magicsword-mcp

magicsword-mcp

A Model Context Protocol server for MagicSword. Lets users running Claude Desktop, Cursor, or any MCP-aware client manage MagicSword conversationally — list endpoints, query alerts, triage findings, mint enrollment tokens, and preview / commit policy changes.

The server is a thin, opinionated wrapper around the /api/public/v1/* customer API exposed by the Magic Portal. It runs on the user's machine, holds an msk_… API key, and speaks MCP over stdio.

Plan gate. The customer API is Enterprise-only. With a pilot-tier key the server returns a clear "this MagicSword org needs the Enterprise plan to use MCP" message; nothing else works until the org is upgraded.

Quick start

Claude Desktop installer

For Claude Desktop on macOS or Windows, download magicsword-mcp.mcpb from GitHub Releases. The installer is included starting with version 0.1.4. If that release is not available yet, use terminal setup below.

1. Generate an Enterprise API key in Magic Portal → Settings → API Keys. Start with read scopes for endpoints, alerts, and policies.
2. In Claude Desktop, open Settings → Extensions → Advanced settings → Install Extension…, then select the downloaded .mcpb file.
3. Review the extension and enter your API key in its secure configuration field. Claude stores this sensitive field using the operating system's secure storage. The installer does not save a separate MagicSword key file.
4. Start a new conversation, then ask MagicSword to call whoami followed by list_endpoints. Confirm the organization and scopes.

The installer includes the server and production dependencies and uses Claude Desktop's bundled Node runtime (22+). It requires no npm installation or JSON editing. Its SHA256 checksum is published beside the installer. A privately downloaded extension must be updated by installing a newer .mcpb file; automatic directory updates require a reviewed directory listing.

Terminal setup

Requires Node.js 22+, an installed MCP client, and a MagicSword Enterprise API key from Magic Portal → Settings → API Keys.

Recommended: install and connect your client in one command. Setup prompts for your key privately and lets you choose Claude Desktop, Codex, Claude Code, or Cursor:

npm install -g @magicsword-io/magicsword-mcp && magicsword-mcp configure

Already know your client? Run the matching one-liner:

Client	Install and configure
Claude Desktop	npm install -g @magicsword-io/magicsword-mcp && magicsword-mcp configure --client claude-desktop
Cursor	npm install -g @magicsword-io/magicsword-mcp && magicsword-mcp configure --client cursor
Claude Code	npm install -g @magicsword-io/magicsword-mcp && magicsword-mcp configure --client claude-code
OpenAI Codex	npm install -g @magicsword-io/magicsword-mcp && magicsword-mcp configure --client codex

Prefer your client's own registration command? After saving your key locally, these also work without a global server install:

codex mcp add magicsword --env MAGICSWORD_BASE_URL=https://www.magicsword.io -- npx -y @magicsword-io/magicsword-mcp
claude mcp add --scope user --transport stdio magicsword --env MAGICSWORD_BASE_URL=https://www.magicsword.io -- npx -y @magicsword-io/magicsword-mcp

These are also a fallback when automatic registration cannot launch a client's CLI on Windows. Desktop launch still requires Node/npx to be available to that client; the recommended setup uses absolute paths instead.

Install buttons: Cursor and VS Code

First save your API key locally (once per machine):

npm install -g @magicsword-io/magicsword-mcp && magicsword-mcp configure --client manual

Then click your client's button and approve its installation prompt:

These buttons register a local npx server pointing at production. They contain no API key; the server reads ~/.magicsword/mcp.json. They require npx on the client's PATH. If a desktop app cannot find Node or npx, use the recommended setup above for absolute launch paths, or copy the absolute command and args printed by configure --client manual into its configuration.

Manual JSON fallback

After saving your key above, merge this entry into Claude Desktop's claude_desktop_config.json or Cursor's ~/.cursor/mcp.json, preserving your other servers:

{
  "mcpServers": {
    "magicsword": {
      "command": "npx",
      "args": [
        "-y",
        "@magicsword-io/magicsword-mcp"
      ],
      "env": {
        "MAGICSWORD_BASE_URL": "https://www.magicsword.io"
      }
    }
  }
}

For VS Code / Insiders, run MCP: Open User Configuration and merge this configuration (VS Code uses servers, not mcpServers):

{
  "servers": {
    "magicsword": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "@magicsword-io/magicsword-mcp"
      ],
      "env": {
        "MAGICSWORD_BASE_URL": "https://www.magicsword.io"
      }
    }
  }
}

Restart your client or start a new session, then ask it to call whoami and list_endpoints. Both must succeed to verify authentication and read access.

The official MCP Registry identifier is io.github.magicsword-io/magicsword-mcp. Client-specific details follow below.

