agentsight

AgentSight

中文版

eBPF-based observability tool for AI Agents on Linux, providing zero-intrusion monitoring of LLM API calls, token consumption, process behavior, and SSL/TLS traffic. AgentSight is an observability component of ANOLISA.

macOS support: On macOS, AgentSight compiles with two commands — agentsight trace (trajectory collector, scans local JSONL session files → ATIF → SQLite, no eBPF) and agentsight serve (Dashboard UI + trajectory viewer). The same source tree produces a full-featured eBPF binary on Linux and a trajectory-only binary on macOS via OS-conditional compilation.

Features

• Zero-Intrusion Monitoring — eBPF kernel probes capture events without modifying agent code or configurations.
• SSL/TLS Traffic Decryption — uprobe-based interception of OpenSSL/GnuTLS library calls to capture plaintext HTTP traffic.
• LLM Token Accounting — Precise token counting with Hugging Face tokenizer support (Qwen series and more).
• AI Agent Auto-Discovery — Scans /proc and monitors execve events to dynamically detect running AI agent processes.
• Streaming Response Support — Parses Server-Sent Events (SSE) for tracking streamed LLM responses.
• Audit Logging — Complete audit trail of LLM calls and process operations with structured records.
• Cloud Integration — Native export to Alibaba Cloud SLS (Simple Log Service) for centralized log analysis.
• GenAI Semantic Events — Builds structured semantic events for LLM calls, tool usage, and agent interactions.

Architecture

AgentSight operates a unified data pipeline:

┌──────────┐    ┌────────┐    ┌────────────┐    ┌──────────┐    ┌───────┐    ┌─────────┐
│  Probes  │───▶│ Parser │───▶│ Aggregator │───▶│ Analyzer │───▶│ GenAI │───▶│ Storage │
└──────────┘    └────────┘    └────────────┘    └──────────┘    └───────┘    └─────────┘
  eBPF events    HTTP/SSE      Req-Resp          Token/Audit     Semantic     SQLite /
  (kernel)       extraction    correlation       extraction      events       SLS export

Stage	Description
Probes	eBPF programs (sslsniff, proctrace, procmon) capture kernel events via ring buffer
Parser	Extracts structured HTTP messages, SSE events, and process exec data
Aggregator	Correlates request-response pairs; tracks process lifecycle via LRU cache
Analyzer	Produces audit records, token usage stats, and LLM API messages
GenAI	Transforms results into semantic events (LLM calls, tool use, agent interactions)
Storage	Persists to local SQLite database and optionally uploads to Alibaba Cloud SLS

eBPF Probes

Probe	Source	Description
sslsniff	src/bpf/sslsniff.bpf.c	uprobe on SSL_read/SSL_write to capture plaintext from encrypted connections
proctrace	src/bpf/proctrace.bpf.c	Traces execve syscalls, captures command-line args, builds process tree
procmon	src/bpf/procmon.bpf.c	Lightweight process monitor for creation/exit events (agent discovery)

Project Structure

agentsight/
├── src/
│   ├── bpf/            # eBPF C programs (sslsniff, proctrace, procmon)
│   ├── probes/         # eBPF probe management and event polling
│   ├── parser/         # HTTP, SSE, and process event parsers
│   ├── aggregator/     # Request-response correlation and process aggregation
│   ├── analyzer/       # Token extraction, audit records, message parsing
│   ├── genai/          # GenAI semantic event builder and SLS uploader
│   ├── storage/        # SQLite-backed stores (audit, token, HTTP, GenAI)
│   ├── discovery/      # AI agent process scanner (/proc + eBPF)
│   ├── tokenizer/      # HuggingFace tokenizer integration for token counting
│   ├── local/          # macOS-only: trajectory viewer server + collector dispatch
│   ├── bin/            # CLI entry points (agentsight, cli subcommands)
│   ├── unified.rs      # Main pipeline orchestrator
│   ├── config.rs       # Unified configuration management
│   └── event.rs        # Unified event type definitions
├── Cargo.toml
├── build.rs            # eBPF skeleton generation for three probes
└── agentsight.spec     # RPM packaging spec

CLI Commands

Commands token, audit, discover, metrics, interruption, skill-metrics, and summary require Linux eBPF and are not available on macOS. trace and serve work cross-platform: on Linux trace runs the full eBPF pipeline, on macOS it runs the trajectory collector only (no eBPF).

agentsight trace

Start tracing of AI agent activity.

