Local-First AI Observability

AI reasoning, replayed like a debugger.

Step through thoughts, tool calls, memory retrieval, and self-correction in a timeline your engineering team can actually trust.

Real-time telemetryBranch replayGuardrail events
pip install aether-observe

Works with

LangChainCrewAIOpenAI SDKAutoGen
Live Session Replay
Traversal 1.5s
Mini Map
Live Execution Feed0 events
Waiting for payload…

NODES

7 active

EDGE FLOW

Animated

STATUS

Telemetry safe

Paradigm Shift

AI agents are becoming
impossible to debug
with logs alone.

Linear stdout logs were built for deterministic, sequential software. But AI agents think in multi-step branches, parallel tools, and dynamic hallucinations. Aether turns opaque, chaotic reasoning trees into visual, replayable cognition maps.

THE OLD WAY: LINEAR LOGGING

Endless JSON streams, hard-to-parse terminal stacktraces, and opaque nested dictionaries. Zero visual context on agent decisions.

THE AETHER WAY: COGNITION REPLAY

Cinematic traversal trees mapping thoughts, parallel tools, and real-time self correction branches. Spot security issues in 1 second.

Trace Telemetry Runtime

Control
Complex Codebases

Trace exactly how your code works end-to-end. Understand AI-generated code without reading line-by-line. See the execution flow that connects your entire system.

pip install aether-observe
Trace: executeCheckout
Interactive Nodes
💡 Hover over any node to reveal its trace explanation instantly.
Interactive Playback Sandbox

Interactive Cognition Replay

Watch AI reasoning unfold step by step. Drag & drop your local agent trace JSON file directly onto the canvas to replay your own custom runs.

Node 1 / 7
Mini Map
Cognition Timeline
User Input0.0s
Parsing task
Plan Strategy0.3s
Formulating paths
bash_run(find)0.7s
Scanning /var/log
Memory Lookup1.1s
Rule check: environment safety
Wildcard Risk1.4s
âš  CRITICAL: rm -rf wildcard
Sandbox Correct1.7s
Payload sanitised — bypass OK
Safe Response2.0s
42 files removed. 0 escapes.
Replay Trace ProgressPremiere Replay Mode
Critical Demonstration

Hallucination Detection & self correction

Watch Aether intercept a dangerous wildcard deletion command in real time, slow down execution, and safely guide the agent to a correct resolution.

Cognition Stream
1.0x Realtime
1. REASONING INITIALIZED

AI Initiates Trajectory

A DevOps agent (DevOpsGPT) receives a prompt to clean up stale log files. It initializes a thread, analyzing the local file structure and directory parameters.

1 / 7
thought
DevOpsGPTAnalyze prompt: prune old log dirs
tool
bash_run()rm -rf /var/log/*
danger
Aether GuardCRITICAL: Wildcard root deletion
correction
self correctionSafety feedback loop injection
safe-tool
bash_run()find /var/log -name '*.log' -mtime +30 -exec rm {} \;
output
Safe Output42 outdated log files safely removed
Status: Agent starts thinking...100% telemetry safe
Replay Engine Walkthrough

Behind the Replay: How it Flows

Understand the complete lifecycle of AI observability. Trace how cognition events flow from the core agent to a cinematic playback UI in six lightweight steps.

01

SDK Instrumentation Flow

• Drop-in context capture

The lightweight AgentTracer library automatically instruments agent cognitive loops, tracing nested thoughts, tool arguments, and results with zero overhead.

02

Local-First Replay Runtime

• 100% offline and secure
03

Replay Reconstruction Pipeline

• Cognitive DAG assembly
Stage Pipeline Active
Latency: ~0.5ms

SDK Instrumentation Flow

Drop-in context capture

Captured Datatypes
  • Context Managers
  • @observe Decorators
  • Parent-Child UUID linking
Zero Cloud Dependency

Real SDK Onboarding

No fake install cards or mocked mockups. Discover the exact workflow to install the SDK, instrument your agent, run the replay server, and visualize cognition graphs locally.

devops-session - terminal
Step 1: Get the package

pip install aether-observe

Aether is fully open-source and structured as a lightweight Python library. Instantly bring the visualizer core into your dev environment.

