Muster.

developers

MCP server

Run Muster inside Claude Code, Codex, Cursor, or Claude Desktop. Your agent can pull a real, proven finding, fix it in place, and mark it resolved — without leaving the editor.

What it is

MCP (Model Context Protocol) is an open standard that lets AI tools call external tools. Muster runs a hosted MCP server that wraps the REST API, so an AI assistant can list your assets, pull findings, read a specific finding's proof and fix, and trigger scans — all scoped to one organization by your API key. ChatGPT and the Claude plugin connect over OAuth instead, read-only unless you allow changes (see OAuth).

Setup — hosted (recommended)

Nothing to install. Point your MCP client at Muster's hosted endpoint and authenticate with an API key (create one under Account → API keys — it acts on the single org that key belongs to).

Claude Code

One command:

terminal
claude mcp add --transport http muster https://mcp.mustersecurity.com \
  --header "Authorization: Bearer mus_xxxxxxxxxxxxxxxxxxxxxxxx"

Codex

Codex reads HTTP MCP servers from ~/.codex/config.toml (its codex mcp add command is stdio-only). Add the server, then put your key in the environment it reads:

~/.codex/config.toml
[mcp_servers.muster]
url = "https://mcp.mustersecurity.com"
bearer_token_env_var = "MUSTER_API_KEY"
then set your key in the shell
export MUSTER_API_KEY=mus_xxxxxxxxxxxxxxxxxxxxxxxx

The key lives in your environment, not config.toml. Restart Codex after adding it.

Cursor / Claude Desktop

Add this to your MCP config (Claude Code also accepts it as .mcp.json in your project). Restart the client to load it.

mcp config
{
  "mcpServers": {
    "muster": {
      "type": "http",
      "url": "https://mcp.mustersecurity.com",
      "headers": {
        "Authorization": "Bearer mus_xxxxxxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}
ℹ
The key already binds to one organization, so the hosted tools auto-scope — your agent doesn't need to pass an org id around. Just ask it about your findings.

ChatGPT and Claude (OAuth)

The Muster plugin for ChatGPT doesn't use an API key. It connects with OAuth 2.1: you sign in to Muster, pick one organization on the consent screen, and approve. ChatGPT identifies itself with a Client ID Metadata Document (CIMD), so there is no client to register, and PKCE with S256 is required. Access tokens are issued for the Muster MCP server only and expire after an hour.

By default the grant has one scope, findings:read, so the client sees only the six read tools: get_summary, list_assets, list_findings, get_finding, list_scans and get_scan. It can't change findings or start scans: a call to update_finding, bulk_update_findings or trigger_scan is refused with 403 insufficient_scope.

To let the client make changes, tick Also allow … to make changes on the consent screen. That adds the findings:write scope and the three write tools: set a finding's status, update many findings at once, and start scans of the organization's own assets. The box is never ticked for you, even when the client asks for write access, and a refreshed token never gains a scope the grant didn't have. To go back to read-only, disconnect and connect again with the box unticked.

Available tools

With an API key you get all of these. Over OAuth you get the read tools, plus the triage and scan tools if you allowed changes; list_organizations is API-key only. The triage and scan tools need the member role or above in the organization: a viewer doesn't see them, whatever the connection.

  • get_summary — your org's score and open findings by severity, at a glance.
  • list_assets — assets under protection.
  • list_findings — findings, filterable by severity, status, engine (code, web, deps, cloud, dns, contract), type, or updated_since.
  • get_finding — one finding's full detail, proof, and fix.
  • update_finding — triage a finding and optionally attach a note (e.g. the PR that fixed it).
  • bulk_update_findings — set one status on up to 1,000 findings at once, with an optional note. Ids from other orgs, or already at that status, are skipped; the result says how many changed.
  • trigger_scan — launch scans of your org's own assets, at most 10 calls an hour per organization.
  • list_scans / get_scan — check scan status; poll after triggering one.
  • list_organizations — the org this key is scoped to.
ℹ
The list tools (list_findings, list_scans, list_assets) paginate with limit (default 25, max 100) and offset. Each result includes total and has_more — raise offset by limit to walk every page.
ℹ
On the free plan, get_summary and list_findings return your counts plus the single most-severe open finding, with upgrade_required: true and an upgrade_url; the other findings stay locked. The finding detail, triage and scan tools need a paid plan.

Tool results

Every tool returns structured JSON (structuredContent, described by the tool's output schema), and the same JSON as text for clients that only read text. Each tool is annotated as read-only or not: the read tools only look at your org's data, update_finding changes one finding, and bulk_update_findings and trigger_scan are flagged as destructive, so clients that honour annotations ask you before running them. A bulk update overwrites up to 1,000 statuses at once, and an active web scan can submit forms on your live app.

Every triage or scan call that reaches Muster's API is recorded in Muster's audit log, whether it succeeds or the API refuses it (for example, an id outside your org or the scan rate limit). Calls turned away earlier, such as a token without the write scope or invalid arguments, are not. Each entry records who made it, whether it came from an API key or OAuth (with the client and scopes), the tool, the finding or asset ids it targeted, and the result.

Text that comes from what Muster scanned — finding titles, descriptions, code snippets, file paths, AI write-ups, and asset and scan names — sits under an untrusted key. Anyone who can put content into a scanned repo or site controls that text, so the tool descriptions tell your agent to treat it as data and never follow instructions in it. It is also cleaned before it reaches the agent: invisible and direction-changing Unicode characters, control characters, and terminal escape codes are removed; snippets are capped at 2,048 characters and other fields at 4,096.

The fix loop

The tools are designed to close the loop from your editor: read a proven finding, fix it in the repo, then mark it fixed — or trigger a scan to auto-verify.

example prompts
"What are my open critical Muster findings? Fix the first one."
"Pull the Muster finding for the IDOR on /api/orders, apply the recommended fix here, then mark it fixed."
"Fix all high-severity broken-access-control findings, then trigger a scan to verify."

Self-hosting the server (advanced)

Running your own Muster instance from the source repo? A standalone stdio server ships at mcp/server.mjs. It talks to the same API over the network, so you can run it locally against any Muster deployment:

.mcp.json (stdio)
{
  "mcpServers": {
    "muster": {
      "command": "node",
      "args": ["/absolute/path/to/muster/apps/dashboard/mcp/server.mjs"],
      "env": {
        "MUSTER_API_URL": "https://app.mustersecurity.com",
        "MUSTER_API_KEY": "mus_xxxxxxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}
⚠
Most users want the hosted setup above — it needs no local file, no Node, and no repo access. The stdio server is only for self-hosters who already have the source checked out.

Security notes

  • The key only reaches the org it's scoped to, with your role there — never your other orgs.
  • Store the key in your MCP client's config; treat it like a password. Revoke anytime from Account → API keys.
  • An OAuth (ChatGPT) connection reaches only the organization you picked on the consent screen, and only with findings:read.
  • Finding text can contain instructions planted in scanned code. Muster fences it under untrusted, but review what your agent proposes before it changes findings or launches scans.

Support

Questions or trouble with the MCP server or the ChatGPT plugin? Email support@mustersecurity.com. Tell us which client you use (ChatGPT, Claude Code, Codex, Cursor or Claude Desktop), roughly when it happened, and the error you saw. Never send your API key or tokens.