mitre-cwe-mcp-server

Public Hosted Server: https://mitre-cwe.caseyjhand.com/mcp

Overview

MITRE's CWE and CAPEC catalogs connect software and hardware weaknesses with attack mechanics. Search root causes, inspect mitigations and code examples, navigate CWE views and live hierarchies, and follow recorded CWE → CAPEC → ATT&CK mappings.

Tools

Related MCP server: mcp-mitre-cwe

Capability reference

cwe_list_reference tool

• topic: views, abstraction_levels, mitigation_phases, platforms, relationship_natures, filter_values, workflow, or versions; optional all-token filter on labels, using the word-start and bounded-prefix rule below (script finds JavaScript)

• views lists every CWE view with its direct member count, and a filter-defined view's filterDefinition

• filter_values lists weakness statuses, mapping usage, and taxonomy names, consequence impacts, and CAPEC statuses, severity, and likelihood

• Lists page with limit (default 20, max 100) and pagination.nextCursor; versions adds each loaded catalog's version, date, counts, and refresh errors

cwe_get_weakness tool

• 1–200 CWE IDs (CWE-79, CWE79, cwe 079, CWE_79, or 79); up to 20 distinct items are processed per call and the rest return in deferredIds. A malformed item comes back in notFound without failing the batch; only an item holding a control character other than whitespace, or a non-ID item over 150 characters, rejects the call

• sections: summary, consequences, mitigations, relations, examples, detection, observed, platforms, mappings, notes (default: the first four); [] returns an outline

• summary includes technical nuance, typed dimensions, recorded alternate terms and roles, ordinalities (Primary, Resultant, Indirect), and background details, platforms adds functional areas and affected resources, and relations names each related weakness. notes returns every note MITRE records for weaknesses, categories, and groupings, such as Relationship, Maintenance, Terminology, and Research Gap notes

• entries[].kind is full, partial (default sections that did not fit the response budget are listed in outline with their sizes, and notice groups re-calls by their complete response size, keeping example parameters only on the examples group), or outline; notFound[].reason is not_found, wrong_type (a view, which belongs to cwe_get_view, or a prefixed ID of the other catalog), or invalid_id (an item that is no ID, echoed trimmed; NVD-CWE-noinfo and NVD-CWE-Other get their own guidance). Categories and groupings resolve here with distinct record types

• Examples filter by exampleLanguage and page with exampleOffset (default 0), exampleLimit (default 3, no fixed cap), and examplePage.nextOffset. With sections omitted, setting any of the three to a non-default value adds examples after the defaults (outlined with its page size and a re-call when it does not fit); explicit sections without examples ignore them, and guidance gives the re-call that adds examples. Each example keeps its ordered prose and code, with code laid out as MITRE's pages show it (4 spaces per indentation level), and lists the references it cites in references

cwe_get_attack_pattern tool

• 1–200 CAPEC IDs (CAPEC-66, CAPEC66, capec_066, or 66), Deprecated and Obsolete included; up to 20 distinct items per call, the rest in deferredIds

• sections: summary, prerequisites, execution, skills, consequences, mitigations, examples, relations, mappings, notes (default: summary, prerequisites, execution, mitigations, and mappings for patterns; summary and relations for categories and views); [] returns an outline

• entries[].type is attack_pattern, category, or view, and only patterns carry abstraction. A category or view's relations list its direct members in source order, or for a filter-defined view its source XPath as filterDefinition; its mappings and notes carry the recorded taxonomy mappings, references, and MITRE notes

• Start at view CAPEC-1000 (Mechanisms of Attack) or CAPEC-3000 (Domains of Attack) to browse patterns by category

• Same entries[].kind and notFound contract as cwe_get_weakness; CAPEC patterns, categories, and views all resolve here, so wrong_type marks only a prefixed CWE ID. Mappings keep the recorded CWE and ATT&CK IDs

cwe_search_weaknesses tool

• Optional all-token filter over the name, MITRE's recorded alternate terms (XSS, ReDoS, UAF), the description, and the extended description, plus exact language, technology, abstraction, consequenceImpact, status, mappingUsage, taxonomyName, taxonomyEntryId, observedCveId, and viewId; no filters browses the catalog

