach the editor and change the open project
  • eda-agentMCP server
    • The agent edits open schematics, boards and libraries and can save documents to disk
    • The `app_` tools simulate real keys and mouse, take focus and act inside the Altium window
    • The autonomous mode drives a project through stages on its own, so the changes can be many
    • The agent acts in the browser under your logins and cookies
    • It can perform actions on sites on your behalf
    • The agent reads your correspondence and attachments and sends and deletes email in your name
    • Access needs an app password or an OAuth token, which opens the whole mailbox
  • emisarMCP server
    • The agent runs commands on production servers, though only declared actions
    • Policy decides actions and packs: a mistake in it or overly broad host permissions widens the damage
    • The runner install needs sudo and an outbound connection to an external control plane
    • In shell mode the agent runs arbitrary commands
    • Sends data to the chosen model provider
  • fauxnixMCP server
    • Runs translated commands directly in PowerShell on your system, including deleting and moving files
    • The experimental mode replaces Claude Code's built-in Bash tool with a built binary based on node.exe
    • The MCP session holds environment variables and the working directory across calls, including values the agent set itself
    • An unofficial route, breakages on Figma updates are possible
    • Edits your design files
  • flow-mcpMCP server
    • The program signs in as your Google account and clicks buttons on the Flow site; the documentation does not discuss the service's rules for such automation
    • Shooting video spends your account credits
    • The local panel and API listen on 127.0.0.1 and are protected by a token in ~/.flow-mcp/token
  • freee-mcpMCP server
    • The server can create, change and delete deals, invoices, requests and documents in your accounting system
    • The token opens payroll, employee and tax return data
    • Start with the app's minimum permissions and check the current company before writing
    • In all mode, it enables financially sensitive VK API methods
    • Can publish, send messages, and manage communities if granted permissions
    • Runs gcloud commands with your account's rights
    • Can delete buckets, objects and backups
    • Runs about twenty Python scripts through uv and drives Chrome or Chromium, and ML packages and model weights are downloaded on first run
    • Reverse image search hands the photo to Baidu and Yandex, and the other scripts call OpenStreetMap and panorama services
    • A dual-use tool: it locates a place from a photo, including someone else's, so personal photos and frames with people require permission
  • gf-data-1cMCP server
    • Installs the GTData extension into the 1C base, which needs the base login and password
    • Runs queries on balances, turnovers, catalogs and accounting totals
    • Rejects data-changing queries before they reach the base
    • The utility drives a browser holding your Google session, and automation falls under Google's terms
    • Video generation spends account credits, and Flow interface errors can interrupt a batch midway
    • The tool creates, edits and deletes issues, gists, Actions secrets and variables, triggers workflows and cuts releases under your `gh` token
    • Through `gh-axi api` the agent gets direct GitHub API access with the rights of your login
    • Reads the encrypted browser cookie store and takes `user_session`, which gives full access to your GitHub account
    • Publishes the chosen files on GitHub under your name, and an upload cannot be removed
    • Calls undocumented GitHub endpoints, and `gh image -- ...` also writes to issues and PRs
    • Access tokens to your health data are stored in plaintext on disk in ~/.config/ghealth
    • Reads sensitive data: heart rate, ECG, glucose, sleep and nutrition
    • Can create and delete exercise, sleep, weight, height and body fat records
    • Xianyu cookies open the account: the agent can publish and take down items and message buyers in your name
    • Too frequent actions break the marketplace rules and may get the account blocked
    • The installer downloads and runs a script
    • The agent runs commands and works with files
    • MCP extensions can access external services
    • Ships releases to Google Play production tracks and changes store listings, subscriptions and prices
    • Manages Play Console users and permissions and reads financial reports, purchases and refunds
    • The service account key grants access to every app it is allowed to touch
    • Stores GigaChat credentials and sends the agent's code and session context to GigaChat