Linux: Full eBPF-based tracing (probes → parser → aggregator → storage). Also runs trajectory collector if features.trajectory_collection.enabled is set.

macOS: Trajectory collection only — scans local JSONL session files (Claude Code, Qoder, Codex, Cursor), converts to ATIF v1.7, and stores in trajectories.db. No eBPF.

# Foreground mode
sudo agentsight trace

# Daemon mode with SLS export
sudo agentsight trace --daemon \
  --sls-endpoint <endpoint> \
  --sls-project <project> \
  --sls-logstore <logstore>

agentsight token

Query token consumption data.

When the Linux systemd service owns the data, run these queries with sudo:

# Today's token usage
sudo agentsight token

# This week, compared to last week
sudo agentsight token --period week --compare

# Detailed breakdown by role and type
sudo agentsight token --detail

# JSON output
sudo agentsight token --json

agentsight audit

Query audit events (LLM calls, process operations).

# Recent audit events
agentsight audit

# Filter by PID and event type
agentsight audit --pid 12345 --type llm

# Summary statistics
agentsight audit --summary

agentsight serve

Start the HTTP API server and serve the embedded Dashboard UI.

macOS: Reads from trajectories.db (populated by agentsight trace). The --db and --config flags are Linux-only.

# Start with default settings (binds to 127.0.0.1:7396)
agentsight serve

# Bind to all interfaces on a custom port
agentsight serve --host 0.0.0.0 --port 8080

# Point to a specific database file
agentsight serve --db /path/to/genai_events.db

agentsight discover

Discover AI agents running on the system.

# Scan for running agents
agentsight discover

# List all known agent types
agentsight discover --list-known

# Verbose output with executable paths
agentsight discover --verbose

Dashboard

The Dashboard is a React-based web UI for visualizing conversation history, trace details, and token statistics. It is embedded into the agentsight serve binary at compile time. By default, the Dashboard follows the browser language; you can switch languages manually, and the choice is persisted across refreshes.

Build the Dashboard

cd src/agentsight

# Build frontend and embed into frontend-dist/ (required before cargo build)
make build-frontend

# Then build the Rust binary with the embedded UI
make build

# Or do both in one step
make build-all

Scenario 1 — Collect data and view the Dashboard simultaneously

Linux (eBPF + trajectory collector):

Run the tracer and the API server in two separate terminals:

# Stop the packaged tracer before starting a foreground tracer
sudo systemctl stop agentsight.service

# Terminal 1: start eBPF tracing (writes to SQLite)
sudo agentsight trace

# Terminal 2: start the API server (reads from the same SQLite)
sudo agentsight serve

macOS (trajectory collector only):

# Terminal 1: start trajectory collection (scans JSONL → trajectories.db)
agentsight trace

# Terminal 2: start the API server (reads from trajectories.db)
agentsight serve

Open http://127.0.0.1:7396 in your browser. The Dashboard auto-refreshes as new data arrives.

Running on a remote server? Bind to all interfaces and access via the server's public IP:

agentsight serve --host 0.0.0.0 --port 7396

Then open http://<server-public-ip>:7396 in your local browser. Make sure port 7396 is allowed in the server's firewall / security group rules.

Scenario 2 — Browse historical data only

No tracing needed. Just start the server pointing at an existing database:

agentsight serve --db /path/to/genai_events.db

Open http://127.0.0.1:7396 to explore recorded conversations and traces.

