Claude Workbench retirement: what teams must move before August 17

Anthropic announced on July 17, 2026 that access to Claude Console's legacy Workbench will end on August 17, 2026. Three experimental prompt-tool API endpoints retire on the same date and will return an error after removal. Saved work that remains only in the legacy surface will become inaccessible.

Teams need to answer two questions before the deadline: does anyone still rely on legacy Workbench data, and does any application call one of the retiring endpoints? If the answer to both is no—because the team never used legacy Workbench and does not call those paths—there is no migration work to perform. Record that result and move on.

The August 17 deadline affects stored legacy work and three API paths

The legacy Workbench retirement covers saved prompts, prompt variables, prompt versions or revisions, completions, and evals kept in that surface. Anthropic's Workbench help page says those materials will no longer be accessible after August 17. The updated Workbench does not support saved prompts, prompt versions, prompt sharing, or evals.

Legacy access is already limited. People who used legacy Workbench before June 16 retain temporary access through August 17 and should see a banner that links to the legacy version. Accounts that started using Workbench on or after June 16 do not have legacy access, so a new account cannot be used to retrieve another user's old work.

The same deadline applies to three experimental APIs for generating, improving, and templatizing prompts. These dependencies can exist in application code even when nobody currently opens Workbench, so the data review and the code search are separate tasks.

Refreshed Workbench does not replace legacy prompt storage

Refreshed Workbench is built on Anthropic's public Messages API. It lets a developer assemble and run a request, inspect the request and response, and export the current request as code. Anthropic says it does not store prompts or conversations on its servers; the current draft remains in the browser.

That design creates a clear persistence boundary. Legacy Workbench could hold reusable prompts, revisions, completions, and evals in the Console. Refreshed Workbench is a stateless place to try a request, so teams must keep durable prompt assets and evaluation material in tools they own. Anthropic gives the application repository as one example.

Claude Apps and Claude Code are separate products, and their storage or history behavior does not change this Workbench decision. The July 17 notice concerns the legacy Workbench in Claude Console and the three experimental API paths below.

The export preserves a copy but does not restore the old workspace

For a person who still has legacy access, Anthropic documents a direct export flow. Open legacy Workbench, select Export data in the top banner, choose whether to include saved model completions and uploaded images or PDFs, and then select Export. Including completions or uploaded files may materially increase the package size. Anthropic packages the data as JSON and emails a download link when it is ready.

The resulting JSON is a retained copy. Anthropic says legacy data cannot be imported into refreshed Workbench because the refreshed version does not save prompts or conversations. An available export therefore proves that data can be copied out; it does not establish a documented import, restoration, or rollback path.

Several operational details remain unpublished on the help page: a formal JSON schema, a checksum, a maximum package size, the email link's expiry, an import target, and a replacement for the three retiring endpoints. Preserve those as unknowns. Do not build the migration around an assumed file shape or an undocumented restoration process.

Search for the exact endpoint paths, not a general Anthropic dependency

The retiring endpoints are:

/v1/experimental/generate_prompt
/v1/experimental/improve_prompt
/v1/experimental/templatize_prompt

Search for each complete string in application repositories, shared libraries, scripts, CI configuration, serverless functions, and automation definitions. A general search for anthropic or Claude will create noise and can still miss a URL assembled from configuration. Check environment-specific base paths and request builders if the complete strings do not appear in source.

Anthropic says calls to these endpoints will return an error after removal. It does not name direct replacements in the cited release note or help page. For each match, identify what the caller expects—prompt generation, improvement, or templatization—then choose a supported approach in your own stack and test the caller's output and failure handling before changing production traffic. The refreshed Workbench's use of the Messages API does not, by itself, document replacement behavior for these prompt tools.

A short migration sequence is enough to make the decision

Keep the work tied to the assets and calls that actually exist:

1. Confirm access and dependencies. Ask prior legacy users whether they still have access, and search owned code and automation for all three exact endpoint paths.
2. Export the content worth retaining. Use the legacy banner before August 17, choosing deliberately whether saved completions and uploaded images or PDFs belong in the package.
3. Inspect and store the result. Open the JSON, confirm that the expected prompts, variables, revisions, completions, eval material, and selected files are present, then keep the package in an access-controlled location the team owns. Store reusable prompt or eval content with the application repository when that fits the team's development process.
4. Trace every endpoint match to a real caller. Identify the application, owner, purpose, inputs, outputs, and failure behavior before replacing the call.
5. Choose and test the replacement. Because Anthropic has not documented direct replacements in these sources, test the team's selected approach against the behavior each caller needs.
6. Record the remaining unknowns. Note any reliance on package shape, link availability, file size, or restoration behavior that Anthropic has not documented.

This is narrower than a general portability project. Our article on keeping support rules, knowledge, and handoffs outside an AI vendor covers the broader ownership problem for a customer-support system; it does not replace this deadline-specific export and code search. The new AI stack is being won in migration windows, not demos addresses model and API retirement planning across a stack, while this decision is limited to one legacy surface and three named endpoints.

Export or replace what exists; otherwise close the review

A prompt visible in a vendor console is useful today, but the work survives a surface change only when the team holds the needed copy, knows where application dependencies live, and can test what replaces them. The export helps with retention; its documented limits prevent the team from treating retention as recovery.

If you need help with one uncertain dependency, bring BaristaLabs one legacy export or one repository suspected of calling a retired endpoint. Our process automation service can help trace what exists, place retained material in an owned location, and test the replacement your team selects without widening the job into a generic AI migration.

If legacy prompts, revisions, completions, uploaded files, or evals matter, export and inspect them before August 17. If any of the three endpoint paths is in use, replace and test that dependency before the same date. If the team never used legacy Workbench and the exact searches return no dependencies, document that no action is required.

Sources

• Anthropic Claude Platform release notes, July 17, 2026 entry
• Claude Help Center: How do I use the Workbench?