MISP威胁情报查询管理MCP服务器

misp-mcp

Query and manage your MISP threat-intelligence platform from any MCP client.

Website & docs → lidless.dev/misp-mcp

misp-mcp is a Model Context Protocol server for MISP, the open-source threat-intelligence sharing platform. It lets an LLM client such as Claude run IOC lookups, manage events, discover cross-event correlations, and export indicators directly against your own MISP instance. Unlike a generic HTTP wrapper, it ships 42 purpose-built tools, MISP-aware resources and prompts, and a confirmation gate that refuses destructive writes (delete, publish, untag) unless you explicitly approve them.

What it does

misp-mcp connects an AI agent to a MISP (Malware Information Sharing Platform & Threat Intelligence Sharing) instance over MISP's REST API and exposes it as Model Context Protocol tools, resources, and prompts. Point it at your MISP server with an API key and an LLM can search threat-intelligence events, look up and add indicators of compromise (IOCs), correlate indicators across events, attach MITRE ATT&CK galaxy clusters, check warninglists for false positives, and export IOCs as Suricata, Snort, STIX, CSV, RPZ, or hash lists. The server is read-only by default: every tool that changes MISP state refuses to run until the operator names it in MISP_ALLOWED_MUTATIONS. Destructive and publishing operations (delete, publish, untag, disabling a feed, false-positive and expiration sightings) additionally need a per-call confirmation flag, and bulk writes need confirmation above a small size.

• 42 MCP tools covering events, attributes, correlations, tags, exports, sightings, warninglists, objects, galaxies, feeds, organisations, and server administration.
• 3 MCP resources for browsing attribute types, instance statistics, and available taxonomies.
• 3 MCP prompts for guided IOC investigation, incident event creation, and threat reporting.
• Read-only by default: state-changing tools run only when listed in MISP_ALLOWED_MUTATIONS; each also accepts dryRun: true to preview a call without sending it.
• Confirmation-gated destructive actions: delete, publish, untag, feed disable, and false-positive/expiration sightings refuse to run without confirm: true (and confirmHard: true for permanent deletes).
• Untrusted-data fencing: MISP content returned to the model is wrapped in <untrusted_misp_data> markers, and every tool declares MCP annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint).
• MITRE ATT&CK integration via galaxy cluster search and attachment.
• Export formats including CSV, STIX, Suricata, Snort, text, RPZ, and hash lists.
• SSL flexibility for the self-signed certificates common in on-prem MISP deployments.

Quickstart

No checkout required. With Node.js 20+ installed, register the published package with any MCP client. For Claude Code:

claude mcp add misp \
  --env MISP_URL=https://misp.example.com \
  --env MISP_API_KEY=your-api-key-here \
  --env MISP_VERIFY_SSL=false \
  -- npx -y misp-mcp

Add --scope user to make it available from any directory instead of only the current project.

The server starts read-only. To let the agent write, add --env MISP_ALLOWED_MUTATIONS=misp_create_event,misp_add_attribute (the tools you want enabled) or --env MISP_ALLOWED_MUTATIONS='*'; see Mutation gate and destructive action safety.

Before wiring a client, verify that the package can reach MISP with the same environment:

MISP_URL=https://misp.example.com \
MISP_API_KEY=your-api-key-here \
MISP_VERIFY_SSL=false \
npx -p misp-mcp mispctrl status

This calls /servers/getVersion and prints the MISP version, recommended PyMISP version, TLS setting, and target URL. If this fails, fix the URL, API key, or TLS setting before debugging the MCP client.

MCP client config (copy-paste)

For Claude Desktop, Cursor, or any client that reads a JSON mcpServers block, add:

{
  "mcpServers": {
    "misp": {
      "command": "npx",
      "args": ["-y", "misp-mcp"],
      "env": {
        "MISP_URL": "https://misp.example.com",
        "MISP_API_KEY": "your-api-key-here",
        "MISP_VERIFY_SSL": "false"
      }
    }
  }
}

Claude Desktop reads this from ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows). Restart the client after editing.

Prerequisites

• Node.js 20 or later.
• A running MISP instance with API access.
• A MISP API key (MISP UI: Administration > List Auth Keys).

Configuration

Boolean variables (MISP_VERIFY_SSL, MISP_ALLOW_DESTRUCTIVE) share one parser: true, 1, false, 0, case-insensitive after trimming; unset or empty means the default. Any other value is rejected at startup with an error that names the variable. The MCP guard and the CLI gate use the same parser, so MISP_ALLOW_DESTRUCTIVE=1 means the same thing in both.

Mutation gate and destructive action safety

The server starts read-only. Read tools always work. A tool that changes MISP state returns a Refused: error and sends no request unless the operator lists it in MISP_ALLOWED_MUTATIONS (or sets it to *). The allowlist is read from the server's environment, so a model cannot change it.