Dashboard Development

To iterate on the frontend without rebuilding the Rust binary:

cd src/agentsight/dashboard
npm install
npm run dev          # starts webpack-dev-server on http://localhost:3004

When finished, run make build-frontend && cargo build --release to embed the updated UI.

Quick Start

Prerequisites

System Packages

Before building, install the required system packages:

Anolis OS / CentOS / RHEL:

sudo yum install -y openssl-devel elfutils-libelf-devel perl-IPC-Cmd libbpf-devel clang llvm bpftool

Ubuntu / Debian:

sudo apt install -y pkg-config libssl-dev libelf-dev libbpf-dev clang llvm linux-tools-common

Package	Required for
openssl-devel	OpenSSL vendored build (used via openssl = { features = ["vendored"] })
elfutils-libelf-devel	libbpf-sys crate (provides gelf.h, libelf.h)
perl-IPC-Cmd	OpenSSL source build (Perl IPC::Cmd module)
libbpf-devel	eBPF program compilation and loading
clang / llvm	eBPF C program compilation to BPF bytecode
bpftool	eBPF skeleton generation

You can verify all dependencies with the included check script:

./scripts/check-deps.sh

Version Requirements

Component	Version
Linux kernel	>= 5.8 (BTF support)
Rust	>= 1.80
clang / llvm	>= 11 (for eBPF compilation)
libbpf	>= 0.8

Install with Anolisa

sudo anolisa --install-mode system install agentsight

AgentSight requires Linux system mode. This installs the AgentSight service and the agentsight-enforcer service together.

Install via RPM

sudo yum install agentsight

Installs:

• /usr/local/bin/agentsight — CLI binary
• /usr/local/bin/agentsight-enforcer — ActPlane enforcement engine
• /usr/lib/systemd/system/agentsight.service — AgentSight system unit
• /usr/lib/systemd/system/agentsight-enforcer.service — enforcement system unit

The RPM is a Linux system package. Its units are installed but not enabled by default; when both units run, AgentSight is ordered after the enforcer.

Start the Service

Both package routes leave the units stopped and disabled. Start the main unit when you are ready to begin collection:

sudo systemctl enable --now agentsight.service
sudo systemctl status agentsight.service

The main unit runs eBPF tracing and the Dashboard together and starts the enforcer dependency in the required order. Open http://localhost:7396 after the service becomes active.

The unit runs as root with UMask=0077, so its data under /var/log/sysak/.agentsight is private. Use sudo for CLI queries and Dashboard access commands that read service-owned data. Stop the unit before starting a foreground tracer.

Build from Source

cd src/agentsight

# Verify dependencies (optional but recommended)
./scripts/check-deps.sh

# Build frontend and Rust binary with embedded Dashboard UI
make build-all

The binary is output to target/release/agentsight. On supported Linux systems, make build-all also invokes scripts/build-enforcer.sh to build the attested ActPlane target/release/agentsight-enforcer binary. make build-mac does not build the enforcer.

cargo build --release only compiles Rust. It does not rebuild the embedded Dashboard UI, so use make build-all for user-facing builds.

Build on macOS

macOS builds agentsight trace (trajectory collector) and agentsight serve (Dashboard viewer). It does not require libbpf, clang/llvm, kernel headers, root, or Linux BPF capabilities.

Prerequisites:

Component	Version	Required for
Rust	>= 1.80	Compile Rust code
Node.js	>= 16	Frontend build
npm	>= 8	Frontend dependency management

Build steps:

cd src/agentsight

# Build the frontend and the macOS binary
make build-mac

The binary is output to target/release/agentsight.

Usage on macOS:

# Terminal 1: collect trajectories (scans JSONL → trajectories.db)
agentsight trace

# Terminal 2: start the Dashboard + trajectory viewer
agentsight serve

# Or bind to a custom host/port
agentsight serve --host 0.0.0.0 --port 8080

Open http://127.0.0.1:7396 to view the Agent Dashboard. trace scans local AI agent session files (Claude Code, Qoder, Codex, Cursor) and stores them as ATIF trajectories in trajectories.db. serve reads from the same database.

macOS limitations: eBPF-dependent commands (discover, token, audit, metrics, interruption, skill-metrics, summary) are Linux-only. The --db and --config flags are also Linux-only. On macOS, trace collects trajectories only (no eBPF), and serve reads from trajectories.db.

