SecOpsAI v1.0.0 is the current stable release.
Evidence-first security operations for agent telemetry, software supply chains, and analyst-reviewed research.
SecOpsAI brings host and AI-agent telemetry, package-registry surveillance, findings triage, guarded automation, and security publishing into one local-first workflow. Deterministic evidence stays authoritative; model review is bounded, optional, and visible to the operator.
Website · Documentation · Install · Mission Control · Security policy
The product tour uses representative sample data. It contains no live credentials, private telemetry, customer records, or local filesystem paths.
Security evidence is usually split across host logs, agent activity, package registries, issue queues, research notes, and deployment tools. SecOpsAI normalizes those signals into a shared evidence model, helps operators decide what deserves attention, and keeps consequential actions behind explicit approval boundaries.
| Outcome | What SecOpsAI provides |
|---|---|
| See the operating picture | Unified OpenClaw, Hermes, macOS, Linux, Windows, Edge, CI, and registry evidence |
| Find risky software early | Multi-ecosystem registry monitoring, no-execution artifact analysis, advisories, and AI Dependency Guard |
| Investigate with context | Durable Research Cases, IOC extraction, correlations, evidence matrices, and bounded model review |
| Respond safely | Explainable triage, reversible low-risk automation, protected actions, and complete audit history |
| Route engineering work safely | Reviewed specialist profiles, persisted OpenCodex model routing, bounded contracts, isolated worktrees, and independent review |
| Publish defensible research | Evidence-linked drafts, media review, editorial approval, archive-safe staging, and separate deployment |
| Area | Capability | Operator guide |
|---|---|---|
| Detection | Cross-platform collection, normalization, correlation, adaptive scoring, and findings | Platform overview |
| Supply chain | npm, PyPI, Packagist, Go, Maven, NuGet, RubyGems, Open VSX, crates, Hugging Face, and container evidence | Supply-chain security |
| AI-built software | Hallucinated, missing, newly registered, lookalike, and source-mismatch dependency review | AI Dependency Guard |
| Artifact analysis | Metadata indexing, quarantine, checksums, static/YARA rules, minimized triage, and analyst escalation | Artifact Fleet |
| Research | Durable cases, competing hypotheses, evidence plans, claim ledgers, blinded review, visual QA, disclosure, and sandbox gates | Research reliability |
| Triage | Evidence bundles, dispositions, queued actions, mitigation, and auditable closure | Findings triage |
| Specialist work | Deterministic expertise routing, explicit OpenCodex model/fallback snapshots, guarded worktrees, and independent review | Specialist Orchestrator |
| Publishing | Source-backed drafts, image review, feeds, archive-safe staging, and protected deployment | Blog publishing |
| Enterprise | Read-only cloud adapters, vulnerability context, Kubernetes posture, authorized DAST plans, and governance records | Enterprise architecture |
Install the complete local-first platform:
curl -fsSL https://secopsai.dev/install.sh | bash
cd ~/secopsai
source .venv/bin/activate
secopsai status
secopsai refresh --platform macos,openclaw,hermes
secopsai triage summaryThe installer checks out the current stable release and runs the documented setup profile. Review the script before running it in a sensitive environment, or use the manual path:
git clone https://github.com/Techris93/secopsai.git
cd secopsai
bash setup.sh --non-interactive --profile default
source .venv/bin/activateThe published npm package is the OpenClaw plugin distribution, not the complete Python platform. Install it through OpenClaw:
openclaw plugins install secopsaiFor repository scanning in CI, use the SecOpsAI Supply Chain Guard:
- uses: Techris93/secopsai-action@v1.0.0
with:
mode: advisory-check
ecosystem: npm
package: node-ipc
version: 12.0.1
fail-on-severity: high| Distribution | Use it for | Status |
|---|---|---|
| Installer / source checkout | Complete SecOpsAI Core and CLI | Recommended |
| Python editable install | Local development and test contributions | Documented in Contributing |
npm secopsai@1.0.0 |
Published OpenClaw plugin | Available |
GitHub Packages @techris93/secopsai |
Authenticated scoped package workflow | Published; access may require read:packages |
Marketplace Action v1.0.0 |
Advisory, package, discovery, and triage checks in GitHub Actions | Available |
See Getting Started for platform-specific setup, GitHub Distribution for the exact npm and GitHub Packages boundaries, and Deployment for long-running services.
