Why We Use Continue.dev for Refactors But Keep Cursor for Quick TypeScript Edits

TL;DR: Continue.dev + Ollama local models handles 3–5 hour refactors with persistent terminal context and 65% cost savings vs Cursor Pro — but zero LSP integration means type errors only caught at build. We use Continue.dev for migrations/debugging; Cursor for <30 min type-heavy edits. Full comparison →

The Context

Three-dev team, M3 Max MacBooks, 32GB+ RAM. Daily AI coding: 4–6 hours. Stack: TypeScript/React/Next.js (strict mode), Go, PostgreSQL. Cursor Pro ($20/mo/dev) was default Jan–May 2026. Pain points: Composer context loss >90 min, $200+ monthly team bill, no token routing. Evaluated Continue.dev (v0.9.x) with Ollama local models June 2026.

What We Tested

Tool	Use Case	Verdict	Why
Continue.dev + Ollama (qwen2.5-coder:32b / deepseek-coder:33b)	3-hr Stripe webhook migration (14 files)	✅	Persisted stripe listen tunnel + Supabase connection entire session; $0 marginal cost
Continue.dev + OpenRouter (Claude 3.5 Sonnet)	Greenfield API endpoint	⚠️	Chat UI friction for terminal work; no autonomous loops
Cursor Composer (Pro, Claude 3.5 Sonnet)	Quick TypeScript edits (<30 min)	✅	LSP diagnostics in Composer panel catch type errors mid-edit; fast for type-heavy work
Cursor Composer	3-hr refactor	❌	Lost tunnel/DB context 3×; 45 min recovery; $12 in API costs for single session
VS Code + Cline + OpenRouter	Greenfield feature	✅	Terminal-native autonomous loops; BYOK cost control

The Pivot Point

June 5: 3-hour auth refactor (14 files, Stripe webhook → new provider). Cursor Composer lost stripe listen tunnel + local Supabase 3×. Each recovery: restart tunnel, re-seed test data, re-explain schema. Same task in Continue.dev (Zed) — agent kept tunnel PID, DB connection, full file context. Finished in one sitting. Cost: $0 (local models). Next day: Quick TypeScript interface refactor (5 files, strict types). Tried Continue.dev — agent wrote code that passed local build but introduced 3 type errors caught only at pnpm build. Switched to Cursor: Composer showed red squiggles inline while typing. Fixed in 12 min. Pattern locked: Continue.dev for long sessions needing context persistence; Cursor for type-heavy edits <30 min.

What We Use Now

Continue.dev (free) + Ollama (local) for sessions >90 min:

• Models: qwen2.5-coder:32b (planning, 8k ctx), deepseek-coder:33b (execution, 16k ctx)
• Config: .continue/config.json with tabAutocompleteModel: "deepseek-coder:6.7b" (fast inline)
• Context providers: terminal, files, codebase, docs — no diagnostics provider (gap)
• Team convention: continue in Zed terminal for refactors/migrations/debugging

Cursor Pro ($20/mo/dev) for <30 min type-heavy edits:

• Model: Claude 3.5 Sonnet (default)
• Used only when: editing TypeScript interfaces, generics, complex type narrowing
• Team convention: cursor for “type surgery” — close immediately after

When You’d Choose Differently

• Aider (free, terminal) if you want git-native BYOK with full repo context — aider --model deepseek/deepseek-coder in tmux, commits each change.
• Cline (free, VS Code) if you want VS Code extension ecosystem + BYOK + terminal-native agent — slower UI than Continue.dev but richer integrations.
• Windsurf ($15/mo) if you want persistent context + Cascade agents + full VS Code extension host — best of both worlds but paid.

Tool Crucible Rating

Dimension	Rating (1–5)	Notes
Overall	4	Best cost/control for long sessions; LSP gap is real blocker for TypeScript teams
Ease of Use	3	Ollama setup friction; config file editing; no GUI for model routing rules
Value	5	Free + local models = $0 marginal cost; 65% savings vs Cursor Pro team plan
Support	3	Active Discord, weekly releases, but docs trail features

---

This is part of our AI Coding Agent Evaluation series. See full comparison: Continue.dev vs Cursor vs Cline vs Aider: Local Model Routing Tested