Start Tracing

# Requires root for eBPF
sudo agentsight trace

Configuration

AgentSight is configured via agentsight.json (default path /etc/agentsight/config.json; falls back to embedded defaults if absent).

Basic Options

Category	Option	Description
Storage	db_path	SQLite database file path
Storage	data_retention_days	Data retention period
Probes	target_uid	Filter events by UID
Probes	poll_timeout_ms	Ring buffer poll timeout
HTTP	connection_cache_capacity	LRU cache size for connection tracking
SLS	sls_endpoint / sls_project / sls_logstore	Alibaba Cloud SLS export settings
Tokenizer	tokenizer_file	Path or URL to HuggingFace tokenizer

Feature Flags (features)

All optional features are enabled by default. Disable them individually via the features block in agentsight.json to reduce memory and I/O overhead:

Feature	JSON Path	Default	Description
Token Stats	features.token_stats	true	Core functionality, not recommended to disable
Local Tokenizer	features.tokenizer.enabled	false	HuggingFace model fallback (50–100 MB per model)
Session Mapping	features.session_mapping.enabled	true	responseId → sessionId correlation (LRU 10,000)
SQLite Storage	features.sqlite_storage.enabled	true	Persist to disk SQLite; disabled uses noop store
Interruption Detection	features.interruption_detection.enabled	true	Dead loop / crash / context overflow detection
Audit	features.audit	true	LLM call audit event persistence
Token Consumption	features.token_consumption	false	Aggregated token consumption records
SLS Logtail	features.sls_logtail	false	Write to SLS log file
Trajectory Collection	features.trajectory_collection.enabled	false	Periodically scan Qoder/QoderWork session JSONL, convert to ATIF v1.7 and store in trajectories.db (trace mode only; scan_interval_secs default 30, scan_dirs overrides scan roots)

Runtime Resource Limits (runtime_limits)

Configure buffer caps to prevent unbounded memory growth:

Option	Default	Description
event_channel_capacity	10,000	Bounded channel capacity for probe events
event_channel_policy	"backpressure"	Full-channel policy: backpressure / drop_newest / sample
pending_genai_max_count	1,000	Max pending events awaiting session_id
pending_genai_max_bytes_mb	64	Max bytes for pending events
pid_cache_size	1,024	PID → agent_name LRU cache size
max_connection_body_mb	8	Per-connection HTTP body buffer cap
connection_idle_timeout_secs	60	HTTP connection idle timeout (seconds)
ring_buffer_mb	32	eBPF Ring Buffer size (must be power of 2)

Minimal Memory Configuration

For resource-constrained environments, disable non-essential features and reduce ring buffer:

{
  "features": {
    "token_stats": true,
    "tokenizer": { "enabled": false },
    "session_mapping": { "enabled": false },
    "sqlite_storage": { "enabled": false },
    "interruption_detection": { "enabled": false },
    "audit": false,
    "token_consumption": false,
    "sls_logtail": false
  },
  "runtime_limits": {
    "ring_buffer_mb": 8,
    "event_channel_capacity": 5000,
    "pending_genai_max_count": 500,
    "pending_genai_max_bytes_mb": 32
  }
}

With this config: idle RSS ~24–30 MB, with event traffic ~35–40 MB.

Supported LLM Providers

Token parsing supports multiple LLM API formats:

• OpenAI / OpenAI-compatible APIs
• Anthropic (Claude, including cache token handling)
• Google Gemini
• Qwen (with native chat template support)

Origins

This project is derived from https://github.com/eunomia-bpf/agentsight.git.

License

Apache License 2.0 — see LICENSE for details.