Compare a decision.
Use Decision Audit with the expected values and restrictions you supply. Inspect mismatches that deserve a closer look.
Open Decision AuditFor developersMake the next failure
Make the next failure
harder to miss.
Start with the structured decision your assistant made. Supply the expected values and restrictions, inspect what differs, fix the cause and run a fresh case. One useful regression caught is worth a hundred unexplained scores.
Toolkit download: 0.2.31. Browser Bill Check: 0.2.30. They are separate releases.
Start with the artifact
Make the failure concrete.
Different questions need different checks. Choose the tool that matches the material you have.
Inspect the change.
Use Reply Compare to see the differences between outputs. Connect a visible change to the change you made in the workflow.
Open Reply CompareTrace the issue.
Use Assistant Audit to check a supplied conversation for its supported patterns. Bring the relevant context when you interpret the finding.
Open Assistant AuditA useful improvement loopA changed answer
A changed answer
is the beginning.
The bigger question is whether the changed assistant handles fresh work better. You run the baseline, make the change and test fresh cases. The tools inspect the material you bring to each step.
- Keep the baseline.
Save the job, the instructions, the assistant setup and the result before making a change.
- Name the intervention.
A corrected tool description, revised instruction or playbook should be a change you can point to.
- Test fresh work.
Use cases the changed assistant has not seen, and inspect critical errors alongside the result.
- Check what carries forward.
Repeat the work across variations before treating the improvement as dependable.
Connector build / private alpha
Six executable tools. No invented integrations.
The v0.1.0 connector build includes local capability discovery, bill arithmetic, Decision Audit, reply comparison, conversation audit and a bounded Handoff Check. It is a development package, not a public Gym installation inside ChatGPT, Claude, Cursor or another host.
Real code, limited reach.
The local stdio and private HTTP entry points and provider adapters belong to a private connector build. Its acceptance tests are internal checks, not customer use or marketplace approval.
The provider is not the product.
Adapters are present for GitHub, Google Drive, Gmail, Google Calendar, Microsoft Graph, Zendesk and Salesforce. Provider authorization, real-account acceptance and tenant approval are not established by the package.
The released browser and local tool downloads below remain separate packages. No new connector download is published on this page.
Read the connector approachKeep your assistantThe job should travel
The job should travel
across the stack.
The Gym is being built around the assistant people already use. Begin with outputs from your own workflow and the supported browser or local tools.
One assistant, more useful work.
A tested playbook should help the same assistant take on more of the jobs its owner needs.
One job, different assistants.
A meaningful job standard should make it possible to compare different configurations against the same requirements.
Native host integrations are part of the broader direction. The tools page gives you the actual browser links and local packages to start with.
Browse tools and downloadsLocal packagesUse a check.
Use a check.
Then build the loop.
Download the versioned packages and use their included instructions for setup and supported inputs.
Assistant Gym local toolkit
Version 0.2.31 · ZIP archive
Toolkit download: 0.2.31. Browser Bill Check: 0.2.30. They are separate releases.
Decision Audit
Version 0.1.0 · tar.gz archive
Build beyond a single reply
The interaction deserves its own test.
The broader Gym includes collaboration, negotiation, visiting assistants, supervision and evidence that survives a handoff.
Test the handoff.
Keep the brief intact when work moves between assistants, people and systems.
Make authority explicit.
Current instructions and permissions need to be available when an assistant considers an action.
Link claims to outcomes.
Use the available system and tool evidence to establish what actually happened.
Start with a failure
you can see.
Inspect a specific artifact and make the next change deliberate.