SecurityClaw SOC Automation Framework
Analyze OpenSearch or Elasticsearch logs with behavioral baselines and LLM-assisted anomaly triage.
- Source repo
- SecurityClaw/SecurityClaw
- Stars
- ★ 279
- Last updated
- 4mo ago
- License
- MIT
- Primary language
- Python
- FA score
- 51/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- High · needs real infrastructure
- You'll need
- Typical use
- A SOC team needs behavioral baselines built from network logs in OpenSearch or Elasticsearch.
- Not a fit if
- Teams seeking a turnkey hosted SOC service
- Teams unable to maintain a database, model service, and local configuration
- Maintenance
- Commits in the last 6 monthslast commit 2026-07-20
- Source review
- 51/100 · Major gaps
What does this agent do, and when should you use it?
SecurityClaw is a self-hosted, modular SOC framework that uses Python skills to read logs and anomaly findings from OpenSearch or Elasticsearch. APScheduler runs anomaly monitoring and baseline-building jobs; NetworkBaseliner summarizes logs into a vector index, and ThreatAnalyst uses retrieved baseline context to produce verdicts. The LangGraph chat_router supports interactive investigations and checkpoints state to SQLite. Users can operate it through a CLI, a FastAPI service, and a React web interface. Deployment requires configuring a database and LLM provider and running the corresponding services.
After database and LLM configuration, the Runner discovers skills under the skills directory and APScheduler runs those with configured intervals. NetworkBaseliner queries recent logs, aggregates behavior, and stores embeddings in the RAG vector index; anomaly_triage can poll findings, add descriptions, and place high-severity results in working memory; ThreatAnalyst retrieves similar baselines and returns TRUE_THREAT, FALSE_POSITIVE, UNKNOWN, or ERROR verdicts. The LangGraph chat_router runs a DECIDE→EXECUTE→EVALUATE StateGraph for conversational investigations and skill calls. Users can inspect status, dispatch skills, and chat through main.py commands, the FastAPI API, or the React UI; SQLite stores conversation and runtime checkpoints.
- A SOC team needs behavioral baselines built from network logs in OpenSearch or Elasticsearch.
- Security operations staff want to poll anomaly findings and send high-severity results for analyst review.
- Analysts need a chat workflow for querying logs, retrieving baselines, and investigating threats.
- Engineering teams want to extend the system with separate skills for database queries, forensic timelines, or field catalogs.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
pip install -r requirements.txt
- Environment variables
OPENAI_API_KEYANTHROPIC_API_KEYSECURITYCLAW_API_ONLY- Config files
config.yaml.env- Runtime
- Python 3.11+
- Install notes cover
- Windows
How do you install or deploy this agent?
Requirements include Python 3.11+, Git, OpenSearch 2.x or Elasticsearch 8.x (a test mock is also available), and Ollama. The README recommends at least 4 GB RAM and about 2 GB of disk space for models and vector indices. Clone the repository, create a virtual environment, and install dependencies:
git clone https://github.com/SecurityClaw/SecurityClaw.git
cd SecurityClaw
python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txtOn Windows PowerShell, activate with .venv\Scripts\Activate.ps1. Run onboarding to configure the database, Ollama, configuration files, and optional external APIs:
.venv/bin/python main.py onboardThe example configuration uses qwen2.5:7b-instruct-q4_K_M and nomic-embed-text:latest; pull Ollama models separately. OpenAI, Anthropic, and other chat providers are also configurable, but RAG embeddings continue to use the configured Ollama embedding model.
How do you use this agent?
Start Ollama and the database, then configure config.yaml and .env; credentials depend on the selected services, such as OpenSearch username/password, OPENAI_API_KEY, or ANTHROPIC_API_KEY. Start the service, which runs the web UI, API, and scheduler:
.venv/bin/python main.py serviceThe current service instructions list the Web UI at http://localhost:5173 and API at http://localhost:7799. The launchers also expose status, stop, chat, skills, and logs. For CLI operation, run the scheduler, start terminal chat, or dispatch a skill:
.venv/bin/python main.py run
.venv/bin/python main.py chat
.venv/bin/python main.py dispatch threat_analyst
.venv/bin/python main.py statusFor API-only mode, set SECURITYCLAW_API_ONLY=1 before running main.py service. Set the database provider to opensearch or elasticsearch in config.yaml, and specify the log, anomaly, and vector indices.
What are this agent's strengths and limitations?
- Skills are isolated folders containing Python
logic.pyand LLM guidance ininstruction.md, making capabilities extensible. - Configuration can switch between OpenSearch and Elasticsearch through the database abstraction.
- Behavioral baselines are stored in a vector index and retrieved as RAG context for analysis.
- LangGraph conversation and runtime memory can be checkpointed to SQLite.
- Provides scheduled operation, CLI chat, a FastAPI service, and a React interface.
- Deployment requires a database and LLM service, several configuration values, and Python dependencies.
