StackGuardian MCP - Deploy Terraform, OpenTofu and Ansible with your AI coding agent
One-command MCP setup and ready-to-use agent skills for Claude Code, Cursor, Codex and more.
Connect your coding agent to StackGuardian, the platform that runs and governs your infrastructure deployments. Then ask in plain language to turn your existing infrastructure code into deployment workflows, review changes, require approvals, add policies and find out why a run failed.
npx @stackguardian/mcp-wizard
No Node.js? Use the one-line installer for macOS, Linux or Windows.
Then try asking your agent:
Deploy the Terraform in my
infrarepository, folderenvs/dev, to AWS.
Create an OpenTofu workflow for this repository that waits for my approval before every apply.
Find out why my last infrastructure deployment failed.
Get started · Supported agents · Skills · Skills guide
What you can do
Creating a workflow doesn't deploy anything. A deployment happens when a run starts: when you or your agent start a plan or apply, or when a schedule you set fires. Most agents ask you to approve each tool call, so you see a run request before it starts.
Get started
Before you start
- A StackGuardian account with access to the organization you want to connect, and roles that let you create and run workflows. Your region is Europe (
app.stackguardian.io) or United States (us.stackguardian.io). - Connectors for what you deploy: a cloud connector for the account you deploy to (AWS, for example), and a git (VCS) connector if your infrastructure code is in a private repository. The agent lists the connectors your organization already has.
- At least one supported coding agent on this machine.
- Node.js 18 or later for
npx, or the standalone executable below.
Install
With Node.js:
npx @stackguardian/mcp-wizard
Without Node.js, the installer downloads a standalone executable, checks it against the release's SHA256SUMS, installs it and runs it:
# macOS and Linux
curl -fsSL https://dl.stackguardian.io/add-sg-mcp/install.sh | sh
# Windows (PowerShell)
irm https://dl.stackguardian.io/add-sg-mcp/install.ps1 | iex
Then:
- Sign in. Pick your region. The StackGuardian dashboard opens in your browser; sign in if needed, then choose the organization, the roles the agent may use and how long the access lasts.
- Connect your agents. Back in the terminal, choose the agents to connect. The ones found on this machine (or those you picked last time) are preselected.
- Restart those agents. Most read their MCP configuration only at startup.
- Ask. Describe what you want to deploy, change or investigate. The agent looks up your workflow groups, templates and connectors, and checks with you before it creates or changes anything.
With the standalone executable, run add-sg-mcp wherever this README says npx @stackguardian/mcp-wizard. See Install script options to pin a version, change the install folder or use a mirror.
Check that it works
- Claude Code: run
/mcp.StackGuardian-<org>should be listed, and the skills show up as/sg-create-workflow,/sg-update-workflow,/sg-upgrade-workflow,/sg-create-stack,/sg-create-policyand/sg-manage-secrets. - Any agent:
npx @stackguardian/mcp-wizard statusshows your saved sign-in and which agents have the server and the skills. - Ask your agent something small, such as "Show my five most recent StackGuardian workflow runs and their status."
Nothing there? See Troubleshooting.
Supported agents
Claude Code, Cursor, Codex, VS Code (Copilot), Gemini CLI, OpenCode, Windsurf, plus GitHub Copilot CLI, Kiro CLI, Zed, Cline, Goose, Grok Build, Kilo Code, Kimi Code, Antigravity, Pi, Mastra Code and MCPorter. Run npx @stackguardian/mcp-wizard list-agents for the current list and each agent's config file.
Every agent gets the MCP server. Skills are installed into each agent's skills directory where the agent has one; Mastra Code and MCPorter get the server only. How a skill starts differs by agent: Claude Code offers each one as a slash command, and other agents pick up a skill when your request matches it. See Agent-specific notes for other differences.
MCP and skills
MCP provides the tools. Skills provide the playbooks.
- The StackGuardian MCP server gives your agent tools for your StackGuardian organization: search workflows, read run logs, create a workflow, start a run, store a secret and more. The server is hosted by StackGuardian; this repository contains the installer that connects your agents to it. In Claude Code,
/mcplists the server's current tools. - The skills are playbooks for multi-step tasks, bundled in
skills/. Each one tells the agent which tools to use, in what order, and what to ask you along the way, so it brings you choices instead of guessing IDs or defaults.
The skills guide walks through each skill step by step, with more examples and the StackGuardian terms you'll meet.
Why run deployments through StackGuardian?
Terraform, OpenTofu and Ansible still do the work. Running them as StackGuardian workflows instead of from your laptop adds the parts a team needs around them:
- Managed execution. Runs happen on StackGuardian-managed runners or on your own private runner group, not on the machine your agent runs on.
- Run history. Every plan and apply is recorded with its logs and outputs, so you, your teammates and your agent can look back at what happened.
- Reviews and approvals. A Terraform or OpenTofu workflow can pause after the plan until named approvers release it; a custom pipeline can pause before any step.
- Policies. Tirith or Rego rules check infrastructure changes and block them, warn or ask for approval.
- Reusable templates. Library templates and stack templates give teams a reviewed starting point, with revisions you can upgrade to.
