Tooling

Claude Code as Paid Companion

When the paid gold standard is worth it, and how it pairs with a self-hosted local stack.

This is the honest "paid AI as companion to your local stack" page. Claude Code (Anthropic, 2025) is, in May 2026, widely regarded as the best agentic coding experience available — paid, closed, but excellent. Cursor is its IDE-flavour competitor. The question this page exists to answer: when does paid earn its slot, given that the rest of this section is OSS-headline?

For the FOSS local picks see Aider, Cline / Roo Code, Continue.dev. Strategic overview at agentic coding overview. For cost math see cost-cloud-vs-local.

What Claude Code is

  • Anthropic's terminal-based agentic coding tool — runs in your shell, edits files, runs commands.
  • Sonnet 4.7 / Opus 4.7 backed; long-context aware.
  • MCP-native — talks to MCP servers.
  • Subscription-based: Pro at ~$20/mo (limited usage), Max at ~$100–200/mo (unlimited).
  • Closed source.

What Cursor is

  • VS Code fork with first-class agent mode and tab completion.
  • Backs onto Claude Sonnet, GPT-5, internal Cursor models for fast tab completion.
  • ~$20/mo Pro; team / business tiers.
  • Closed source (the IDE; agent backends are upstream models).

When paid is worth it (honest)

  • Hardest tasks: novel architecture, ambiguous specs, gnarly debugging across many files. Frontier models are still genuinely better here than the best open weights.
  • Long-context whole-codebase work. 200K+ tokens of context with strong recall is something local 32K models can't match.
  • Speed. 80–100 tok/s makes agent loops tolerable in wall-clock time.
  • Reliability over many turns. Frontier APIs malform fewer tool calls.
  • No hardware investment. Skipping the Tier 2+ build is real money saved.
  • Heavy use justifies subscription tiers. $200/mo unlimited beats $0.50–2/hr API metering at heavy usage.

When local wins anyway

  • ★ ★ Privacy-sensitive code. Source code that must not leave the org.
  • ★ ★ No-budget / hobby use. Hardware amortizes; subscriptions don't.
  • ★ ★ Offline / network-limited. Local works on a plane.
  • Routine work. Qwen2.5-Coder 32B in Aider handles the boring 80% fine.
  • Learning / hobby. Building this is the hobby.

Hybrid: the practical answer

The most productive setup for many in May 2026:

  • ★ ★ Aider routing to Claude Sonnet 4.7 for hard tasks via OpenRouter or direct API; Qwen2.5-Coder 32B local for routine.
  • LiteLLM as the proxy with a budget cap.
  • Aider architect mode with Claude as architect and local Qwen as editor — premium reasoning, free execution.
  • Claude Code as the second tool — invoked when Aider's diff workflow isn't right for the task; rarely used but worth having.

Or:

  • Cursor for IDE work, Claude Code for terminal-side complex tasks, local Aider for offline / privacy-sensitive code.

Cost honest framing

  • Claude Code Max ~$200/mo unlimited Sonnet/Opus access — for heavy users, this beats per-token API; for light users, overkill.
  • Cursor Pro ~$20/mo with included usage; pay-per-use beyond.
  • OpenRouter / direct API at moderate use — $20–80/mo typical.
  • Pure local — $0/mo + electricity + amortized hardware.

For a heavy daily coder spending 4hrs/day with AI assistance, the breakeven for buying a 3090 vs. paying Claude Code Max at $200/mo is roughly 5–8 months. Hardware wins long-term if usage is steady; subscriptions win for inconsistent use or when frontier capability matters.

See cost-cloud-vs-local for the deeper math.

Honest framing

  • Claude Code is genuinely the gold standard in agentic coding in May 2026. It's worth saying outright; the rest of this section's OSS-headline posture doesn't require us to pretend otherwise.
  • It's also paid, closed, and rate-limited at the lower tiers. Those are the real tradeoffs.
  • The local stack closes 80–90% of the practical workload — for routine work, your own Aider + Qwen2.5-Coder is fine.
  • For the remaining 10–20% — the truly hard tasks — paid frontier earns its slot honestly.

Pick this if…

  • Heavy daily AI coder, no privacy constraint, budget allows: Claude Code Max.
  • Heavy IDE-flavoured user: Cursor.
  • Mix of routine + occasional hard: ★ ★ Hybrid — local Aider + Qwen routing to Claude API for hard tasks via LiteLLM.
  • Privacy / offline / hobby: stay local; Aider + Qwen2.5-Coder 32B is the answer.
  • You want one tool to rule them all: there isn't one. The honest setup is a stack.

On this page