Continue

Continue was the open alternative to Copilot, and its configuration wired every model you could name, including the ones on your own machine, which is why people loved it and why leaving hurts. The trait that decides it now is maintenance: there is none, and an editor extension with no maintainer is a countdown you cannot see. It still works today, which is the most dangerous state for a dependency.

If you are on it, move to Cline this quarter; it covers the same ground with a live upstream and the migration is an afternoon. If you are not on it, do not start.

The risk is decay. A VS Code extension that no longer ships releases keeps working until the editor API moves under it, and then it breaks on a Tuesday with no one to fix it and no release to roll forward to. That is the failure mode of every archived extension, and it is certain; only the date is unknown, and it is not a date you choose.

Pin the editor version if you must stay, and put a migration ticket on the board now. What it did right: Apache-2.0, which leaves the door open for a fork, and a fork is the only fix that will ever ship.

Continue's design is documented and, with no vendor steering it, effectively frozen. Models are declared per role, chat, edit, autocomplete, from providers including Ollama, LM Studio and llama.cpp, a reproducible configuration surface uncommon in its clarity: a reader can state exactly which model handled which step. Maintenance ended with the acquisition, so the design will not change and neither will its bugs.

No benchmark was published. The consequence is that the architecture is a reference, not a product. The observation: the configuration format will outlive the tool, being the part people copied.

Continue is my case for open source as a marketing plan. Apache-2.0, roughly 36,000 stars, and a thin hosted service for revenue, which is a funnel with nothing at the bottom. In June 2026 Cursor acquired the company, shut the hosted services and stopped selling the product within the same announcement. A talent purchase, not a product purchase; the product got switched off because the buyer already had one.

The acquirer is named and the pivot is complete. Position: closed on the product; long the lesson, which is that stars do not convert to runway.

The demo was local-model autocomplete, which had our data-residency people interested, because a model on the laptop is a model that never sees a vendor. Procurement has no counterparty: the company was acquired, hosted subscriptions ended, and nobody signs a security questionnaire for a codebase with no owner. Sixty engineers on frozen code with sixty keys is a liability with no vendor to name, which is the worst kind.

Onboarding cost is now migration cost, and the sooner the cheaper. Not yet, not again in this form; fund the move and keep the residency requirement for the replacement.

The company got bought and the repo got frozen; my disk did not notice. Continue is Apache-2.0, and config.yaml declares every model by role, Ollama and llama.cpp included, so it runs air-gapped from a file I already version. MCP servers go in the same file under mcpServers. Read-only means no fixes when VS Code changes, so I am the maintainer now, which I have been for worse software.

Grudging respect for the config format; the rest is a fork waiting for a name, and the license means nobody has to ask.