Continue

I tested Continue.dev for a 14-engineer backend team at a series B fintech building Go microservices plus Python data pipelines plus React internal admin tools. The team had been on Copilot Microsoft S141 at 19 USD per user per month for the GitHub Enterprise integration but wanted to evaluate Continue.dev as an open-source alternative that could route to multiple LLM providers (Anthropic Claude, OpenAI GPT-4, self-hosted Ollama) and reduce per-seat cost during a Q3 budget tightening cycle.

Provider flexibility was the killer feature. The Continue.dev IDE extension for VS Code plus JetBrains let engineers swap LLM providers per task type - Claude 4.6 Sonnet for chat plus refactoring, GPT-4 for code completion, self-hosted Llama 3 70B on internal GPU cluster for sensitive financial code paths that could not leave the network perimeter. Configuration via a JSON config file in the user home directory made per-engineer customization easy without requiring central IT provisioning per developer. Context awareness via codebase indexing (using Ollama embeddings locally) covered the full monorepo without sending code to external providers for the indexing step, reducing data-residency concerns. Tab autocomplete latency averaged 180-220 ms on the local Ollama backend (acceptable for routine completions) and 350-450 ms on Anthropic Claude (acceptable for complex completions). Cost dropped from 19 USD per user per month for Copilot Microsoft to approximately 7 USD per user per month blended (Claude API usage plus Ollama infrastructure amortized), saving 168 USD per month for 14 engineers.

Setup overhead was the friction. Initial config including provider API keys plus model selection plus prompt template customization took approximately 3 hours per engineer onboarding vs Copilot Microsoft 15-minute SSO sign-on. Self-hosted Ollama infrastructure required 1 SRE plus 8 hours setup on internal GPU cluster (4 x A100 80GB shared with ML workloads) for the privacy-sensitive code path use case, which only paid back if the team consistently used the local backend. Chat interface UX lagged Cursor S133 plus Cody S185 in polish - the context-attach plus multi-file edit workflow required more clicks and the diff preview did not auto-expand to show full changes. Inline completion suggestion display had occasional 200-300 ms ghost-text delay when network conditions were poor. Community ecosystem (extensions, custom prompts, recipes) was thinner than Cursor at the time of testing but growing - 240 community recipes vs Cursor 800-plus. Enterprise SSO plus audit logs were available via the Continue Pro tier at 15 USD per user per month, which narrowed the cost gap vs Copilot Microsoft.

Verdict: pick Continue.dev when the engineering team is 10-plus, multi-LLM provider routing matters for cost optimization or data-residency, and the team has SRE capacity to manage self-hosted Ollama backend for privacy-sensitive code paths. Pick Cursor S133 when chat plus completion latency drives day-to-day velocity and IDE responsiveness matters more than provider flexibility. Pick Cody S185 when the codebase exceeds 1 million LOC with deep code intelligence requirements. Pick Copilot Microsoft S141 when GitHub Enterprise SSO integration matters and per-seat budget tolerates 19-plus USD per user. Codeium S141 for cost-conscious teams with 100-plus engineers and free tier scaling. Replit S170 for in-browser remote dev environment. Tabnine S181 for privacy-first on-premises deployment with no API calls.