选型 · 按 2026-08 价目表选档与分工Choose · Tiers and the Split, Priced as of 2026-08.
选档先分清你缺的是额度还是管控:Free 试水(现在带 Warp Agent CLI 与按量充值),Build 是多数个人的答案,Max 是重度并发的容量档,Business 管 SSO 与团队可见,Enterprise 才是 BYOLLM 与自托管。再划清 Warp、Cursor、CLI agent 与普通终端的分工。数字截至 2026-08-06。To pick a tier, first sort out whether you're short on quota or on controls: Free to test (now with the Warp Agent CLI and pay-as-you-go reloads), Build is most individuals' answer, Max is the capacity tier for heavy concurrency, Business is SSO and team visibility, Enterprise is BYOLLM and self-hosting. Then the split between Warp, Cursor, CLI agents, and plain terminals. Numbers as of 2026-08-06.
这一章不问「Warp 值不值」,而问三个更实际的问题:你该付哪一档,模型钱该走 Warp credits 还是自己的 key,这台终端形状的元工具又该和 Cursor、CLI agents、普通终端怎么分工。数字都写死日期 —— 本章是 2026-08-06 的价目表,因为 Warp 的计费和模型表改得很勤。This chapter does not ask whether Warp is worth it. It asks three more useful questions: which tier to pay for, whether model spend should go through Warp credits or your own key, and how this terminal-shaped meta-tool splits work with Cursor, CLI agents, and a plain terminal. Every number is dated — this chapter is the 2026-08-06 price list — because Warp changes pricing and model availability often.
— I
先看撞的是额度墙,还是管控墙First: Quota Wall or Control Wall.
截至 2026-08-06,Warp pricing 页显示 Free / Build / Max / Business / Enterprise 五档。1注 1Note 1Warp · Pricing —— 截至 2026-08-06(逐字重取):Free $0,含「Bring your own AI inference」「Limited cloud agents access」「Warp Agent CLI access」「Reload credits at pay-as-you-go rates」「Use any harness in the cloud (beta)」,Drive 与云端会话存储有限;Build 月付 $20 / 年付折算 $18,含 1,500 credits(按 API 价格折 $20 的 agent 用量)、frontier + 开源模型、加购(volume discounts + auto-reload + team spend cap)、扩展 cloud access、最高 codebase indexing、无限 Drive objects;Max 月付 $200 / 年付折算 $180,18,000 credits(Build 的 12 倍);Business 月付 $50 / 年付折算 $45,pricing 页写至多 25 座席(pricing FAQ 文档页仍写 50,本书以 pricing 页为准并标日期),1,500 credits/座、团队指标、管理员数据控制、SAML SSO;Enterprise 定制:共享额度池、BYOLLM、自托管 cloud agents、Analytics API、custom indexing、cross-harness agent memory(Research Preview)。未用完的月度 credits 不结转;Build 是 2025-10-30 发布、2025-12-01 起替换旧 Pro / Turbo / Lightspeed 的档。Warp · Pricing — as of 2026-08-06 (verbatim re-fetch): Free $0 with 'Bring your own AI inference,' 'Limited cloud agents access,' 'Warp Agent CLI access,' 'Reload credits at pay-as-you-go rates,' 'Use any harness in the cloud (beta),' and limited Drive / cloud conversation storage; Build $20 monthly / $18 annualized with 1,500 credits ($20 of included agent usage at API rates), frontier + open-source models, reloads (volume discounts + auto-reload + team spend cap), extended cloud access, highest codebase indexing, unlimited Drive objects; Max $200 monthly / $180 annualized with 18,000 credits (12× Build); Business $50 monthly / $45 annualized — the pricing page says up to 25 seats (the pricing-FAQ docs page still says 50; this book follows the pricing page, dated) — 1,500 credits/seat, team metrics, admin data controls, SAML SSO; Enterprise custom: shared pools, BYOLLM, self-hosted cloud agents, Analytics API, custom indexing, cross-harness agent memory (Research Preview). Unused monthly credits do not roll over; Build replaced the legacy Pro / Turbo / Lightspeed tiers (announced 2025-10-30, effective 2025-12-01).