MCP

AiSOC: Getting Started with an Open-Source, Self-Hosted AI Security Operations Center

AiSOC is best suited to security teams or individuals who want to self-host an AI security operations center and investigate threats with auditable AI agents. The key prerequisite before starting is a Docker VM with 8 GB of memory and 20 GB of disk (plus python3 3.9+ and bash), and the first run downloads roughly 2 GB of models into a volume that only make clean removes.

Official source · GitHub - beenuar/AiSOC

Install

git clone https://github.com/beenuar/AiSOC && cd AiSOC
make up

Who It's For

  • Security teams or individual researchers who want to run an AI security operations center (SOC) locally or on their own servers: AiSOC is an open-source, self-hostable project that ingests security telemetry, detects and correlates threats, investigates with AI agents whose reasoning is auditable, and proposes response actions that require human approval.
  • Users who want it to run without depending on cloud API keys: the official documentation states it runs without any API key.
  • Users of clients such as Claude, Cursor, or Continue who want to connect to it via MCP.

Not a good fit when:

  • You don't have a Docker environment meeting the resource requirements (see prerequisites below) — installation cannot be completed.
  • The evidence does not state whether it is suitable as a production-grade SOC replacement, or which specific telemetry source formats it supports. These are unknowns; verify them in the repository before committing.

First Try in 3 Minutes

The recommended main path is the official Quick start: clone the repository and start it with make.

Prerequisites:

  • Docker Compose v2, with 8 GB of memory and 20 GB of free disk space in the Docker VM.
  • python3 (3.9 or later) and bash.
  • If you're unsure whether your environment qualifies, run the official make doctor first — it checks all of the above.

Install and start (official commands):

git clone https://github.com/beenuar/AiSOC && cd AiSOC
make up

The first run downloads roughly 2 GB of models into a volume that only make clean will remove — a side effect you should know about before starting.

Success signal: once the services are running locally, you can interact with them through the UI or the MCP interface. The evidence does not provide specific UI steps or a default access URL after the first run; follow the repository README for that step.

Low-risk first try: the evidence does not list the specific MCP tool names or their read/write properties, so no particular "read-only tool" can be recommended as a first call. Until tool side effects are confirmed, it's advisable to start with display-oriented features such as the UI and the decision ledger, and avoid executing response-type actions directly.

Use Cases

Case 1: Check your environment before starting. Scenario: you're not sure whether your Docker VM has enough resources. Approach: after cloning the repository, run make doctor first — it checks Docker Compose v2, 8 GB of memory, 20 GB of disk, python3 3.9+, and bash one by one. Expected outcome: run make up only after the checks pass, avoiding a first-run failure midway due to insufficient resources.

Case 2: Connect AiSOC via MCP in Claude, Cursor, or Continue. Scenario: you want to query AiSOC's detection results or investigation conclusions directly from these clients. Approach: the project officially ships with an MCP server for Claude, Cursor, and Continue; the exact configuration fields and connection steps in each client are not given verbatim in the evidence, so defer to the corresponding MCP configuration instructions in the repository. Confirmable boundary: it can be confirmed that an MCP server and the target clients are provided, but the list of available tools after the first connection, their parameters, and output formats are not covered by the evidence and should not be assumed.

Limits and Boundaries

  • Read/write nature: it ingests security telemetry and detects threats; AI agents investigate and "propose" responses, but responses require human approval — i.e., write operations have a confirmation point. The evidence does not state which MCP tools are read-only and which trigger actual actions.
  • Local data and disk side effects: the first run downloads roughly 2 GB of models into a volume that only make clean removes; this consumes real disk space, and the cleanup itself deletes the volume's contents.
  • Credential requirements: the official documentation says it runs without an API key; no other credential requirements are mentioned.
  • Explicit requirements: Docker Compose v2 (8 GB memory / 20 GB disk), python3 3.9+, and bash.
  • Not yet confirmed: the specific MCP tool list, each tool's read-only properties, the default access port, supported telemetry source types, and the upgrade/uninstall procedures are not given on the official page in this evidence; consult the latest repository README before use.