Skip to main content

Installation

Requirements​

  • .NET 8, .NET 9 or .NET 10 target framework
  • Microsoft.Extensions.AI ≥ 9.x (the IChatClient abstraction)

That's it. No native dependencies, no out-of-process services, no databases unless you opt into the SQLite audit store.

Core install​

The minimum viable install — gets you the pipeline, all 60 detectors (56 on by default, 4 opt-in compliance detectors), the intervention engine, and the in-memory ring-buffer audit store:

dotnet add package AI.Sentinel

Optional packages​

Pick what you need; everything is independently versioned and additive.

Dashboard​

dotnet add package AI.Sentinel.AspNetCore

Embedded real-time dashboard. No JS framework — HTMX + Server-Sent Events. Mount with app.MapAISentinel("/route-prefix").

Persistent audit​

dotnet add package AI.Sentinel.Sqlite

SqliteAuditStore — single-file SQLite database. Hash-chain integrity preserved across process restarts. Time-based retention sweep.

Audit forwarders to external SIEMs​

dotnet add package AI.Sentinel.AzureSentinel    # Azure Monitor Logs Ingestion API
dotnet add package AI.Sentinel.OpenTelemetry    # OTel collector (vendor-neutral)

NDJSON file forwarder is in the core package — no separate install.

Custom detector SDK​

dotnet add package AI.Sentinel.Detectors.Sdk

SentinelContextBuilder, FakeEmbeddingGenerator, DetectorTestBuilder. You only need this package if you're writing custom detectors — IDetector itself lives in AI.Sentinel.

CLI tools (dotnet tool install)​

dotnet tool install -g AI.Sentinel.Cli              # offline replay for forensics + CI
dotnet tool install -g AI.Sentinel.Mcp.Cli          # stdio MCP proxy
dotnet tool install -g AI.Sentinel.ClaudeCode.Cli   # Claude Code hook adapter
dotnet tool install -g AI.Sentinel.Copilot.Cli      # GitHub Copilot hook adapter

TFM matrix​

Packagenet8.0net9.0net10.0
AI.Sentinel✓✓✓
AI.Sentinel.AspNetCore✓✓✓
AI.Sentinel.Detectors.Sdk✓✓✓
AI.Sentinel.Sqlite✓✓✓
AI.Sentinel.AzureSentinel✓✓✓
AI.Sentinel.OpenTelemetry✓✓✓
AI.Sentinel.Mcp (library behind the sentinel-mcp CLI; no public embedding API)✓✓✓
AI.Sentinel.ClaudeCode✓✓✓
AI.Sentinel.Copilot✓✓✓
AI.Sentinel.Approvals.Sqlite✓✓✓
AI.Sentinel.Approvals.EntraPim✓✓✓
AI.Sentinel.Cli✓—✓
AI.Sentinel.ClaudeCode.Cli, AI.Sentinel.Copilot.Cli, AI.Sentinel.Mcp.Cli——✓

AI.Sentinel.Cli ships as a dotnet tool for net8.0 and net10.0; dotnet tool install picks the newest one your machine can run. The hook and MCP CLIs are .NET 10 RID-specific tools: installing them needs the .NET 10 SDK, and on linux-x64, linux-arm64, linux-musl-x64, linux-musl-arm64 (Alpine), osx-arm64 and win-x64 you get a Native AOT build. Library packages multi-target, so a net8, net9 or net10 host gets a matching build.

What you get out of the box​

Once AI.Sentinel is installed and AddAISentinel() is called, these services are wired into DI as singletons:

  • IDetectionPipeline — runs the 56 default-on detectors in parallel per call, plus any opt-in compliance detectors you enable
  • the intervention engine (internal) — applies the configured OnCritical/OnHigh/OnMedium/OnLow action
  • IAuditStore — RingBufferAuditStore (in-memory, bounded by opts.AuditCapacity, default 10,000)
  • IAlertSink — NullAlertSink unless opts.AlertWebhook is set
  • IToolCallGuard — DefaultToolCallGuard (allow-by-default unless you wire policies)

60 official detectors auto-register via the [Singleton(As = typeof(IDetector))] source generator — no manual AddTransient<IDetector, ...>() calls.

Next: Quick start — wire it into your chat client in 5 minutes​