DNS monitor
Matches LLM APIs and agent infrastructure using exact domains and provider-aware suffix rules.
Open-source AI infrastructure discovery
AgentSniff is the asset-intelligence and attack-surface platform for AI agents.
AgentSniff combines active probing, passive observation, protocol detection, and signal fusion to surface AI-agent infrastructure—with evidence you can inspect.
curl -fsSL https://agentsniff.org/install.sh | bash
Linux & macOS · requires Cosign to verify release provenance. Windows binaries are available on GitHub Releases.
agentsniff / scanlistening
Target10.0.0.0/24
Hosts256
Detectors9 / 9
Correlated evidence03 signals
mcp_server · 10.0.0.42:8080protocol initialized · 12 tools
confirmed
a2a_agent_card · 10.0.0.18agent card · trusted signer
high
llm_api_dns · 10.0.0.73known provider domain
high
+Signals grouped by source host and fused into inspectable confidence.
9 detectorsCorrelated concurrently
Rust-nativeOne focused implementation
Signed catalogsPinned trust anchor
Apache-2.0Open source
Operator view
From raw signals to a clear inventory.
Launch and stop scans, choose detectors, follow live results, inspect evidence, and return to stored scan history from one local dashboard.
Live scan streamSSE updates results as detector signals arrive.
Evidence, not guessesExpand each host to review its contributing signals.
Portable findingsExport the latest scan as JSON, CSV, or SARIF 2.1.0.
Detection model
One signal is a clue. Nine tell the story.
Detectors run concurrently and correlate evidence by source host with noisy-OR scoring. Weak indicators stay weak; corroboration raises confidence.
Port scanner
Finds MCP, inference, vector, agent-platform, and IDE-agent services without treating an open port as identity.
AgentPin prober
Discovers AgentPin documents as self-asserted evidence, without overstating unsigned metadata as verified identity.
A2A detector
Validates Agent Cards and can verify ES256 signatures against operator-configured trusted keys.
MCP detector
Negotiates current and legacy MCP protocol versions, then enumerates exposed tools, resources, and prompts.
Endpoint prober
Checks framework routes, response headers, OpenAPI documents, metadata standards, and client signatures.
TLS fingerprint
Uses JA3 and JA4 client-library hints as supporting evidence, not standalone agent identity.
Traffic analyzer
Looks for burst behavior, observe-reason-act timing, API diversity, and known LLM connections.
SSE detector
Surfaces long-lived event-stream traffic associated with streamed model responses and agent activity.
Secure by default
Built for discovery without blurring the trust boundary.
AgentSniff separates discovery evidence from verified identity, constrains network-facing behavior, and keeps optional privileged capture explicit.
The dashboard binds to loopback by default; non-loopback access requires a strong API token.
Detection datasets are signed, verified against a pinned key, and activated atomically.
Passive eBPF and raw-network capabilities are optional—active scanning works without root.
release / provenance
Authenticated binaries
The one-line installer fails closed unless Cosign verifies the archive against the expected GitHub Actions release identity.
signal / confidence
Honest confidence
Low-confidence catalog matches remain hints. Protocol responses and locally trusted signatures can produce stronger conclusions.
Quick start
Scan in a few commands.
Install from crates.io, scan a network you are authorized to test, or start the loopback dashboard.
Full installation guide
terminal
# Install the Rust CLI $ cargo install agentsniff --locked # Discover agent infrastructure $ agentsniff scan 192.168.1.0/24 # Scan selected detectors $ agentsniff scan 10.0.0.0/24 \ --detectors mcp_detector,a2a_detector # Start the local dashboard $ agentsniff serve --port 9090