MISP_ALLOWED_MUTATIONS=misp_create_event,misp_add_attribute,misp_tag_event

State-changing tools (gated by MISP_ALLOWED_MUTATIONS): misp_create_event, misp_update_event, misp_add_event_report, misp_publish_event, misp_tag_event, misp_delete_event, misp_add_attribute, misp_update_attribute, misp_add_attributes_bulk, misp_delete_attribute, misp_add_sighting, misp_add_object, misp_delete_object, misp_attach_galaxy_cluster, misp_toggle_feed, misp_fetch_feed, misp_cache_feed.

Every one of them accepts dryRun: true. A dry run validates the arguments, reports whether the tool is currently allowed, and sends nothing to MISP, so it works in read-only mode too.

Also require confirm: true (or MISP_ALLOW_DESTRUCTIVE=true):

• misp_delete_event
• misp_delete_attribute
• misp_delete_object
• misp_publish_event
• misp_tag_event when remove: true
• misp_toggle_feed when enable: false
• misp_add_sighting with type 1 (false positive) or 2 (expiration)
• misp_add_attributes_bulk with more than 10 attributes

Setting MISP_ALLOW_DESTRUCTIVE=true pre-authorizes the confirmations above so the per-call confirm flag can be omitted (useful for trusted automation). It does not enable any tool by itself; the tool must also be in MISP_ALLOWED_MUTATIONS.

Permanent hard deletes (hard: true on misp_delete_attribute / misp_delete_object) require a second confirmation, confirmHard: true, in addition to confirm: true. The env opt-in does not bypass confirmHard.

Future destructive CLI commands use the same policy: MISP_ALLOW_DESTRUCTIVE=true plus --confirm --destructive, with --confirm-hard still required for permanent deletes.

Untrusted MISP content

Event titles, comments, IOC values, tags, export bodies, and similar fields are written by other organizations. Tool results wrap them in <untrusted_misp_data> markers, the JSON resources carry them under an untrusted_misp_data key beside a notice so they stay valid JSON, any marker-like text inside the content is escaped, and tool descriptions and the server instructions tell the model to treat the content as data. Prompt arguments are delimited the same way.

Bounds

• Responses are read through a strict byte cap (MISP_MAX_RESPONSE_BYTES) and the request timeout stays active until the body is fully read.
• misp_add_attributes_bulk takes at most 100 attributes and stops after five consecutive failures; the result lists succeeded, failed, and skipped items and warns about partial updates.
• Search tools take an integer limit from 1 to 500 and an integer page from 1; fractional, zero, negative, and oversized values are rejected. When a page comes back full, the result ends with a note that more results may exist and names the next page.
• misp_get_related_events accepts maxAttributes up to 50 and perAttributeLimit up to 500. Tag lists are capped at 20 and object attribute lists at 100.
• Upstream error bodies are not returned to the model. Errors carry the status class, and for 4xx responses a short, redacted message field. The API key is redacted from every error.

A guarded call returns an error (isError: true) with a Refused: message and performs no MISP request.

Tools

Verified against the server source (42 tools, registered in src/tools/).

Event tools (6)

Event report tools (3)

Attribute tools (6)

Correlation & intelligence tools (3)

Tag & taxonomy tools (2)

Export tools (2)

Sighting & warninglist tools (3)

Object tools (4)

Galaxy tools (4)

Feed tools (4)

Organisation tools (2)

Server & admin tools (3)

Resources

Prompts

Usage examples

Search for an IOC

"Search MISP for the IP address 203.0.113.50"

Uses misp_search_events and misp_search_attributes to find all events and attributes referencing this IP.

Investigate a suspicious domain

"Investigate suspicious-domain.example in MISP"

Triggers the investigate-ioc prompt workflow: searches for the domain, checks correlations, queries warninglists, and provides a structured threat assessment.

Create an incident event

"Create a MISP event for a phishing campaign targeting our finance team. The phishing emails came from attacker@phish.example and linked to https://harvest.phish.example/login"

Uses misp_create_event followed by misp_add_attributes_bulk to create a fully populated event.

Export Suricata rules

"Export all IOCs from the last 7 days as Suricata rules"

Uses misp_export_iocs with format "suricata" and last "7d".

Check for false positives

"Is 192.0.2.123 on any MISP warninglists?"

Uses misp_check_warninglists to verify if the value is a known benign indicator.

Find MITRE ATT&CK techniques

"Search for phishing techniques in MITRE ATT&CK"

Uses misp_search_galaxy_clusters to find relevant ATT&CK techniques, then misp_attach_galaxy_cluster to link them to events.

Add structured objects

"Add a file object to event 1 with filename encrypt.exe, SHA256 hash, and file size"

Uses misp_add_object with the "file" template to create a structured group of related attributes.

