toolstile MCPサーバー

toolstile

Check the MCP server configs your AI agents use, in one command:

npx toolstile audit

It reads the configs of Claude Code, Claude Desktop, Cursor, Codex, VS Code, Windsurf, Devin Desktop, Gemini CLI and Cline and grades them for:

Check	Severity
API keys, tokens and passwords written into config files, URLs or arguments	high
Remote servers reached over plain http://	high
Containers run with --privileged, the Docker socket or host networking (--net=host)	high
The same credential pasted into several clients' configs	medium
Config files holding secrets that other accounts can reach (file and folder permissions)	medium
npx, uvx and docker servers that run whatever version is newest	medium
Filesystem servers that can read your whole home folder or disk	medium
Permission rules that approve every tool of a server without asking	medium
Tools on a client's auto-approve list	low

It also lists servers set up separately in several clients. That is context, not scored: remote servers need an entry per client.

toolstile audit · local MCP config check · nothing leaves this machine

  Grade C  71/100   2 high · 1 medium
  10 MCP servers in 4 files from Claude Code, Cursor, Codex

HIGH  Plaintext secrets (2)
  Anything that can read these files can use the credentials, including
  backups, dotfile syncs and other tools.

  • github (Cursor)  GitHub token
    ~/.cursor/mcp.json  mcpServers.github.env.GITHUB_PERSONAL_ACCESS_TOKEN
  • linear (Claude Code)  authorization header
    ~/.claude.json  mcpServers.linear.headers.Authorization

  Fix Move them into environment variables, or use the server's OAuth sign-in.
      Rotate any secret in a file that was shared or synced.
…
Fix them: npx toolstile audit --fix (shows every change; you pick which to apply)

Fix what it finds

npx toolstile audit --fix

lists every change it can make, as a diff with secret values hidden, then lets you pick which to apply (space toggles, enter applies):

  1.  HIGH   github (Cursor): move secret to $GITHUB_PERSONAL_ACCESS_TOKEN
             ~/.cursor/mcp.json  mcpServers.github.env.GITHUB_PERSONAL_ACCESS_TOKEN
             - "••••"
             + "${env:GITHUB_PERSONAL_ACCESS_TOKEN}"
  2.  MEDIUM linear (Claude Code): pin linear-mcp to 1.4.2
             ~/.claude.json  mcpServers.linear.args[1]
             - "linear-mcp"
             + "linear-mcp@1.4.2"

Finding	What --fix does
Plaintext secret in env or a header	Replaces the value with a variable reference in the client's own syntax (${NAME} for Claude Code and Gemini CLI, ${env:NAME} for Cursor, VS Code and Windsurf, env_vars / bearer_token_env_var / env_http_headers for Codex). The value goes into ~/.config/toolstile/secrets.env (mode 600), and one line in your shell profile (~/.zshrc, ~/.bash_profile or ~/.bashrc, ~/.profile, or fish's conf.d) loads it. A secret shared by several clients becomes one variable.
Config file with secrets readable by other accounts	chmod 600
Unpinned npm or PyPI package	Pins the version that is current now, looked up on npm or PyPI
Rule that approves a whole server, or enableAllProjectMcpServers	Removes it. These start unselected, since the agent then asks before each tool.

Everything else (secrets in arguments or URLs, Claude Desktop and Cline, which cannot read variables, Devin Desktop, Windows, http:// servers, Docker images, broad filesystem roots and auto-approve lists) is listed with what to change by hand.

Each edited file is checked before it is written: it must parse to exactly the planned result, or it is left alone. JSON files keep their comments and layout; a Codex server's table is rewritten in place. The original files are copied to ~/.config/toolstile/backups/ first. They still hold the old secrets, so delete them once your agents work. After moving secrets, open a new terminal and restart your agents so they start with the variables, and rotate any secret that sat in a shared or synced file.

--dry-run only shows the changes; --yes (or -y) applies the ones that start selected without asking.

The audit only reads

• Reads config files on your machine. Writes nothing unless you ask for a card, or run --fix and confirm the changes.
• Makes no network requests and collects no telemetry. (--fix looks up the current version of the unpinned packages it offers to pin on npm or PyPI. Only the separate auth and mcp commands below talk to Toolstile, and only when you run them.)
• Never prints, stores or sends secret values. Findings name the file and key, for example mcpServers.github.env.GITHUB_TOKEN, and the kind of secret.
• Never quotes a config file: one that cannot be read or parsed is listed with a reason and position only, for example invalid TOML at line 3, column 12.
• Plain TypeScript with two runtime dependencies (smol-toml, for Codex's TOML config, and jsonc-parser, so --fix keeps comments and layout in JSON files) and an optional one (resvg-js, only to render --card as PNG). Read it before you run it.

Options

--fix               Suggest fixes and apply the ones you select
-y, --yes           With --fix: apply the fixes that start selected, without asking
--dry-run           With --fix: only show the changes
--json              Print results as JSON
--card [file]       Write a shareable grade card (default toolstile-audit.png;
                    use a .svg name for SVG). Contains only the grade and counts.
--redact            Hide your home folder and project paths, for posting the output
--project <dir>     Also scan project configs in <dir> (default: current directory)
--no-project        Scan only user-level configs
--fail-on <level>   Exit with code 1 if any finding is at or above high, medium or low

--fail-on high makes it usable in CI or a pre-commit hook for repositories that check in .mcp.json, .cursor/mcp.json or .vscode/mcp.json.