Mission Control is maintained separately so the static Cloudflare-compatible console and local helper can evolve without coupling UI delivery to Core:
git clone https://github.com/Techris93/secopsai-dashboard.git
cd secopsai-dashboard/secopsai-dashboard
cp .env.example .env
./start-local-dashboard-stack.shOpen http://127.0.0.1:45680. Local helper actions require the configured
action token; hosted mode fails safely when a helper-backed capability is not
configured.
In Work, Specialist Orchestrator can preview a deterministic route for each item, save a recommendation, queue read-only analysis, or prepare an explicitly approved isolated worktree. The specialist defines the expertise; the persisted operator-selected OpenCodex model performs the work. OpenClaw and Hermes remain separate optional telemetry and compatibility runtimes.
Mission Control is the operator-facing companion to the SecOpsAI CLI. It keeps the dark green navigation and restrained green action language of the product concept while using a bright, high-contrast workspace for dense operational data.
The full-width overview above shows priorities, investigation queues, research production, and service health without exposing internal implementation views as competing products.
Work the latest evidence-backed detections first. Each row exposes severity, confidence, environment impact, evidence state, ownership, and the next safe action without turning scanner output into an automatic verdict.
Persist the model you chose, see its current health, and decide explicitly whether an ordered fallback policy may be used. Primary-only mode leaves work queued rather than silently consuming another provider.
Index registry metadata, run deterministic static and YARA checks, minimize the
context sent to optional model triage, and escalate only suspicious or
inconclusive artifacts. Source-First Artifact Research is the single
adapter-driven workflow for npm, PyPI, crates.io, Packagist, Go, Maven, NuGet,
RubyGems, Open VSX, GitHub, Hugging Face, containers, and approved local
artifacts. Package code, lifecycle scripts, extensions, and binaries are never
executed. Use secopsai research investigate --ecosystem <name> --package <identifier> --json; the existing research rust-package command remains a
compatibility alias.
Every artifact result is context-calibrated: structured manifests determine lifecycle hooks, language-aware rules inspect executable call sites, source and documentation URLs stay separate from attacker IOC candidates, and repeated observations are deduplicated with an audit count. Mission Control shows a decision card with priority, detection confidence, assessment, potential impact, local exposure, evidence quality, contradictions, and next action. See Artifact Signal Calibration.
For material investigations, the Research Case reliability workspace adds competing hypotheses, versioned evidence plans, scaffold/transition/full safe gates, tamper-evident run bundles, claim-level support checks, specialist and blinded independent review, adjudication, completeness/originality/visual audits, and resource accounting. Unsupported claims are removed or qualified with an auditable revision diff before a review-only draft can be created. See Execution-Grounded Research Architecture and Research Reliability Operations.
Use Run Safe Automation to advance every deterministic gate that the current evidence supports. The same guarded coordinator runs during the enabled daily workflow, queues at most one selected-model read-only review, and resumes idempotently when evidence or review state changes. It stops for unsupported claims, reviewer disagreement, real visual evidence, publication approval, sandbox submission, disclosure, publishing, deployment, destructive response, and any external communication.
Turn a package lead into a durable case with quarantined artifact evidence, checksums, rule hits, comparison results, IOCs, readiness gates, and a review-only publication handoff.
Review claims, references, and media; approve editorial content; stage approved posts; and deploy the complete archive as a separate protected action. Older published posts are preserved during rebuilds.
