PROTOTYPE / DESIGN PARTNER PREVIEW
What we actually reproduced.
Functional evidence from local emulation and deployed Cloudflare infrastructure.
On September 27, 2026, we tested two things separately: an Effect-native application with persistent tools, and Alchemy’s infrastructure engine embedded in a Worker. Dependencies were pinned; source is included in the prototype download.
Application results
| Experiment | Observation |
|---|---|
| 14 HTTP checks, local emulation | All passed with dummy cloud credentials. |
| Same 14 checks, live deployment | All passed; repeated after redeployment. |
| Existing records after redeployment | Original version and revoked-tool status preserved. |
| Existing records after local restart | Original version and revoked-tool status preserved. |
| Application and stack typecheck | Passed strict TypeScript checking. |
Engine results
A deployed Worker ran Alchemy’s planner and apply engine to create, rename and delete a real KV namespace. The rename preserved the namespace ID. A separate cloud inventory check verified cleanup.
The working integration bypasses default CLI/platform setup and injects services explicitly. The ordinary empty-stack deployment succeeded in Bun but failed in the local Workers runtime while trying to create /bundle/.alchemy.
Browser result
Effect bundled successfully for a browser target. Direct Alchemy deployment and planner/apply bundles failed on transitive Node imports. We did not demonstrate browser-hosted provisioning.
What these tests do not establish
No production SLA, independent sandbox audit, multi-provider compatibility, durable deployment recovery, infrastructure concurrency safety, cost benchmark or autonomous model loop was tested. The repeated live suite is still the same 14 cases.
The duplicate-request check repeats an application mutation; it does not simulate a lost cloud-provider response. The concurrent-update check concerns D1 records, not competing infrastructure plans.
Inspect it yourself
Run the prototype. The downloadable bundle includes a sanitized evidence summary and the test drivers. The temporary cloud resources used for the original experiment were deleted.