Detectors¶
AgentSniff uses nine detection modules that run concurrently. Each produces signals with confidence levels that are fused into a single score per service asset. Presence confidence does not imply verified identity.
| Detector | Method | Requires Root | Confidence |
|---|---|---|---|
| DNS Monitor | Passive DNS monitoring for LLM API domain queries | Yes* | High |
| Port Scanner | Async TCP scanning of agent-related ports | No | Low-Medium |
| AgentPin Prober | .well-known/agent-identity.json discovery |
No | Medium (unsigned) |
| A2A Detector | Agent Card discovery and trusted JWS verification | No | High-Confirmed |
| MCP Detector | Version-negotiated JSON-RPC / SSE probing for MCP servers | No | High-Confirmed |
| Endpoint Prober | HTTP probing for agent framework signatures | No | Medium-High |
| TLS Fingerprint | JA3/JA4 client-library hints with capture provenance | Yes* | Low |
| Traffic Analyzer | Behavioral pattern analysis (burst detection, LLM call patterns) | Yes* | Medium-High |
| SSE Detector | Server-Sent Events pattern detection | Yes* | Medium |
* Falls back to non-root alternatives automatically.
The four detectors marked "Requires Root" can use kernel-level eBPF capture when the binary is built with --features ebpf and run with the necessary capabilities. Without eBPF, the detectors use userspace fallback paths where available.
DNS Monitor¶
Passively captures DNS queries on the network and matches against 110+ known LLM API domains including OpenAI, Anthropic, Google, Mistral, Groq, Together, Cohere, DeepSeek, xAI, and Chinese providers (DashScope, Moonshot, Zhipu, MiniMax, Baidu, ByteDance, Tencent Hunyuan).
Also matches label-safe patterns for Azure Foundry (*.services.ai.azure.com), AWS Bedrock and AgentCore, GCP Vertex (*.aiplatform.googleapis.com), and others. Wildcards match exactly one DNS label.
Additionally tracks 150+ agent infrastructure domains — skill registries (ClawHub, Smithery, Glama), agent observability platforms (Langfuse, LangSmith, Helicone), tool connectivity services (Composio, Moltyverse), AI coding-agent backends (Cursor, Devin, Zed, Cline, Warp), and vendor-hosted remote MCP servers (GitHub, Slack, Figma, Notion, Linear, Stripe, Datadog).
Fallback: When raw sockets are unavailable, resolves the top 20 LLM API domains and cross-references their IPs against active connections in /proc/net/tcp.
Port Scanner¶
Async TCP scanner targeting 24+ ports associated with:
- MCP servers: 3000, 3001, 6274, 8080
- Agent frameworks: 8000 (FastAPI/vLLM), 8001, 5000 (Flask), 8888 (Jupyter)
- LLM inference: 11434 (Ollama), 1234 (LM Studio), 4000 (LiteLLM), 30000 (SGLang)
- Vector databases: 6333 (Qdrant), 50051 (Weaviate gRPC), 19530 (Milvus)
- Agent platforms: 3100 (Dify), 3080 (LibreChat), 8501 (Streamlit)
- IDE agents: 65432 (Continue.dev), 8443 (code-server), 2024 (LangGraph Studio)
The signed catalog can assign multiple candidates to ambiguous ports such as 8080. Port-only matches remain low confidence; a banner, protocol response, or another detector must corroborate them.
AgentPin Prober¶
Probes hosts for AgentPin discovery documents at /.well-known/agent-identity.json. The scanner validates the AgentPin v0.1 document shape, including entity metadata, P-256 public-key declarations, delegation depth, timestamp, and agent declarations. Discovery does not prove domain ownership or private-key possession, so the prober reports an explicit unsigned, medium-confidence discovery rather than a verified identity. The optional v0.3 a2a_endpoint is retained as cross-protocol evidence.
