STEP 01
ON YOUR MACHINE
Connect
Point it at a local stdio server, a remote SSE or HTTP MCP endpoint, an A2A card, or an imported OTEL trace.
→AGENT STUDIO OPEN · MCP AND A2A DEBUGGING IDE
Open the agent. Own the trace.
Connect to your MCP and A2A agents, inspect the tool path, replay the run that failed, and export a regression suite your CI can run. It is a desktop IDE. The trace stays on your disk.
Source listed for review. Package, license, and desktop release proof must be verified before they are marketed as available. It is the local inspector that sits behind the Models app in App Data OS.
On disk
the trace never leaves
Portable
exports your CI can run
No account
to debug or export
HOW IT WORKS
Three steps. No orchestration framework.
Connect the server. Replay the path. Ship the regression. The trace stays on disk.
STEP 02
THE AGENT CAN'T GAME IT
Replay
Record the agent path. Replay it with mocked tool outputs. Compare per-turn and per-tool across model versions.
→STEP 03
A REVIEWER CAN OPEN IT
Export
Generate a GitHub Action, JUnit report, PR comment, trace card, or AuraOne intake packet for the next reviewer.
THE PRODUCT SURFACE
The agent debugger should show the path it is debugging.
Captured from the running app: connect the endpoint, inspect the tool trace, replay the failed run, compare model behavior, and export the CI suite. One desktop IDE. Nothing here touches our servers.
CONNECT MCP/A2A
Connect to a local server or remote endpoint.
The workbench shows transport, command, manifest, risk scan, and lifecycle state before any trace is recorded.
INSPECT TOOL TRACE
Inspect the actual tool-call path.
Tool inputs, outputs, retries, timing, and state transitions stay visible without pushing the run into a hosted debugger.
REPLAY DETERMINISTIC RUN
Turn a failed run into a deterministic replay.
Mock tool outputs, lock the path, and make the next model or prompt revision prove it still clears the case.
COMPARE BEHAVIOR
Compare model behavior against the same trace.
Replay diffs, model deltas, latency, and outcome changes sit beside the trace so reviewers can isolate what moved.
EXPORT CI REGRESSION
Export the trace card and CI regression suite.
Ship repo-ready artifacts: trace cards, JUnit, GitHub Actions, PR comments, and AuraOne intake packets.
WHAT COMES OUT
Every run leaves a portable artifact.
Repo-ready files. No hosted account. After a competitor lost four terabytes — including who its workers were — nobody wants tooling that pools their data. This never does.
01
Trace cards
Portable Markdown and JSON for one agent run: tools, retries, data touched, outcome, failure mode.
↳ ARTIFACT
02
Regression suites
Every failed tool call becomes a deterministic replay the next release candidate must clear.
↳ ARTIFACT
03
GitHub Actions
A drop-in workflow file that runs the replay set on every push and posts findings to the PR.
↳ ARTIFACT
04
JUnit reports
Standard XML for the CI dashboard your team already runs. No new viewer required.
↳ ARTIFACT
05
Intake packets
Packaged .auraonepkg with a privacy preview before handoff to AuraOne reviewers.
↳ ARTIFACT
SOURCE AND RELEASE
Read the source. Check the license. Then install.
MIT-oriented source is listed on GitHub. Package, release, checksum, desktop trust, and platform proof are required before install or binary availability claims are marketed.
License
MIT — source link listed; license proof required.
Source link listed
Source
github.com/auraoneai/agent-studio-open
Source link listed
Install (macOS)
Package release proof required before install commands are marketed.
Release proof required
Desktop artifact
Binary artifact, signing, and notarization proof required.
DMG proof required
Checksum
Checksum proof required before checksum claims are marketed.
Checksum proof required
Browser IDE
Run it in the browser. Nothing leaves the tab.
Preview proof required
Platforms
Desktop platform proof required.
Desktop trust proof required
Changelog
Release and changelog proof required.
Platform proof required
RELATED OPEN SURFACES
Next to this in AuraOne Open.
RUBRIC STUDIO OPEN
Author criterion-level evals on disk.
File-based, git-friendly. Write the rubric, run it, keep the code.
See the page →ROBOTICS STUDIO OPEN
Review robot datasets without uploading the robot.
Scrub synchronized sensor streams. Cluster failures. Export reviewed subsets.
See the page →TRUST TOOLKIT
The provenance machinery an audit asks for.
Eval manifests, regression banks, contamination audits. Run them yourself before the EU AI Act clock hits August 2026.
See the page →AGENT STUDIO OPEN
Your work. Your data. Your tools.
Your tool calls, your replay artifacts — all on disk, no telemetry by default. LangSmith, Langfuse, Braintrust, and Arize trace what happened in the cloud. Agent Studio inspects it locally. Tracing is not release governance: send the intake packet to the Models app in App Data OS and the failed run becomes a signed release gate — and weights you keep.