Configure

magicsword-mcp configure

Setup asks for an API key with hidden input (mint one in Magic Portal → Settings → API Keys) and which client to connect: Claude Desktop, Codex, Claude Code, Cursor, or manual configuration. It saves the key only in ~/.magicsword/mcp.json (mode 600). The Portal defaults to https://www.magicsword.io, replacing any saved preview origin. A custom origin requires an explicit --base-url option.

Setup registers the selected client at its standard user-level location, preserves other servers and settings, and creates a private backup of existing client configuration. It updates an existing magicsword entry for JSON clients and Codex. Claude Code requires that name to be available; use its own MCP commands to remove an old registration before retrying. Codex and Claude Code registration require their respective CLIs to be installed.

To choose a client directly, or reuse a saved key without prompting:

magicsword-mcp configure --client codex
magicsword-mcp configure --non-interactive --client claude-desktop
magicsword-mcp configure --client cursor --client-config /path/to/mcp.json

The launch entry uses absolute Node and installed server paths so desktop apps can start it without inheriting your terminal's PATH. Re-run setup if you move your Node installation or install the package under a different Node version. Manual setup prints a credential-free launch entry for other local stdio MCP clients. Unattended setup without --client writes only the MagicSword file.

Restart the selected client or start a new session to load magicsword and its 22 tools. Ask it to call whoami and then list_endpoints to verify access; discovery alone does not verify API-key scopes. Claude Desktop logs are at ~/Library/Logs/Claude/mcp*.log on macOS. Browser-only clients require a remote MCP transport; this package provides a local stdio server.

If you need a per-client override instead of ~/.magicsword/mcp.json, set MAGICSWORD_API_KEY and optionally MAGICSWORD_BASE_URL in that MCP host's environment. Keep API keys out of shared config snippets and screenshots.

Tools

Tool	What it does
whoami	Returns organization, key id, and scopes. Always call first when troubleshooting.
list_endpoints	Lists endpoints with optional platform / status / hostname-glob filtering.
show_endpoint	Fetches one endpoint with OS, policy, compliance, upgrade, heartbeat, and AMSI details.
list_agent_releases	Lists available agent versions and per-platform latest releases.
find_alerts	Searches alerts by severity / ack state / since / hostname / MITRE technique.
list_events	Searches Investigate events by name, filename, path, hash, publisher, computer, or package family, with type/status/policy/endpoint/time filters.
triage_alert	Fetches one alert directly with triggering events, metadata, and file context; optionally acknowledges or dismisses it.
list_policies	Lists policies with current version + mode.
show_policy	Shows one policy by id.
manage_policy_rules	Lists rules or adds explicit / event-derived rules to a policy by id or name. Avoid explicit flat hash rules for Windows WDAC.
apply_policy_to_endpoints	Resolves endpoints and previews a policy assignment; confirm=true applies it after approval.
flip_to_enforcing	Two-step preview/confirm flip with a server-issued one-time confirm token. Safety-critical.
list_intel_sources	Discovers enabled global and organization-owned intel sources and IDs, with platform and pagination filters.
list_customer_intel_feeds	Lists private intel feeds or feed items.
manage_customer_intel_feed	Creates or updates feeds; deletion requires a preview followed by confirm=true.
manage_customer_intel_item	Edits feed items; deletion requires a preview followed by confirm=true.
upsert_customer_intel_items	Adds indicators extracted from reports into a private feed.
manage_policy_intel_sources	Lists, attaches, or detaches intel feeds on a policy.
upgrade_endpoints	Previews upgrades for an endpoint selection; confirm=true queues them after approval.
request_endpoint_checkin	Queues an endpoint check-in command.
mint_enrollment_token	Mints a one-time agent enrollment token.
agent_install_instructions	Returns the install one-liner for macOS / Linux / Windows. No API call.

Write tools require matching Customer API scopes in Magic Portal, such as alerts:write, policies:write, endpoints:write, or intel:write. Event review uses alerts:read; turning selected events into policy rules uses policies:write.

Tools publish standard MCP safety annotations. Read-only discovery tools are marked read-only and idempotent; enforcement, deletion, policy assignment, rule changes, and agent upgrades are marked destructive so MCP clients can apply appropriate confirmation UX.

Adding and listing policy rules

manage_policy_rules accepts one input source per add: rules for explicit rules, or event_ids for rules derived from telemetry. Each accepts 1–100 items. Use policy_id, or policy_name with an optional platform to resolve a name. The policy must already exist; ambiguous names require an ID or platform.