- After the deploy. Schedules, drift detection for Terraform and OpenTofu workflows, notifications, webhooks and chained workflows.
- Secrets and connectors. Workflows use cloud and git connectors and vault secrets by reference, so credentials don't have to sit in your repository.
These are per-workflow settings: none is switched on just by connecting your agent, and some apply only to certain workflow types.
Choose how your agent signs in
Only the grant sign-in opens the dashboard. An API key is never fetched for you: pass it with --token (or SG_API_KEY). Without --auth, a key you pass, or one saved by an earlier --token run, selects the API key; otherwise you sign in for a grant.
The tools that create, change, run or delete things act on your real StackGuardian organization, with the access of the credential you chose. Read what the agent is about to do before you approve a tool call.
Grant tokens
npx @stackguardian/mcp-wizard (optionally --region eu --org my-org) asks the StackGuardian OAuth broker for a grant token instead of using your API key. The browser opens a consent page where you pick the organization, the roles the agent may use and how long the grant should last; the CLI receives an sgat_ token on a loopback callback and writes it to each agent as an Authorization: Bearer header.
- The token is bound to one organization and the roles you approved — it is not your full user key, so an agent can be given less than you have.
npx @stackguardian/mcp-wizard statusshowsexpires <date>,no expiryorexpired. An expired grant is never reused; the next run asks for a new one.logoutrevokes the saved grant and deletes the local copy. Revoke a grant any time from Profile → Connected apps in the dashboard.npx @stackguardian/mcp-wizard loginrequests a fresh grant, then asks which agents to install it to.- If an API key is saved from an earlier
--tokenrun,npx @stackguardian/mcp-wizardkeeps using it. Pass--auth grantto switch to a grant.
OAuth (preview)
npx @stackguardian/mcp-wizard --auth oauth --org my-org --region eu writes the server URL without a credential for agents that implement MCP OAuth themselves (Claude Code, VS Code, Cursor, Codex, Gemini CLI, Windsurf). The agent then signs you in on first use. This needs the StackGuardian OAuth broker to be enabled for your environment.
Everyday tasks
Headless / CI
Skip the browser with an API key from Profile → API keys in the dashboard:
npx @stackguardian/mcp-wizard --token sgu_… --org my-org --region eu -y --all
# or
SG_API_KEY=sgu_… SG_ORG=my-org SG_REGION=eu npx @stackguardian/mcp-wizard -y -a claude-code
--api-base https://…/api/v1 replaces --region for non-standard environments. In a Docker image, curl -fsSL https://dl.stackguardian.io/add-sg-mcp/install.sh | ADD_SG_MCP_NO_RUN=1 sh installs the executable without running it.
Troubleshooting
The server doesn't show up in my agent. Restart the agent; most read their MCP config only at startup. Then run npx @stackguardian/mcp-wizard status. If the agent isn't listed, it wasn't selected: run npx @stackguardian/mcp-wizard -a <agent>. With --project, the server is only in that project's config.
"No coding agents detected on this machine." With -y, or when there's no terminal to prompt in, only detected agents are used. Name them instead (-a claude-code -a cursor) or pass --all.
The browser doesn't open, or sign-in times out. Add --no-browser and open the printed link in a browser on the same machine; the sign-in returns to a listener on 127.0.0.1. Sign-in waits five minutes (--login-timeout <seconds> changes that). On a remote server or in CI, use an API key instead (Headless / CI).
My agent gets "unauthorized" errors. The credential in the agent's config no longer works:
- API key rotated or deleted: run
npx @stackguardian/mcp-wizard --token <new-key> --org <org> --region <eu|us>to write the new key to your agents. - Grant expired:
npx @stackguardian/mcp-wizardasks for a new one. - Grant revoked: run
npx @stackguardian/mcp-wizard loginand pick your agents.
The skills don't appear. Restart the agent. Mastra Code and MCPorter have no skills support, so they get the server only. --skip-skills skips them on purpose.
"… exists and was not installed by add-sg-mcp." A skill folder with the same name is already in the agent's skills directory. Move it away and run again.
Claude Desktop. Not supported here: Claude Desktop only takes local servers from this tool. Add StackGuardian through Settings → Connectors instead.
"this Linux has no glibc". The standalone executables need glibc, which Alpine and other musl-based systems lack. Use npx @stackguardian/mcp-wizard with Node.js there.
Reference
Options
npx @stackguardian/mcp-wizard [options]
--region <eu|us|qa> StackGuardian region (prompted when omitted)
--org <org> Organization to connect (pre-selected in the browser)
-a, --agent <agent> Agents to install to (repeatable; default: the ones detected)
--all Every supported agent
-y, --yes No prompts (installs to all detected agents)
--project Project scope (current directory) instead of your user profile;
the generated files are added to .gitignore
-n, --name <name> Server entry name (default: StackGuardian-<org>)
--skip-skills Do not install the skills
--token <api-key> Use this API key instead of signing in (or SG_API_KEY); needs --org
--api-base <url> API base for --token or --auth grant, in place of --region
--no-browser Print the sign-in link instead of opening a browser
--login-timeout <seconds> How long to wait for the browser (default 300)
--dashboard-url <url> Another dashboard (other environments, local dev)
--auth <apikey|grant|oauth> Credential mode (default: apikey when a key is given, otherwise grant)
remove and logout --purge take every StackGuardian-<org> entry and the skills out. remove --name <entry> takes out just that entry and keeps the skills.
