v0.3 The agent execution workstation

Nothing hidden.
Every step in view.

A local workstation for agent execution. Follow the trace, intercept a request, fork a run, and verify what changed. Free and open source.

View on GitHub
MITFree & open sourceLocal-first
FLIGHT DECK / v0.3LOCAL-FIRST · OPEN SOURCE

From the first call
to the final verdict.

Observe

Trace agents, models, tools, retries, HTTP requests and errors in a live DAG.

Intercept

Pause execution, inspect state, mutate prompts and mock tool responses.

Replay

Fork runs and reuse cached execution prefixes for controlled experiments.

Compare

Align spans semantically to reveal changes in behavior, cost and latency.

Evaluate

Check reliability, tokens and topology with structured assertions.

Optimize

Find cost hotspots, critical paths, retry chains and prompt bloat.

THE v0.3 WORKSPACE

From debugging to verified experiments.

Change a prompt. Try another model. Replay the variant. Compare what changed and evaluate what improved.

06 / Experiment Studio

One run. Multiple possibilities.

Create persistent experiments from existing traces. Mutate prompts, models, tool arguments or mock responses, batch-replay variants and preserve their ancestry. Review replay history and send results to Diff Arena.

07 / Evaluation Studio

Better agents. Concrete evidence.

Evaluate cost, latency, tokens, reliability and execution topology with safe, structured assertions. Inspect error clusters and root causes, review optimization signals and export redacted JSON or Markdown reports.

08 / Semantic Diff Arena

Not just different JSON. Different behavior.

Align executions semantically to compare tool calls, models, retries and graph structure. See cost and latency deltas alongside lineage and provenance differences before deciding whether a change is a regression.

01 / Visual debugging

Follow every step. Find the failure.

Navigate the live DAG and timeline, pin important spans and aggregate selected metrics. Follow retry chains, critical paths and causal relationships back to the source.

02 / Time travel

Rewind and fork with context intact.

Fork a previous execution while preserving lineage and provenance. Reuse cached prefixes where available. Choose cache-only, safe or live replay; live replay requires explicit confirmation.

03 / Breakpoints

Pause and edit thoughts mid-flight.

Live breakpoints pause matching requests for inspection. Mutate prompts, tool arguments or mock responses before resuming. Conditional expressions run without arbitrary JavaScript evaluation.

04 / Chaos engine

Break your agents before production.

Inject synthetic 429 rate limits, 502 upstream crashes, latency delays and malformed JSON.

Chaos Engineering Engine
Global Chaos Fault Injection
Probabilistically inject failures across all agent calls
Artificial Latency Delay:2500 ms
Synthetic HTTP 429 Rate Limit:25%
Synthetic HTTP 502 Upstream Failure:10%
Malformed / Corrupted JSON Rate:15%
Cancel Arm Chaos Engine

05 / Trace bundles

Share bugs like Wireshark packets.

Export runs as portable .trace.json bundles and import them into another Flight Deck. Preserve lineage, provenance and replay context. Existing v0.1.0 traces remain compatible.

ESC
status:errorevent:toolcost:>0.10
  • Pause Execution
  • Toggle Execution Timeline
  • Toggle Breakpoints Interception
  • Open Run Diff Arena
  • Open Chaos Engineering Engine
  • Export Trace (.trace.json)↵
K to open

How it works

Start local. Trace your first run.

  1. 01

    Start the proxy and Flight Deck

    npx aethertrace start
  2. 02

    Point your SDK baseURL at the proxy

    baseURL: "http://localhost:4000/v1"

    Works with OpenAI, Anthropic, Ollama or local models.

  3. 03

    Open the Flight Deck

    http://localhost:3000

Optional: npx aethertrace sim dispatches a 5-agent DAG simulation.

Integrate your stack

Your agents. Your telemetry. One local view.

OpenTelemetry

Import OTLP/HTTP Protobuf-JSON traces and map GenAI attributes into native agent events. Imported telemetry is redacted before storage.

POST /v1/traces

LangChain & LangGraph

TypeScript and Python callback handlers preserve parent-child relationships across LLM, tool and chain execution.

LLM · TOOL · CHAIN

SDK instrumentation

Wrap OpenAI calls, tools and agents. Context propagation preserves ancestry across nested TypeScript and Python execution.

wrapOpenAI · wrapTool · wrapAgent
  • OpenTelemetry
  • LangChain
  • LangGraph
  • TypeScript
  • Python
  • OpenAI
  • Anthropic
  • Ollama

Local-first by design

Your traces. Your workstation.

SQLite (WAL) on your machine

Everything lives locally in .trace.sqlite.

No cloud account required

Run locally or in a developer-controlled environment. Docker support includes persistent storage and healthchecks.

MIT licensed

100% free and open source. No subscription or hosted backend required.

Security built in

Recursive secret redaction, SSRF protection and a security audit trail. Search and exported reports are scrubbed.

Reproducibility with provenance

Capture Git, source, dependency and environment context. See warnings when the conditions for reproduction change.

Replay with guardrails

Distinguish pure, read-only, mutating and unknown tools. Mutating tools can be mocked; live replay requires confirmation.

Your agents.
Understood.

GitHub