무제한 Claude로 Factory Droid 실행하기

Droid는 계획·수정·재시도를 루프로 수행하는 에이전트 CLI라, 하나의 작업이 끝나기까지 많은 토큰을 움직일 수 있습니다. 바로 그 작업에서 정액 무제한 Claude 요금제가 이깁니다 — Droid를 끝까지 실행해도 가격은 변하지 않습니다.

One config, no meter

Droid reads ~/.factory/settings.json. Add a customModels array and each entry defines one model the CLI can use. With an unlimited plan behind it, you set this once and stop thinking about per-run cost.

settings.json is the modern path; the legacy config.json lacks some of its features. Keep your unlimited setup here so every Droid session draws from the same flat-rate plan.

Because Droid is built to run long autonomous sessions, the one-time config and the flat-rate plan complement each other well. You configure it once, then leave the cost question behind entirely and let the agent work as long as the task demands. Configure once, forget the meter, and let the agent run is the whole shape of this setup. The long autonomous sessions Droid favours fit a flat rate perfectly, since their cost is otherwise the hardest to predict.

Describe the Claude model

Each entry takes model, baseUrl, apiKey, provider, and maxOutputTokens. Point model at a gateway id like claude-sonnet-4-5 and baseUrl at the unlimited gateway host.

Set maxOutputTokens generously. On metered billing people cap output low to save money; on unlimited you can let Droid produce full, complete responses because the cost is the same either way.

A higher maxOutputTokens matters for an agent that writes code, since a clipped response can cut off mid-edit and force another round to finish. With the cost concern removed you size it for complete outputs, which keeps the agent loop smoother and reduces wasted turns. Complete outputs mean fewer wasted turns, which matters more the longer a session runs. A response cut off mid-edit forces another round you would rather avoid, so a generous limit keeps the loop moving.

Use the anthropic provider

Set provider to anthropic for the gateway's native Messages surface. The other modes, openai and generic-chat-completion-api, target OpenAI-style endpoints; for Claude on the unlimited gateway, anthropic is the match.

Getting the provider right keeps the agent loop solid through long runs. You want it correct so an hours-long unlimited session never trips on a wire-format mismatch.

This is more important on unlimited precisely because you intend the sessions to be long. A provider mismatch that you might catch quickly on a short metered test could otherwise surface deep into an extended run, so it is worth confirming on a small task before turning the agent loose. A small confirmation run up front is cheap insurance against a mismatch surfacing hours in. A quick sanity check costs nothing on an unlimited plan, and it spares you a failed long run.

Key handling

Put your unlimited key in apiKey, ideally as ${AIPRIME_API_KEY} so the secret lives in the environment. ${VAR} expansion works in settings.json but not the legacy config.json, and apiKey must not be empty.

That one unlimited key backs every model entry and every Droid session for the plan's period. There is no separate meter per model to watch.

So if you keep both a Sonnet and an Opus entry, they share the single unlimited key and the single flat rate between them. Switching models never changes what you pay, which makes the choice purely about which model suits the task in front of you. Sharing one key and one rate across model entries makes switching purely a quality decision. Cost never enters the choice between Sonnet and Opus, so you simply pick the model the task deserves.

Let the agent loop run

In the CLI, switch models with /model. Keep a claude-opus-4-8 entry for the hardest tasks and, because the plan is unlimited, leave it selected without worrying about token cost.

This is where flat-rate really pays off: an agentic CLI that retries and explores would be unpredictable on metered pricing. On unlimited you point Droid at a hard task, let it iterate as long as it needs, and pay the same fixed amount.

It also frees you from the instinct to interrupt a long run to save money, which often produces worse results than letting the agent finish its plan. When a five-minute run and an hour-long run cost the same, you make that call on whether the work is done, not on what the next stretch of tokens would add. Judging a run on whether it is done, not on token cost, is the mindset a flat rate unlocks. You stop interrupting good work to save money.

// ~/.factory/settings.json
{
  "customModels": [
    {
      "model": "claude-sonnet-4-5",
      "baseUrl": "https://claudeapikey.dev",
      "apiKey": "${AIPRIME_API_KEY}",
      "provider": "anthropic",
      "maxOutputTokens": 8192
    }
  ]
}
// export AIPRIME_API_KEY=<key>   (apiKey must NOT be empty)
// Switch models in the CLI with /model

FAQ

무제한 Claude로 Factory Droid를 실행할 수 있나요?
네 — provider를 anthropic, baseUrl을 {SITE}, 무제한 키를 넣은 customModels 항목을 추가하세요. 정액 요금제가 모든 실행을 커버합니다.

왜 무제한이 Droid에 좋은가요?
Droid의 에이전트 루프는 계획·수정·재시도를 하며 작업마다 높은 토큰 양을 움직입니다. 정액 무제한은 실행이 아무리 길어도 비용을 고정으로 유지합니다.

provider 값은 무엇인가요?
네이티브 Messages 표면에는 anthropic입니다. apiKey는 비어 있으면 안 되며, settings.json에서는 ${AIPRIME_API_KEY}를 사용하세요.

토큰 한도가 있나요?
요금제 기간 동안 토큰 한도는 없으며, 게이트웨이를 빠르게 유지하기 위한 공정 사용 속도 제한만 있습니다.