Install script options
Pass options to the first run:
curl -fsSL https://dl.stackguardian.io/add-sg-mcp/install.sh | sh -s -- --region eu -a claude-code
& ([scriptblock]::Create((irm https://dl.stackguardian.io/add-sg-mcp/install.ps1))) --region eu -a claude-code
Environment variables (with curl | sh, set them on the sh side: … | ADD_SG_MCP_NO_RUN=1 sh):
Builds exist for macOS (Apple Silicon, Intel), Linux with glibc (x64, arm64) and Windows x64. To install by hand, download https://dl.stackguardian.io/add-sg-mcp/v<version>/add-sg-mcp-<os>-<arch>.tar.gz (or add-sg-mcp-windows-x64.zip) and put the executable on your PATH; https://dl.stackguardian.io/add-sg-mcp/latest names the newest version, and SHA256SUMS sits next to the archives.
gh attestation verify <file> --repo StackGuardian/stackguardian-mcp confirms that a file was built by this repository's workflow. Releases up to and including 1.0.6 were built before the repository was renamed from add-sg-mcp; verify those with --repo StackGuardian/add-sg-mcp.
dl.stackguardian.io has every release from 1.0.4 on.
The executables are not yet notarized by Apple or code-signed for Windows. The install scripts are not affected, but a browser download is quarantined on macOS and may show a SmartScreen prompt on Windows (More info → Run anyway).
Agent-specific notes
- Codex and Grok Build store MCP servers in a TOML file that is rewritten as a whole; comments in
~/.codex/config.tomlare not preserved. - Claude Desktop and fx are skipped: Claude Desktop only supports local (stdio) servers here and fx only sends bearer tokens. Add StackGuardian to Claude Desktop through Settings → Connectors instead.
- Claude Code and GitHub Copilot CLI share
.mcp.jsonin project scope. - Mastra Code and MCPorter get the server but have no skills directory.
How it works
-
npx @stackguardian/mcp-wizardstarts a one-time callback listener on127.0.0.1and opens the StackGuardian sign-in for your region in your browser. -
You sign in if needed and choose the organization, the roles the agent may use and how long the grant lasts.
-
The browser returns a one-time code to the loopback callback, and the CLI exchanges it for a grant token. The token is stored in
~/.config/add-sg-mcp/credentials.json(mode0600). With--token, steps 1 to 3 are skipped and the key you passed is stored instead. -
The CLI writes the server entry into each selected agent's own config file:
{ "mcpServers": { "StackGuardian-<org>": { "type": "http", "url": "https://api.app.stackguardian.io/api/v1/orgs/<org>/mcp/", "headers": { "Authorization": "Bearer sgat_…" } } } }(Codex gets the TOML equivalent, Goose YAML, and so on — each agent's native format. With an API key the header is
apikey sgu_….) -
The skills are copied to
~/.agents/skills/and symlinked into each agent's skills directory (~/.claude/skills,~/.codex/skills,~/.cursor/skills, …), so a single copy serves every agent.
Security
- On Windows there is no
0600equivalent; the file relies on the ACL of your profile folder. - With an API key (
--token) the credential is your user API key for the selected organization: it acts as you, with your roles, and it does not expire on its own. Rotate it from Profile → API keys in the dashboard if a machine is lost. - With a grant (the default) the credential is a grant token bound to one organization and the roles you approved, and it can carry an expiry.
logoutrevokes it; you can also revoke it from Profile → Connected apps. For an API keylogoutonly deletes the local copy — it never invalidates the key. - By default nothing is written into project directories.
--projectwarns and adds the generated files to.gitignore. - The CLI accepts a callback only when its one-time
statematches, and sends the code and the credential only to API hosts understackguardian.io(orlocalhost, for local development). statusnever prints the key or the grant token in full; the callback page strips the one-time code from the browser history, and the grant flow keeps its PKCE verifier off the front channel (only the one-time code travels through the browser).
Contributing
Pull requests are welcome: fixes to the installer, support for more agents, and improvements to the skills in skills/. For a problem with your StackGuardian account or the hosted MCP server, contact [email protected].
bun install
bun run dev -- --help # run from source
bun run typecheck && bun run test
bun run build:binary # standalone executable for this machine
Tests are plain tsx scripts with node:assert (no mocks; the login test spins up a real loopback server), run by tests/run.mjs under a throwaway HOME. See AGENTS.md for the workflow, docs/RELEASING.md for releases and docs/superpowers/specs/ for the design.
If this project is useful to you, starring the repository helps other developers find it.
This project is a fork of neon-solutions/add-mcp (Apache-2.0); the agent config writers come from upstream unchanged. Passing a URL or package name as the first argument still installs any other MCP server the way add-mcp does.