Supported attribute types

Use misp_describe_types for the complete list of supported types and categories.

Other clients & local development

git clone https://github.com/lidless-labs/misp-mcp.git
cd misp-mcp
npm install
npm run build

claude mcp add misp \
  --env MISP_URL=https://misp.example.com \
  --env MISP_API_KEY=your-api-key-here \
  --env MISP_VERIFY_SSL=false \
  -- node /absolute/path/to/misp-mcp/dist/mcp-bin.js

openclaw mcp set misp '{
  "command": "npx",
  "args": ["-y", "misp-mcp"],
  "env": {
    "MISP_URL": "https://misp.example.com",
    "MISP_API_KEY": "your-api-key-here",
    "MISP_VERIFY_SSL": "false"
  }
}'

systemctl --user restart openclaw-gateway
openclaw mcp list   # confirm "misp" is registered

codex mcp add misp \
  --env MISP_URL=https://misp.example.com \
  --env MISP_API_KEY=your-api-key-here \
  --env MISP_VERIFY_SSL=false \
  -- npx -y misp-mcp

mcp_servers:
  misp:
    command: "npx"
    args: ["-y", "misp-mcp"]
    env:
      MISP_URL: "https://misp.example.com"
      MISP_API_KEY: "your-api-key-here"
      MISP_VERIFY_SSL: "false"

docker build -t misp-mcp .
docker run -e MISP_URL=https://misp.example.com -e MISP_API_KEY=your-key -e MISP_VERIFY_SSL=false misp-mcp

MISP_URL=https://misp.example.com MISP_API_KEY=your-key node dist/mcp-bin.js
MISP_URL=https://misp.example.com MISP_API_KEY=your-key node dist/cli.js status

MISP_URL=https://misp.example.com MISP_API_KEY=your-key npm run dev

Testing

npm test                  # Unit tests (mocked)
npm run test:integration  # Integration tests against a live MISP instance
npm run test:watch        # Watch mode
npm run lint              # Type check

Integration tests require MISP_URL, MISP_API_KEY, and optionally MISP_VERIFY_SSL=false environment variables.

Project structure

misp-mcp/
  src/
    cli.ts                # mispctrl CLI entry point
    mcp-bin.ts            # MCP stdio entry point
    mcp-server.ts         # Shared MCP server factory
    index.ts              # Library export surface
    config.ts             # Environment config + validation
    client.ts             # MISP REST API client
    guards.ts             # Mutation allowlist + destructive-action confirmation guards
    tool-specs.ts         # Per-tool annotations and mutation flags
    tool-registry.ts      # Tool registration: gate, fence, error redaction
    untrusted.ts          # <untrusted_misp_data> fencing helpers
    limits.ts             # Bulk and fan-out maxima
    types.ts              # MISP API type definitions
    resources.ts          # MCP resources
    prompts.ts            # MCP prompts
    tools/                # One module per tool family (events, attributes, ...)
  tests/                  # Unit + integration tests
  Dockerfile
  package.json

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md for the contribution path and SECURITY.md for how to report a vulnerability privately. By participating you agree to the Code of Conduct.

Why not the MISP web UI or raw API?

The MISP web UI and PyMISP are excellent for analysts working a console. misp-mcp is for the case where the operator is an LLM agent, not a human at a browser. It gives the model a typed, named tool for each common MISP operation, validates inputs with Zod, and shapes responses so a model can reason over them, rather than handing the model a raw REST endpoint and an OpenAPI dump. It also adds a guardrail layer the bare API does not: destructive operations fail closed unless the caller opts in, so an over-eager agent cannot publish or delete intelligence by accident. If you want a human-driven console, use the MISP UI. If you want PyMISP scripting, use PyMISP. Use misp-mcp when you want a chat or agent client to drive MISP safely.

What misp-mcp is not

• Not a MISP server. It talks to an existing MISP instance over the REST API; it does not store, host, or replace MISP.
• Not a general HTTP proxy. It exposes a curated set of MISP operations as MCP tools, not arbitrary passthrough to every MISP endpoint.
• Not a SIEM, EDR, or alerting system. It reads and writes threat intelligence; it does not collect logs or generate alerts on its own.
• Not a credential vault. It reads MISP_URL and MISP_API_KEY from the environment; manage and rotate those keys with your own secrets tooling.
• Not an autonomous deleter. Destructive and publishing actions are confirmation-gated by design.

Container root access

The installer creates no guest root password and locks password login after starting the container. Manage the guest from its Proxmox host with pct enter. Service API credentials continue to use protected stdin delivery and root-only storage.

License

MIT

Part of Lidless Labs · the eye does not close

_{Security / SOC: soc-stack · wazuh-mcp · suricata-mcp · thehive-mcp · cortex-mcp · mitre-mcp · zeek-mcp · hotwash}

_{All tools · Lidless Labs on GitHub}