Exit codes:

Code	Meaning
0	The audit ran, and no finding reached the --fail-on level (or --fail-on was not set)
1	A finding is at or above the --fail-on level
2	Invalid options, or the audit itself failed

Scoring

Each distinct problem has a weight: 10 for a leaked secret, plain-http remote or privileged container; 4 for readable secret files, unpinned packages, broad filesystem access and whole-server auto-approval; 2 for each copied secret; 1 for auto-approved tools. A secret pasted into several files counts as one leak. The score is 100 / (1 + weight / 50): one leaked secret scores 83 (B), four score about 56 (D). Grades: A 90+, B 75+, C 60+, D 40+, F below 40.

Share your grade

npx toolstile audit --card

writes a 1200×630 image with your grade and counts (no server names, paths or values) and prints a ready-to-post link. To post the text report instead, add --redact so it shows ~ and <project> in place of your folders.

Files it reads

Client	User config	Project config
Claude Code	~/.claude.json, ~/.claude/settings.json	.mcp.json, .claude/settings*.json
Claude Desktop	claude_desktop_config.json in the app's config folder
Cursor	~/.cursor/mcp.json	.cursor/mcp.json
Codex	$CODEX_HOME/config.toml (default ~/.codex)	.codex/config.toml
VS Code	user settings.json and mcp.json	.vscode/mcp.json
Windsurf	~/.codeium/windsurf/mcp_config.json
Devin Desktop	~/.config/devin/mcp_config.json (%APPDATA%\devin on Windows)	.devin/mcp_config.json, .devin/mcp_config.local.json
Gemini CLI	~/.gemini/settings.json	.gemini/settings.json
Cline	cline_mcp_settings.json in VS Code or Cursor storage

Connect your agents to Toolstile

If you use the managed Toolstile service, the CLI can also sign in and set up the agents on a machine for you (self-hosted instances use each agent app's own sign-in instead):

npx toolstile auth login
npx toolstile mcp setup

auth login prints a link and a code. Open the link on any device with a browser (your laptop or phone) and sign in. The page fills in the code and this machine's name; edit the name if you like, check the code matches and approve. The machine appears under that name in the Toolstile dashboard. Nothing has to reach the machine you are signing in, so it works over SSH, in containers and on servers without a browser.

mcp setup lists the agents it finds (Claude Code, Claude Desktop, Cursor, Codex, VS Code, Windsurf, Devin Desktop, Gemini CLI, Cline) and lets you turn each one on or off. Run it again any time to change the selection. Each enabled agent appears under this machine in the Toolstile dashboard with its own access, so you can grant, limit or revoke it there; then choose which MCP servers it can use.

Which agents on this machine should use Toolstile?
› ◉ Claude Code  ~/.claude.json
  ○ Cursor  ~/.cursor/mcp.json
  ◉ Codex  ~/.codex/config.toml
↑/↓ move · space toggle · a all · enter save · esc cancel

Without a terminal, pass the choice as flags: npx toolstile mcp setup --enable claude-code,codex --disable cursor.

How it stays out of your configs: an enabled agent's config gets a toolstile server that runs npx -y toolstile@<version> mcp proxy --agent <agent>. The proxy relays the agent's MCP messages to Toolstile using that agent's own token, which lives only in the CLI's credentials file (~/.config/toolstile/credentials.json, readable by you alone; %APPDATA%\toolstile on Windows). No token is written into the agent's config, so toolstile audit has nothing to flag. JSON configs with comments are not rewritten; setup prints the entry to add yourself.

The CLI keeps the credentials file, --fix's secrets.env and its backups in one folder: $TOOLSTILE_CONFIG_DIR when set, otherwise %APPDATA%\toolstile on Windows and $XDG_CONFIG_HOME/toolstile (default ~/.config/toolstile) elsewhere.

npx toolstile auth logout ends the CLI's sign-in. Agents you enabled keep working until you turn them off with mcp setup or delete the device in the dashboard.

Option	Command	Meaning
--server <url>	auth login	Toolstile address (default https://app.toolstile.com, or TOOLSTILE_URL)
--no-browser	auth login	Do not try to open a browser on this machine
--enable <agents>	mcp setup	Turn these on without prompting (comma-separated)
--disable <agents>	mcp setup	Turn these off without prompting
--workspace <slug>	mcp setup	Workspace the agents use (default: the last one, or personal)
--device-name <name>	mcp setup	Name for this machine in the dashboard (default: its hostname)

Use it from code

import { audit } from "toolstile";

const { grade, findings } = audit();

About

Built by Akshit Kr Nagpal as part of Toolstile, an MCP gateway that keeps upstream credentials off your machines and lets you choose which tools each agent can use. Toolstile is available as a managed service and as an open-source self-hosted edition at toolstile/toolstile. The audit works on its own and does not require a Toolstile account.

Questions and reports: support@toolstile.com.

Development

Development uses Vite+. Its vp env shims pick Node.js from .node-version and pnpm from packageManager, so install Vite+, then run:

pnpm install
pnpm typecheck
pnpm test    # builds dist, then runs the tests with vp test
pnpm build

Development needs Node.js 22.18+, 24.11+ or 26+. The published CLI still runs on Node.js 18 and later.

Releases

The npm package is published by GitHub Actions with npm trusted publishing and provenance, so each release links back to the commit and workflow that built it. No npm token is stored anywhere.

License

MIT