Presented credentials use a separate fail-closed verifier: agentsniff verify-agentpin checks the signature, time and key validity, revocation, declared authority, delegation chain, audience, and persistent TOFU state against an authenticated out-of-band trust bundle. This explicit boundary prevents an untrusted scan target from supplying its own trust anchor or causing issuer-driven network requests. See AgentPin credential verification.
Follows the AgentPin spec's no-redirect security policy. Probes HTTPS and HTTP across common ports (443, 8443, 8080, 3000, 8000, 80).
AgentPin itself is a discovery and identity layer rather than an MCP server. The MCP detector recognizes servers that advertise the AgentPin credential-binding capability and reports mcp_agentpin_binding_advertised at high confidence. This means the server claims support for carrying AgentPin credentials in MCP metadata; it does not mean a credential was presented or an identity was cryptographically verified. Separately, agentsniff serve can host an MCP verification tool that checks a presented credential against operator-configured trust material.
A2A Detector¶
Fetches the permanent /.well-known/agent-card.json discovery resource and validates the required Agent Card fields. The legacy /.well-known/agent.json path is accepted as compatibility evidence and identified as such.
Unsigned cards and cards with unknown signers report A2A presence without claiming verified identity. A valid ES256 JWS against an out-of-band public key configured in a2a_trusted_keys produces trusted_verified identity assurance and a confirmed-presence signal. Keys advertised by the scanned host are never trust anchors.
MCP Detector¶
Probes for Model Context Protocol servers using the signed protocol catalog. It tries 2026-07-28 server/discover first, then compatible initialize probes for 2025-11-25 and 2024-11-05. Requests carry MCP-Protocol-Version and Mcp-Method; legacy session identifiers and the initialized notification are preserved when required. On confirmed servers, it enumerates available tools, resources, and prompts.
Detects both HTTP+SSE and direct JSON-RPC transports.
Endpoint Prober¶
Probes HTTP endpoints for signatures of 100+ agent frameworks. Pre-filters closed ports with a quick TCP connect check before firing HTTP probes.
For each reachable port, checks:
- Framework-specific and well-known paths (e.g.,
/crew/statusfor CrewAI,/runsfor LangGraph) - Framework-identifying response headers (e.g.,
x-langchain-*,semantic-kernel-version) - Response body markers and JSON predicates from the signed schema
- Agent metadata documents (
/.well-known/agents.json,/AGENTS.md,/SKILL.md) - OpenAPI specs with AI-related content (
/openapi.json,/docs)
Detected frameworks include: LangChain, CrewAI, AutoGen, Dify, Flowise, n8n, PydanticAI, LangGraph, AG2, Haystack, Composio, Letta, Mastra, and many more. Request User-Agent and MCP client-name indicators are kept as passive indicators and are never searched in server response bodies.
TLS Fingerprint¶
Computes JA3 and JA4 fingerprints from TLS ClientHello messages to produce low-confidence client-library hints. Each catalog entry records library versions, TLS backend, capture time, and source; legacy entries are explicitly marked for recapture. These generic fingerprints never establish agent identity.
Fallback: Active TLS server probing on agent-associated ports when passive capture isn't available.
Traffic Analyzer¶
Profiles network hosts by behavioral patterns characteristic of AI agents:
- Burst detection — Bursty tool invocation sequences
- ORA loop — Observe-reason-act timing patterns (tool call interspersed with LLM API call)
- API diversity — Connections to many different API targets
- LLM API connections — Active TCP connections to known LLM API IPs
Also analyzes /proc/net/tcp for established connections to known LLM API IP addresses.
SSE Detector¶
Detects Server-Sent Events (SSE) streaming patterns characteristic of LLM responses. Looks for long-lived HTTP connections with text/event-stream content from known LLM API IPs.
Requires: Raw socket access (or eBPF with --features ebpf) for passive capture.
Selecting Detectors¶
Run specific detectors only:
Available detector names: dns_monitor, port_scanner, agentpin_prober, a2a_detector, mcp_detector, endpoint_prober, tls_fingerprint, traffic_analyzer, sse_detector.