GitHub Spark

Do not adopt. Spark was a tidy way to turn a sentence into a small hosted app with auth and data already wired, and it is over: existing users lose access on August 31, 2026, so the only action left is exporting what you built to a repository before then.

If you liked the loop, the loop exists elsewhere. Pick Lovable for the hosted experience where the backend comes with the app, or v0 if the front end is the whole product and you will host it yourself. Treat Spark as a lesson in reading the deprecation notice before the pricing page.

The worst part happened before the shutdown. Spark's llm() function ran on GitHub Models, retired on July 30, 2026, so deployed apps with AI features broke a month before the product itself. That is what a single-vendor stack does: the pieces die in the wrong order, and users find out from an error message rather than an email.

The consequence for anyone with a spark still running is that the AI path is already dead and the hosting follows. There is nothing to fix here, only to leave. What it does right: the export path is one menu, open the workbench, select Create repository.

The design was closed on purpose. 1. Hosting, a data store, GitHub authentication and model inference were pre-configured, so the app never chose a dependency and never had to. 2. Generation ran on Claude Sonnet 4, a single fixed model, with no route to a different one. 3. No terminal tool and no MCP client, so verification was the browser preview and nothing else.

The consequence is that every layer had exactly one supplier, and when one supplier changed, nothing could be swapped. As an experiment in how far pre-wiring can go, it answers the question. The observation: a stack with no seams cannot be kept alive.

GitHub launched Spark in July 2025 as a Copilot Pro+ perk at $39 a month and pulled it in August 2026: thirteen months, no standalone price, no defensible revenue line. A feature inside a bundle never gets to prove it could sell itself, and when the bundle owner reprioritizes, it is the first thing cut.

Microsoft does not need the product, it needs Copilot, and the deprecation notice sends users to Copilot, which tells you what the experiment was for. No acquirer, since nothing is for sale. Position: zero, and a data point on bundle economics for anyone evaluating a perk as a product.

The demo made an app from a sentence. Procurement cannot proceed: no new users since August 4, 2026, so there is no way to onboard sixty engineers, and no way to renew the ones who had it. While it lived, usage drew on the Copilot plan's premium requests and later AI Credits, a meter shared with everything else Copilot does, so nobody could have budgeted it separately.

No enterprise tier existed, no SSO story, no retention statement specific to sparks. Nothing ran in CI. The one useful outcome is that any spark a team did build can be exported into a repository we govern. Not yet.

Nothing to bend. No model choice, no key field, no MCP client, no shell, and no source, so the only thing I ever controlled was the prompt. The one useful line is in the deprecation notice: to keep AI features working after export, replace the inference with your own provider and supply your own API keys. BYOK arrives as a funeral instruction.

That is the pattern to remember: the escape hatch a closed tool never offered while it was alive becomes the migration guide when it dies. Export, own the code, wire your own key, move on. Next time check for the key field first.