CALDERA 的 AI 辅助运营工作空间。聊天驱动的工作流程创建对手内容，将 CTI 导入 STIX，并规划和运行对手仿真，LLM 基于 MCP 工具和实时 CALDERA 状态，而非操作员复制粘贴。

工作流程

在“计划与执行”中选择STIX可启用该次运行的CTI检索;若未选中，工作流程则在不选CTI的情况下运行。每次运行都会保留其转录记录、工具调用和工件以供审查。

管道生产的

威胁报告告诉你该跑什么。你的环境告诉你在哪里。

提取只得到 ATT&CK 技术、命名的威胁行为者和文件哈希可观测量，除此之外什么都没有。

Caldera MCP Plugin

An AI-assisted operations workspace for CALDERA. Chat-driven workflows create adversary content, ingest CTI into STIX, and plan and run adversary emulation, with the LLM grounded in MCP tools and live CALDERA state instead of operator copy-paste.

Workflows

Selecting STIX in Plan and Execute enables CTI retrieval for that run; with none selected the workflow runs without it. Every run keeps its transcript, tool calls and artifacts for review.

What the pipeline produces

A threat report tells you what to run. Your environment tells you where.

Extraction yields ATT&CK techniques, the named threat actor and file hash observables, and nothing else. Hosts, accounts, domains, services and software describe the report’s victim rather than your estate, so they never become CALDERA facts. CALDERA discovers those at runtime, which is why the abilities that use them carry has_agent_copy and no_backwards_movement requirements.

The one insertion point into core CALDERA is the adversary profile:

upload report(s) -> clean -> IR -> ATT&CK techniques -> STIX bundle
                                                            |
  "build an adversary from this and run it"                 v
      build_adversary   attack-pattern -> ability.technique_id -> Adversary
                        + unmatched_techniques + platform_excluded
      run_operation     Adversary + your agents + runtime-discovered facts

Installation

Install like any CALDERA plugin, from the CALDERA root:

git clone https://github.com/mitre/mcp.git plugins/mcp
source venv/bin/activate
pip install -r plugins/mcp/requirements.txt
python -m spacy download en_core_web_lg
cp plugins/mcp/.env.example plugins/mcp/.env

The spaCy model needs its own command because pip cannot install it from requirements.txt. Then add mcp to the plugins: list in conf/local.yml, alongside magma, sandcat and stockpile. Do not commit that file.

Optional extras:

brew install poppler                              # macOS, supplies pdftotext
apt install poppler-utils                         # Debian/Ubuntu
pip install -r plugins/mcp/requirements-dev.txt   # to run the test suite

Without poppler, TXT, MD and HTML still ingest and PDF uploads report the missing dependency rather than failing silently.

LLM credentials