See the difference between an available adapter, a configured connector, and a source producing fresh evidence. Cloud ingestion is read-only by default, Kubernetes assessment is non-mutating, and active DAST requires recorded authorization.
flowchart LR
A[Agent and host telemetry] --> C[Normalize and correlate]
B[Registries and advisories] --> D[Static, policy, and YARA checks]
C --> E[Findings and evidence]
D --> E
E --> F[Bounded model review]
F --> G[Analyst decision]
G --> H[Guarded response]
G --> I[Review-only publication]
Safety invariant: untrusted artifacts are never executed by this workflow; models receive bounded evidence, and response, disclosure, and publication stay behind explicit operator approval.
- Collect: adapters ingest supported host, agent, Edge, cloud, CI, and registry signals.
- Normalize: SecOpsAI maps evidence into shared event, finding, asset, and research records.
- Detect: deterministic rules, advisories, artifact checks, correlations, and policy gates produce explainable evidence.
- Investigate: Research Cases preserve provenance while optional models receive only bounded, minimized context.
- Decide: an analyst verifies evidence, disposition, mitigation, disclosure, and publication readiness.
- Act: only allowlisted, approval-appropriate responses are applied and recorded.
Coverage labels are intentionally strict: Complete is production-usable in the documented local workflow, Partial requires scoped configuration or does not cover the full platform surface, Experimental is pilot-stage, and Planned describes target architecture that is not presented as available.
| Source or surface | Support level | Notes |
|---|---|---|
| OpenClaw | Complete | Audit telemetry, plugin workflow, detections, and response guidance |
| Hermes Agent | Complete | Read-only log and tool-call collection with persistent monitoring |
| macOS | Complete | Unified log, process, file, persistence, and network evidence |
| Linux | Partial (beta) | Auth, process, file, persistence, and network adapters |
| Windows | Partial (beta) | Event, process, PowerShell, persistence, and network adapters |
| SecOpsAI Edge | Experimental (pilot) | Normalized asset graph and findings import; raw scan logs remain at the sensor |
| Package registries | Complete | Registry metadata and safe artifact inspection across the documented ecosystems |
| AWS, GCP, Kubernetes | Partial (read-only) | Normalized connectors and non-mutating posture checks; no infrastructure changes |
| PostgreSQL data plane | Partial (optional) | Pooled adapter is available; SQLite remains the authoritative local default |
| Managed hosted control plane | Planned | Target architecture only; not presented as an available hosted service |
- Package artifacts are inspected without installing them, importing modules, running lifecycle scripts, activating extensions, or executing binaries.
- Browser actions call fixed helper routes; the browser cannot provide arbitrary shell commands.
- Model analysis receives minimized evidence and cannot independently publish, disclose, submit to a sandbox, or enable unverified rules.
- Active DAST, cloud mutations, Kubernetes changes, ticket creation, disclosure, and publication remain approval-gated.
- Source references are kept separate from attacker-controlled IOCs.
- Credentials stay server-side or in the operator's local runtime and must never be committed to the repository.
Read the Security Policy, Threat Model, Security and Data Handling, and Operator Runbook before enabling protected actions.
| Start with | When you need |
|---|---|
| Documentation home | The complete operator documentation |
| Getting Started | Installation and first-run checks |
| Intelligence integrations | Local model bridge, model routing, and the read-only ChatGPT app |
| Specialist Orchestrator | Work routing, reviewed profiles, automation tiers, approvals, recovery, and profile updates |
| Research discovery | Watchlists, candidate intake, orchestration, and promotion |
| Triage orchestrator | Evidence collection, action queues, and closure |
| Rules registry | Detection rule lifecycle and validation |
| API reference | Protected Core and integration contracts |
| Security and data handling | Local-first storage, credentials, models, artifacts, and approval boundaries |
| Repository layout | Canonical folders, compatibility entry points, and duplication rules |
| Background monitoring | Long-running services and scheduled operation |
| GitHub Action | Versioned repository security checks |
| Marketplace | Install SecOpsAI Supply Chain Guard |
| Issue tracker | Bugs, feature requests, and operator feedback |
| Security reporting | Private vulnerability reporting and response expectations |
| Contributing | Development setup, checks, and review expectations |
| MIT License | Open-source terms |