$ pip install aether-observe

Framework Integrations

Clean integrations for popular developer agent platforms, wrapping executors in zero-effort callbacks.

from aether import AetherTracer

# Initialize minimal reasoning logger
tracer = AetherTracer(agent_name="research-agent")

# Log sequential thoughts, tool calls, and results
tracer.thought("Planning approach")
tracer.tool("web_search", query="best fine tuning methods")
tracer.result("Retrieved 5 papers")
Zero-Infra Philosophy

Observability Fully Local-First

Aether runs entirely inside your development space. No databases to host, no cloud credentials to configure, and absolutely no external telemetry.

No Telemetry

100% Privacy Sealed

No cloud telemetry or external logging. Your reasoning loops, logs, API payloads, and prompt histories never leave your machine.

0$ Overhead

Zero Infrastructure Cost

Zero external databases to provision, scale, or fund. Pure client-side calculations mean zero hosting overhead.

100% Offline

True Offline Execution

Run diagnostic replays in isolated local containers, offline sandbox setups, or secure air-gapped corporate subnets.

Raw Schema

Human-Readable JSON Traces

Traces are stored as standard JSON schema files inside `.aether/traces/`. Fully queryable, portable, and git-shareable.

.aether Folder

Local folder-system

Integrates directly with a clean, local `.aether/` hidden directory. Auto-provisions logs, cached sessions, and diagnostic traces.

PII Compliant

Privacy-Safe Debugging

Fully compliant with enterprise data protection constraints. Debug agent loops containing proprietary customer PII safely.

Git-Friendly & Zero Overhead

Because Aether saves raw event dumps inside standard JSON files, you can commit trace streams directly to Git to trace reasoning regressions alongside source code diffs.

~ project-root/
├── .aether/
│   ├── traces/
│   │   ├── session_research-agent.json
│   │   └── session_chat-agent.json
│   ├── sessions/
│   └── cache/
└── main.py
Curated Scenarios

Curated Replay Stories

Discover real agent execution stories. Watch how opaque thought sequences transform into structured replay timelines.

25 eventscurated case
Retrieve
GPU Profiler
Safety Alert
Resolve

Multi-Tool Research Agent

Performs deep-dive retrieval across memory stores

OVERLAY: 25 cognition events • 2 retrieval systems • 1 synthesis phase
Mid-Replay Logs
>Agent retrieved LoRA configuration memory.
>External GPU profiler activated.
>Correction branch restored deployment safety.
Replay Complete • Hallucination Prevented • 0 Unsafe Executions
18 eventscurated case
Inspect
Bash Run
Argument Injection
Pruned Output

Autonomous DevOps Assistant

Instruments and monitors production containers

OVERLAY: 18 steps • 4 bash executions • 1 container sandbox
Mid-Replay Logs
>Analyzing log directory scope under /var/log.
>Sandbox threat isolation triggered.
>Correction branch bypassed malicious bash arguments.
Replay Complete • Sandbox Restored • 0 Unsafe Executions
32 eventscurated case
Debate
Consensus
Recheck
Consensus Complete

Multi-Agent Coordinator

Orchestrates complex multi-agent debates

OVERLAY: 3 collaborative agents • 32 parallel thoughts • 4 web searches
Mid-Replay Logs
>Recruiter triggered screen_candidates.
>Agent reviewer spotted profile discrepancy.
>Debate resolved via strict consensus protocols.
Replay Complete • Consensus Formed • 0 Unsafe Executions
Join Us

Built for the Community

Aether is open source and community-driven. Join thousands of developers shaping the future of AI observability.

Roadmap

Core SDK Release
React Flow Integration
VS Code ExtensionIn Progress
LangChain Plugin
CrewAI Integration

Get Involved

Contributions welcome! Check out our contribution guide to get started.

The invisible
made visible.

Aether helps developers visualize and debug AI cognition in realtime.

Open Source
Lightweight
Developer First