- Ollama is the default model service; remote chat providers still use the configured Ollama embedding model.
- The README marks anomaly_triage, threat_analyst, and forensic_examiner as under validation or in progress.
- The README lists Elasticsearch compatibility testing as an improvement area and says in-progress modules weigh down aggregate test coverage.
How does this agent compare with similar options?
OpenSearch and Elasticsearch are documented as selectable database backends. The README does not compare SecurityClaw with other SOC products.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| SecurityClaw SOC Automation Framework This agent | 51 · Major gaps | Self-hosted serviceFree + model costs | ★ 279 | 4mo ago | Python | OpenAI API · Claude API |
| Home Generative Agent | 69 · Some gaps | — | ★ 383 | today | Python | OpenAI API · Claude API |
| AiSOC — Open-Source Self-Hosted AI Security Operations Center | 84 · Good | CLIFree + model costs | ★ 7.1k | 1d ago | Python | — |
| Local OpenAI LLM (Home Assistant Custom Integration) | 68 · Some gaps | Self-hosted serviceFree + model costs | ★ 317 | 3d ago | Python | OpenAI API |
How does FollowAgents rate this agent?
Why each dimension lost points
README, SECURITY.md, and server.py show that the service binds to 0.0.0.0 by default and has no authentication or authorization. The documentation clearly warns against untrusted access and recommends firewall isolation, but access control and least privilege remain weak. Startup is operator initiated and external APIs require configured keys, but the packet does not establish confirmation before consequential actions or a recovery path. Data-flow documentation distinguishes local storage from optional external threat-intelligence lookups and names IPs and domains that may be sent. .env isolation, secret masking, and retention notes help, although stored data has no automatic expiry. Dependencies have minimum versions rather than a demonstrated lock or security audit. Capabilities, endpoints, and suggested actions are described, but attribution is limited and no clear rollback mechanism is shown.
The README gives conflicting service ports: quick-start instructions list 5173/7799 while later examples use 3000/500, weakening consistency. requirements.txt specifies open-ended minimum versions and the packet provides no lockfile evidence, so dependency availability is only partially addressed. SECURITY.md says failed connection checks may proceed and describes API errors without stack traces, but concrete error-handling evidence is limited in the supplied implementation excerpt.
The material targets single-operator SOC use and covers OpenSearch/Elasticsearch, Ollama and several LLM providers, CLI and web interfaces, Docker, and multiple operating systems. Skill listings, manifest contracts, and manual versus scheduled status describe capability boundaries; optional threat-intelligence integrations and configuration switches help control triggers. The sample does not include enough implementation detail to establish that all boundaries and trigger conditions are enforced consistently.
The README provides a directory map, skill descriptions, installation steps, configuration, command examples, and notes on in-progress skills, context budgets, and security limitations. Conflicting service-port guidance and older frontend/backend examples reduce naming and information stability. The complete MIT license text earns full marks. The supplied material does not clearly establish a changelog or routine maintenance responsibility/update path. It does provide a vulnerability-reporting route, but that alone does not establish ongoing ownership.
The skill-based SOC workflow, RAG behavioral baselines, anomaly analysis, and verdict/recommendation outputs have a clear operational use; CLI, API, and dashboard options support different workflows. The README also says some skills remain under validation and coverage is held back by unfinished modules. Deployment requires OpenSearch, a model runtime, and notable resources, while the packet does not quantify costs or compare them with benefits, limiting these scores.
The README gives checkable detail on configuration, endpoints, data flows, skills, and limitations; SECURITY.md and server.py provide some corroboration on binding, CORS, and configuration masking. Many performance, security, and test-coverage statements remain documentation assertions, and the conflicting port descriptions further limit corroboration. The materials distinguish current state from planned improvements and known risks sufficiently to support basic fact/inference separation. No execution or independent verification is claimed, so confidence remains low as required.
- The service binds to 0.0.0.0 by default and has no API authentication; follow SECURITY.md and restrict it to trusted networks behind firewall controls.
- Documented service ports conflict; verify the current configuration and startup implementation before deployment.
- Configuring external threat-intelligence services may send IPs and domains from logs to third parties; conversation and runtime SQLite data have no automatic expiry.
- Dependencies are not locked. Security scans, runtime behavior, and test results were not executed or independently verified in this static review.
FAQ
Is Ollama mandatory?
Does it automatically respond to threats?
Can I run it from the command line only?
main.py run starts the scheduler, main.py chat starts terminal chat, and main.py dispatch <skill> runs one skill.Which credentials are needed?
Is SecurityClaw SOC Automation Framework free?
Is SecurityClaw SOC Automation Framework open source?
Is SecurityClaw SOC Automation Framework safe to use?
Is SecurityClaw SOC Automation Framework still maintained?
Who makes SecurityClaw SOC Automation Framework?
Thanks — we use this to improve the page.