For example, to block the filenames from issue #5 on a Windows blocklist policy:

{
  "action": "add",
  "policy_name": "Workstations",
  "platform": "windows",
  "rules": [
    { "type": "filename", "value": "rustdesk.exe", "status": "blocked" },
    { "type": "filename", "value": "mshta.exe", "status": "blocked" },
    { "type": "filename", "value": "wscript.exe", "status": "blocked" }
  ]
}

Explicit rules use type and value, with optional status (defaults to blocked). Unknown rule fields are rejected; rule_type / file_name are not substitutes. Filename and path values must be non-empty strings or supported value objects. Publisher values require publisherName (or publisher) and a valid tbsHash (or tbs_hash / certificate_thumbprint). Optional rule metadata includes name, mode, version bounds, signer/file-attribute references, package_version, and target_platforms. The Portal validates whether a rule can be enforced on the resolved policy's platform. Flat Windows WDAC hashes remain unsupported; use telemetry-derived rules or private intelligence feeds.

For telemetry, use event_ids with top-level status (defaults to allowed) and an optional type override. Top-level status and type cannot be used with explicit rules; set those fields on each rule instead.

Adds preserve existing rules and intelligence-source links. Matching customer rules with a different status are updated; rules already at the requested status are reported as unchanged duplicates. Feed-owned rules retain their provenance; unsupported conflicts with a feed rule return an API error. The tool reports the Portal's actual added_count, updated_count, and duplicate_count, and retains its version information in structured output. An unchanged duplicate batch reports “No changes needed.” Unexplained zero changes or malformed responses are errors: inspect the current rules/version before retrying because the request may have created a version. Writes are never retried automatically. For event-derived batches, a count total smaller than the number of unique requested events is also an error: some events may not have produced rules. Successful changes and version details remain available in structured output.

To verify the current version's stored rules:

{ "action": "list", "policy_id": "<policy UUID>", "limit": 500, "offset": 0 }

List requires an ID and policies:read. It returns stored customer rules and materialized feed rules for the current version, with total, limit, and offset. The default page size is 100, maximum 500; increase offset until all total rows have been read. An empty page does not describe other versions or indicators that have not been materialized into this version.

Example transcript

User:   Show me unack'd critical alerts from the last 24h.
Claude: [calls find_alerts severity=critical acknowledged=false since=...]
        → 3 critical alerts on 2 endpoints. Want me to triage the top one?

User:   Yes.
Claude: [calls triage_alert]
        → Evidence chain: cmd.exe spawned by winword.exe with -enc base64.
          MITRE T1059.001. Acknowledge with comment "office macro chain — under investigation"?

User:   Yes, and extract the IOCs from this report into our SOC feed.
Claude: [calls upsert_customer_intel_items]
        → Upserted 47 indicators into feed 8d6...

User:   Show me audited or blocked events from the last day.
Claude: [calls list_events hours=24 status=audited,blocked]
        → 31 events. Here are the file paths, publishers, and hashes.

User:   Allow the first 5 on the Workstations policy.
Claude: [calls manage_policy_rules action=add policy_name=Workstations event_ids=[...] status=allowed]
        → Created a new policy version with 5 allowed rules.

Investigate events and attach intelligence

Ask “Show audited filename events for image.png from the last week.” The agent calls list_events with type=filename, status=audited, file_name=image.png, and hours=168. file_name is an exact, case-sensitive match, including literal punctuation. For broader discovery, q searches the same seven fields as Portal Investigate: name, filename, path, hash, publisher, computer, and package family. It is a case-insensitive literal substring search accepting up to 120 characters, including Unicode, quotes, commas, parentheses, and Windows backslashes. %, _, and * are literal characters; wildcard and field-query syntax are not supported. Control characters are rejected. A blank query applies no text filter. Filters combine, so type=filename selects filename events rather than all event types that happen to include a filename. Omit type to search across event types.

For example, “Find events from publisher Contoso, Inc. (R&D)” uses list_events q="Contoso, Inc. (R&D)". To block a file found in those results, review the returned IDs and rule context, choose the intended rule type, and call manage_policy_rules action=add policy_id=... event_ids=["<selected-event-id>"] status=blocked type=filename. Search does not change policies or automatically block every match. Searching a publisher does not imply blocking that publisher; the selected rule type determines what is blocked. Event IDs can also derive other supported rules when appropriate, subject to Portal platform validation.

The default window is 24 hours with statuses audited,blocked; use since for an explicit lower bound. The window uses the last observation time, falling back to creation time. Results include file paths, publishers, hashes, computers, occurrence counts, rule context, and pagination. Use limit and offset for additional pages. This is normalized event data rather than a raw event log.

For “Attach LOLDrivers intel to my Windows policy”, the workflow is:

1. list_policies platform=windows identifies the policy name and ID.
2. list_intel_sources platform=windows returns available source names and IDs. It defaults to enabled sources; use is_enabled=false to inspect disabled ones. Use limit (up to 100) and offset to discover additional pages.
3. manage_policy_intel_sources action=list policy_id=... shows current attachments.
4. After identifying the intended policy and sources, use manage_policy_intel_sources action=attach policy_id=... source_ids=[...]. The result reports new attachments and sources already attached separately.
5. List the policy attachments again to verify the result.

