Guides · 4 min read ·

Continue.dev Model Config: Roles, Providers and Teams

Continue.dev Model Config: Roles, Providers and Teams

Continue.dev puts model configuration where it belongs: in a file you can read, version and review. Each role — chat, autocomplete, embeddings — takes its own provider, endpoint and key, which makes Continue the most cost-tunable client in existence. Most users accept defaults and overpay on every keystroke; the configured minority spends fractions for identical work. This guide shows per-role setup, custom provider wiring, verification routines and team patterns that keep billing sane permanently.

Per-role models, the core skill

Configure each role deliberately instead of accepting one model everywhere. Chat and debugging deserve flagship reasoning — this is where judgment errors cost hours. Autocomplete deserves the cheapest fast model that completes sensibly; developers accept far more from ghosts than from agents. Embeddings deserve whatever costs least per token, full stop — no human ever reads them.

Write the reasoning as config comments so future-you and teammates understand each choice. Review assignments quarterly like code, because this config spends money with every keystroke. Teams that configure by role typically halve bills versus single-model setups with zero perceptible difference in daily work. The config file is a budget document wearing technical clothes.

Wire custom providers cleanly

Add custom providers with base URL, key and exact model strings copied from catalogs — never from memory, as identifiers churn constantly. Verify each role separately with a trivial task after configuring; chat, autocomplete and embeddings fail independently with confusing aggregate symptoms. Back up the last working config before every change so rollbacks take seconds.

Change one provider or role per round with tests between; compound changes create compound mysteries. Compare new configurations against known-good baselines to split client versus credential instantly. The whole setup takes under an hour methodically and repays forever. Document working combinations per provider so team setups copy instead of being rediscovered painfully.

Prove savings with telemetry

Continue surfaces acceptance rates, request counts and error data — use them. Run identical work periods on default versus tuned configs and score acceptance plus task success before price. A cheap configuration developers constantly override is not cheap; it taxes attention on every keystroke. Add dollars per developer per week as the final column from each key's price list.

Require two full weeks before standardizing any configuration change. Keep the previous config warm for a month regardless; providers have outage days and the rollback is a file swap. Publish results internally with methodology — telemetry-driven decisions stick where opinions get relitigated. The data habit matters more than any single configuration.

Teams without secret chaos

Layer configurations: shared base in git holding everything except secrets, personal keys per developer, one team gateway key with budgets for shared automation. Secrets never enter version control; routing never lives in heads. New hires clone the config, paste one key, and match team economics immediately — onboarding measured in minutes, not mentoring sessions.

Budget every shared key before growth reaches it; unmetered team automation funds expensive lessons overnight. Review per-developer spend monthly to spot misrouted roles and runaway loops early. The system stays boring deliberately: files, variables, quarterly reviews. Boring billing infrastructure is what lets exciting agent work happen without finance involvement.

Keep configs optimal forever

Review every role assignment quarterly against current prices and capabilities. The value ranking rotates fast; last quarter's bargain becomes this quarter's overpay silently. Use telemetry to justify downgrades — steady acceptance on cheaper models is permission to switch permanently. Check all keys' prices, not just headline models; providers move quiet ones without announcements.

Keep fallback configurations warm for outages and deprecations. The maintenance burden is an hour per quarter protecting savings that compound every keystroke of every developer. Teams that review regularly spend less every single quarter while changing nothing about daily work. Configuration is a living budget document — treat it that way permanently.

Frequently asked questions

How do I configure models in Continue?

Edit the config file: each role (chat, autocomplete, embeddings) takes a provider, endpoint, key and model. Validate syntax, test each role with a trivial task, and keep a backup of the last working config always.

Which model per role?

Flagship for chat and debugging, cheap fast models for autocomplete, cheapest embeddings available. Per-role assignment is Continue's killer cost feature — one flagship everywhere wastes most of the budget.

What breaks custom provider setups?

Base URL, key and exact model string, each verified alone. Compare against a known-good config to split client versus credential in minutes instead of hours.

How do teams share Continue configs?

Yes. Version-control the base config with secrets overlaid per developer, plus one shared gateway key with budgets for automation. New hires clone and paste one key.

How do we keep configs optimal?

Review every role assignment quarterly against current prices, track acceptance rates per role, and downgrade where telemetry permits. An hour per quarter protects compounding savings.

Related guides

Keep your coding workflow moving

Kintio provides an Anthropic-compatible API for Claude Code and other developer tools. Review the fit for your workload before switching.

Explore Kintio →

Last updated: