Continue kept swapping to Claude when I set Ollama
Spent last week on Continue + Ollama (qwen2.5-coder:14b) for a Rust CLI refactor.
Config clearly pointed at local. Half the turns still billed Anthropic until I noticed the usage graph. Free Continue tier is fine — the silent fallback is what got me.
Anyone found a hard "never leave localhost" switch, or is that just vibes?

5 comments
Join the discussion
Log in to comment.
Same. I pinned
provider: ollamainconfig.jsonand it still "helpfully" fell back when the local server took >8s. Only saw it in the provider logs, not the UI.Silent cloud fallback is a bug wearing a feature badge.
yeah the 8s thing is real. i wrapped ollama in a tiny proxy that always returns 200 within 200ms with a keep-alive, then let continue talk to the proxy.
still not a real "never leave localhost" switch. just stopped the silent swap.
yeah i mostly demoted Continue to commit messages on Windows. anything multi-file and it invents paths that never existed in the repo.
do you keep a separate profile just for local, or one config fighting itself?
Separate profile, always. One for local-only, one for cloud — mixing them is how you get that silent fallback.
For the path inventing: I log every file touch and abort the session if the path is not in
git ls-files. Half my "multi-file" runs die early now, which is honestly better than cleaning up invented modules. Un po' drastic, but the eval numbers got cleaner.I instrumented this last month for an eval harness. Continue's "auto" mode treats Ollama timeout as a soft fail and retries Anthropic if an API key is in the env — even when
provideris set.Only hard lock I found: strip every cloud key from the process env, then run with
CONTINUE_GLOBAL_DIRpointing at a folder that has no secrets. Ugly, but my routing sheet finally stopped showing surprise Claude spend.