Source discovery requires intel:read, policy discovery/attachment listing requires policies:read, and attachment requires policies:write. Source visibility does not guarantee premium entitlement; the API checks it when attaching. Attachments create a policy version and take effect as endpoints receive it.

Configuration sources, in order of precedence

1. MAGICSWORD_API_KEY / MAGICSWORD_BASE_URL env vars (set by the MCP host).
2. ~/.magicsword/mcp.json (or $MAGICSWORD_CONFIG).
3. Default base URL https://www.magicsword.io.

Optional transport controls are MAGICSWORD_REQUEST_TIMEOUT_MS (default 30 seconds), MAGICSWORD_RESPONSE_MAX_BYTES (default 4 MiB), and MAGICSWORD_GET_RETRIES (default 2, maximum 3). Only idempotent GET requests are retried; write actions are never retried automatically.

Safety notes

• flip_to_enforcing is two-step. The first call returns a server preview and one-time confirmation token; you must show the preview to a human and pass the token back to commit. Tokens are short-lived and single-use. This is not MFA or server-verified human approval: the same API key can perform both calls. The MCP host must obtain human approval before committing. A commit response must confirm success for the requested policy and identify its new version; otherwise use show_policy to verify the outcome before retrying. Success confirms the server policy version change. Endpoint enforcement takes effect through check-in; this response does not confirm endpoint delivery.
• Intel attachments use enabled sources in batches of at most 60. Capture requests share the API key creator's rate budget with the Portal UI. A 429 response includes retry guidance; wait before submitting another batch. Attachment summaries report newly attached and already-attached counts separately. Invalid or incomplete counts require listing attachments before retrying.
• Fleet changes and deletion are explicit. Endpoint upgrades, policy assignments, and private-intel deletion return a no-op preview unless the approved follow-up call includes confirm=true.
• Windows WDAC policy edits should not use explicit flat file hashes. Use manage_policy_rules with event_ids when approving audit events so the Portal can derive supported path / publisher / filename rules, or use upsert_customer_intel_items for hash, AuthentiHash, page-hash, and TBS intelligence in a private feed.
• Secrets stay on the user's machine. The MCP server is a local stdio process; the API key is read from ~/.magicsword/mcp.json (mode 600) or an env var passed by the MCP host. Nothing is sent off-machine except the requests to the configured MAGICSWORD_BASE_URL.
• Remote Portal URLs must use HTTPS. Plain HTTP is accepted only for localhost development. Configured URLs cannot contain credentials, paths, query strings, or fragments.

Develop

npm install
npm run build
npm test
npm run test:package
npm run release:verify
node dist/index.js --version

Tests validate --help, --version, production-default migration, private client registration, preservation and rollback, configure validation, secret redaction, modern and legacy MCP startup, focused mutation confirmation, retry and timeout semantics, response-size bounds, malformed responses, write non-retry behavior, all 22 tools' HTTP contracts and confirmation flows over modern and legacy stdio, explicit filename round trips, invalid rule shapes, ambiguous inputs, actual mutation counts, and npm pack contents. test:package installs a production-only tarball and checks the installed CLI and MCP tool discovery; release:verify checks all release metadata. See docs/RELEASING.md for npm and MCP Registry releases.

An opt-in live policy test is also available after npm run build:

MAGICSWORD_LIVE_TEST=1 MAGICSWORD_TEST_POLICY_ID="<disposable-policy-uuid>" node scripts/live-policy-rules.mjs

Use your private saved credentials or MAGICSWORD_API_KEY; do not put keys in commands or shared logs. This test writes policy versions and requires a disposable, unassigned Windows audit policy named with the prefix MCP issue 5 disposable verification, seeded with more than 1,000 baseline rules with at least one attached intelligence source, and no existing rustdesk.exe, mshta.exe, or wscript.exe customer rules. It verifies append behavior, pagination, intelligence-source preservation, repeat adds, status updates, invalid inputs, and event-derived filenames when suitable telemetry exists. It also exercises discovery and read tools through stdio. It prints assertion names/counts only. Remove the disposable policy and its versions after inspection; the test does not create or delete the fixture. Set MAGICSWORD_TEST_PROTOCOL=legacy to exercise the legacy live transport.

With the same disposable policy and opt-in environment, run node scripts/live-demo-tools.mjs to check assignment, upgrade, and enforcement previews, all three platforms' installer instructions, and a disabled temporary feed's create/update/upsert/list/item-delete/attach/detach/feed-delete flows. The feed is removed in finally; the policy fixture still requires manual cleanup. These checks never commit endpoint assignments, upgrades, or enforcement. Actual endpoint commands, enrollment tokens, and alert acknowledgements/dismissals are covered by the local HTTP fixtures.