    • Client key authorization is off by default, and DEV mode exposes /docs and the /logs endpoints
  • gstackSkill
    • Setup runs scripts and builds its own browser
    • Adds hooks to ~/.claude/settings.json and a session-start auto-update
    • Skills create commits and PRs and drive a browser with your sessions
  • gwqCLI
    • Commands create and delete worktrees and branches, and remove with a flag forces deletion of an unmerged branch
    • `gwq exec` and `gwq tmux run` run arbitrary commands in the chosen folders
  • gwsCLI
    • Gets OAuth access to mail, files and calendar
    • Can send email and modify documents
  • h1-brainMCP server
    • The HackerOne API token opens your reports, program scopes and conversations with security teams
    • The briefing from hack puts the agent in offensive mode, and targets may be tested only within an authorized program
    • The personal h1_data.db holds details of vulnerabilities you found, including ones not yet fixed
  • ha-mcpMCP server
    • Controls real smart-home devices
    • Uses a long-lived Home Assistant access token
    • Connected users share the server credentials
    • Runs pentest tools and network attacks on hosts, including scanners, exploits and secret handling
    • Many commands need elevated privileges and act on real targets; use only with written authorization
  • hcomCLI
    • Writes hooks into agent configs
    • Agents launch and kill other agents; with the relay this extends to other devices
    • apply, chat_send and cv_push write to the live account immediately, with no confirmation or dry-run
    • Access goes through a session cookie rather than an official API, which may violate hh.ru's rules
    • The hhtoken cookie is effectively an account access key and must be kept secret
    • apply_to_vacancy, apply_two_step and send_message write directly to the live account and employer correspondence
    • Logging in via the official hh.ru mobile app's OAuth instead of your own app may violate the service's terms
    • The token grants broad access to the job seeker's resumes, applications and correspondence
    • Can invite a candidate and send a message on the employer's behalf (hh_employer_negotiations invite, message_send)
    • Viewing a resume spends a paid employer contact, so a mistaken call costs money
    • The employer OAuth tokens grant broad access to vacancies, applications and candidate resumes
    • Automates applications on your behalf, which may lead to hh.ru account restrictions under its rules
    • Stores the browser session and CV locally, and sends vacancy and CV text to an external LLM via OpenRouter
    • The hh-apply-chat flow can automatically send an application and a message to the employer unless --dry-run or --no-submit is passed
    • Bulk application automation, per the README's own warning, can lead to hh.ru account restrictions
    • The Home Assistant token opens control of locks, cameras, climate and every other device in the house
    • The tools install and remove add-ons and HACS packages and change automations in the working configuration
    • The HTTP interface on port 3000 exposes home control to network clients until you restrict access to the port
    • bind_payment_card and charge_saved_card can bind a card and charge money from it
    • create_topup_card and create_topup_crypto initiate top-ups and monetary operations
    • The key is issued automatically with no strict identity check, up to twice a day per IP
    • The demo execute tool runs arbitrary 1C code on the server or thin client
    • Authentication is off by default; anyone reaching the port gets access to every tool
  • hyper-mcpMCP server
    • Plugin signatures can be turned off
    • Plugins by http links are not checked by signature
    • There are no stable releases
  • iMCPMCP server
    • Opens messages, contacts, calendar and reminders to the agent
    • macOS only
  • ida-mcp-rsMCP server
    • Runs IDAPython scripts and can modify the analysis database
    • A reverse engineering tool; use it on your own or authorized binaries
    • Working with untrusted binaries requires environment isolation
  • ida-multi-mcpMCP server
    • The py_eval tool runs arbitrary IDAPython inside the IDA process, and patch, patch_asm and put_int change bytes and database data
    • The ida-multi-mcp --install command installs a plugin into the IDA directory and edits the settings of every MCP client it finds
    • The agent analyzes binaries that may include malware samples, so work in an isolated environment
    • Can change recipe cards and record cash encashment in the live restaurant system
    • Keeps the iiko login and password in plain text, with SSL verification off by default
    • Raw calls reach any iikoServer endpoint beyond the tool list

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.