• A filter token matches at a word start, after _, or at a camelCase hump (dos finds ReDoS). Tokens of at least four Unicode letters also match after exactly one un or mis prefix at a word start (trusted finds untrusted; config finds misconfiguration). Prefixes are never stripped recursively; spelling coincidences remain possible. Other compounds are searched as written: serialization does not find CWE-502, deserialization does. A token ending in a digit never matches before another digit (79 skips RFC 793)

• Each result carries MITRE's mappingUsage, such as Allowed, Allowed-with-Review, or Prohibited, which says whether the entry may be used to map a vulnerability

• taxonomyName (such as CERT C Secure Coding or ISA/IEC 62443) and taxonomyEntryId (such as ARR30-C or A1, any case) find the weaknesses that record a taxonomy mapping, both holding on one mapping when given together; observedCveId (such as CVE-2021-44228) finds the weaknesses MITRE cites that CVE on as an observed example, not NVD's CVE-to-CWE assignments. Each result then lists every mapping (matchedTaxonomyMappings) or observed example (matchedObservedExamples) that matched

• viewId includes same-view membership closure, or for a filter-defined view the weaknesses its source XPath filter selects (CWE-658 C, CWE-702 Implementation, …). Categories and groupings are excluded; every status is included by default

• Ranked by all tokens in the name or in one alternate term, then in the description, then any other match, then numeric ID; previews hold the first sentence, and cwe_get_weakness returns the terms in summary. limit default 20, max 100; continue with pagination.nextCursor

cwe_search_attack_patterns tool

• Optional all-token filter over the name, recorded alternate terms (MITM), and the description, using the same word-start and bounded-prefix rule as cwe_search_weaknesses, plus cweId, attackTechniqueId, abstraction, severity, and likelihood; CWE mappings match exactly, and a parent ATT&CK technique includes its recorded subtechniques

• includeInactive (default false) adds Deprecated and Obsolete patterns. Ranked by all tokens in the name or in one alternate term, then in the description, then any other match, then numeric ID; cwe_get_attack_pattern returns the terms in summary. limit default 20, max 100; continue with pagination.nextCursor

cwe_get_view tool

• One view id (from cwe_list_reference topic views); limit default 20, max 100, continued by pagination.nextCursor

• Objective, audience, status, filter definition, recorded notes, and direct members in published order. Navigation views list only top-level members; cwe_get_hierarchy reaches descendants

• A filter-defined view lists no members: filterDefinition is its source XPath, and cwe_search_weaknesses with its viewId returns the weaknesses that filter selects. A filter with a part the server does not evaluate is named as such in guidance, and that search returns none with a notice

• found: false carries not_found or wrong_type

cwe_get_hierarchy tool

• One CWE id and a direction: parents, children, ancestors, or descendants; viewId defaults to 1000, or all for every view. primaryOnly applies to ancestors only. Parents and children include category, grouping, and view relations and match the first level of the same view's tree

• View-scoped nodes, parent-to-child edges, found, retrievedAt, and hierarchyFingerprint. Empty relations succeed; upstream failures are errors, and a temporary MITRE refusal (HTTP 403) is retryable after a few minutes. Node names come from the loaded catalog and can lag the live API

• The total deadline includes queue wait; expiry is retryable hierarchy_timeout, while caller cancellation is nonretryable. Source-recorded relations remain available through cwe_get_weakness, but do not replace live children or descendants

• limit default 20, max 100 edges; each page lists only the nodes its edges reference, plus any root without edges on the first page. Each pagination.nextCursor call re-fetches the traversal; invalid_cursor means the topology or catalog changed, so restart without a cursor

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

MITRE-specific:

• CWE schema 7 and 8 catalogs are supported. CWE groupings retain their own record type alongside older categories; both organize weaknesses rather than identify root causes. Sparse schema-8 metadata stays in existing sections, including custom phase domains, platform notes, note versions, observed-example rationales, and alternative reference links

• No API key. Both XML catalogs load into memory before the server accepts requests. Each successful download is saved to a cache directory (CWE_CACHE_DIR), and a later start loads the saved copy without waiting on MITRE, so a restart works while MITRE is unreachable. The first start, or any start without a usable saved copy, needs MITRE's CWE and CAPEC download sources

• CWE checks for a new version every 24 hours and CAPEC refreshes weekly; a catalog loaded from the cache is checked right after start and, while those checks fail, again after 5, 10, 20, 40, then every 60 minutes. A failed refresh keeps the loaded data and shows in cwe_list_reference topic versions, which also reports whether each catalog came from the cache or a download

• MITRE can refuse a burst of requests from one address with HTTP 403 for a few minutes. Catalog downloads, version checks, and hierarchy calls report it as a temporary, retryable failure and do not retry it immediately

• Hierarchy calls go to the live CWE REST API and keep multiple parents and view identity; search and detail run against the local catalogs

• MITRE statuses and absent sections stay explicit. CAPEC mappings and execution flows are incomplete: a missing mapping does not establish that no relationship exists, or that an ATT&CK ID is current

Agent-friendly output:

• The published mitre-cwe-guide supplies classification gates and multi-step research workflows through plugin discovery and the MCP Skills extension

• Provenance on every response: catalog versions and dates, load time, source links, and attribution

• A 24,000-byte budget across both output surfaces, enrichment included. Whole IDs, items, or edges defer to the next call, and a default detail selection omits whole sections that do not fit, naming them for a re-call; explicit selections stay complete, and a single record or item over budget returns whole with a notice

• Continuation through deferredIds, pagination.nextCursor, or examplePage.nextOffset; stale cursors fail with invalid_cursor, unknown vocabulary with unknown_filter, whose message names the field and value and lists the supported values or the cwe_list_reference topic that does

• CWE and CAPEC ID inputs accept an optional prefix in any case, followed by -, _, a space, or nothing, and drop leading zeros. ATT&CK technique IDs take Tdddd or Tdddd.ddd, with or without the T in either case (T1059, t1059.001, 1059). A malformed single ID is an invalid argument whose message names the accepted forms

• content[] and structuredContent carry the same values, ordered example code included. Catalog prose and code are reference data, never instructions

Getting started

Connect to the hosted instance, no install:

claude mcp add --transport http mitre-cwe-mcp-server https://mitre-cwe.caseyjhand.com/mcp

Or run it locally:

claude mcp add mitre-cwe-mcp-server -- npx -y @cyanheads/mitre-cwe-mcp-server

For other clients, add this to the MCP config file:

{
  "mcpServers": {
    "mitre-cwe-mcp-server": {
      "command": "npx",
      "args": ["-y", "@cyanheads/mitre-cwe-mcp-server"]
    }
  }
}

No API key is required. For Docker or a source build, see Running the server.

Configuration

No environment variables are required. Catalog sources, refresh intervals, and response limits are fixed; framework storage settings do not persist these catalogs.

See .env.example for the full list of optional overrides.

Running the server

Local development

Requires Bun v1.4.0 or higher (or Node.js v24+), outbound HTTPS access to cwe.mitre.org and capec.mitre.org at least once to fill the cache, and cwe-api.mitre.org for version checks and live hierarchy calls.

• Clone and install:

  git clone https://github.com/cyanheads/mitre-cwe-mcp-server.git
  cd mitre-cwe-mcp-server && bun install
  cp .env.example .env # optional transport, authentication, and logging settings

• Build and run the production version:

  # One-time build
  bun run rebuild
  
  # Run the built server
  bun run start:http
  # or
  bun run start:stdio

• Run checks and tests:

  bun run devcheck  # Lints, formats, type-checks, and more
  bun run test      # Runs the test suite

Docker

docker run --rm -p 3010:3010 -v mitre-cwe-cache:/usr/src/app/.cache ghcr.io/cyanheads/mitre-cwe-mcp-server:latest

The image defaults to HTTP transport, stateless sessions, and logs to /var/log/mitre-cwe-mcp-server. The named volume keeps catalog downloads across containers. For stdio, use "command": "docker" with "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "-v", "mitre-cwe-cache:/usr/src/app/.cache", "ghcr.io/cyanheads/mitre-cwe-mcp-server:latest"]. To build locally, run docker build -t mitre-cwe-mcp-server .; add --build-arg OTEL_ENABLED=false to omit OpenTelemetry peers.

Project structure

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic

• Use ctx.log for logging, ctx.state for storage

• Register new tools and resources in the createApp() arrays in src/index.ts

• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.

CWE data: Copyright © 2006–2026, The MITRE Corporation. CAPEC data: Copyright © 2007–2026, The MITRE Corporation. CWE and CAPEC are MITRE trademarks; their data licenses and disclaimers are reproduced in THIRD_PARTY_NOTICES.md.