不要按「越贵越强」理解它们,按「我现在缺什么」来读:Free 和 Build 答额度,Max 答容量,Business 答管控,Enterprise 答采购。As of 2026-08-06, Warp's pricing page shows five tiers: Free / Build / Max / Business / Enterprise.1注 1Note 1Warp · Pricing —— 截至 2026-08-06(逐字重取):Free $0,含「Bring your own AI inference」「Limited cloud agents access」「Warp Agent CLI access」「Reload credits at pay-as-you-go rates」「Use any harness in the cloud (beta)」,Drive 与云端会话存储有限;Build 月付 $20 / 年付折算 $18,含 1,500 credits(按 API 价格折 $20 的 agent 用量)、frontier + 开源模型、加购(volume discounts + auto-reload + team spend cap)、扩展 cloud access、最高 codebase indexing、无限 Drive objects;Max 月付 $200 / 年付折算 $180,18,000 credits(Build 的 12 倍);Business 月付 $50 / 年付折算 $45,pricing 页写至多 25 座席(pricing FAQ 文档页仍写 50,本书以 pricing 页为准并标日期),1,500 credits/座、团队指标、管理员数据控制、SAML SSO;Enterprise 定制:共享额度池、BYOLLM、自托管 cloud agents、Analytics API、custom indexing、cross-harness agent memory(Research Preview)。未用完的月度 credits 不结转;Build 是 2025-10-30 发布、2025-12-01 起替换旧 Pro / Turbo / Lightspeed 的档。Warp · Pricing — as of 2026-08-06 (verbatim re-fetch): Free $0 with 'Bring your own AI inference,' 'Limited cloud agents access,' 'Warp Agent CLI access,' 'Reload credits at pay-as-you-go rates,' 'Use any harness in the cloud (beta),' and limited Drive / cloud conversation storage; Build $20 monthly / $18 annualized with 1,500 credits ($20 of included agent usage at API rates), frontier + open-source models, reloads (volume discounts + auto-reload + team spend cap), extended cloud access, highest codebase indexing, unlimited Drive objects; Max $200 monthly / $180 annualized with 18,000 credits (12× Build); Business $50 monthly / $45 annualized — the pricing page says up to 25 seats (the pricing-FAQ docs page still says 50; this book follows the pricing page, dated) — 1,500 credits/seat, team metrics, admin data controls, SAML SSO; Enterprise custom: shared pools, BYOLLM, self-hosted cloud agents, Analytics API, custom indexing, cross-harness agent memory (Research Preview). Unused monthly credits do not roll over; Build replaced the legacy Pro / Turbo / Lightspeed tiers (announced 2025-10-30, effective 2025-12-01). Don't read them as more-expensive-is-stronger; read them as answers to what you're missing: Free and Build answer quota, Max answers capacity, Business answers controls, Enterprise answers procurement.
档位Tier
截至 2026-08-06 的官方价目Official price as of 2026-08-06
适合谁Best for
别误会成什么Do not mistake it for
FreeFree
$0;无自带 agent 用量 —— 接 BYOK / custom endpoint / SuperGrok,或按 PAYG 价格充值 credits;含 Warp Agent CLI、有限 cloud access、cloud harness(beta)。$0; no bundled agent usage — connect BYOK / a custom endpoint / SuperGrok, or reload credits at pay-as-you-go rates; includes the Warp Agent CLI, limited cloud access, cloud harnesses (beta).
先验证 Warp 是否真能替你省终端工作 —— 手里有 provider key 的人最顺;只有 SSH 场景的人,光 Agent CLI 就值回安装。Testing whether Warp genuinely saves terminal work — smoothest with a provider key in hand; for SSH-only scenarios, the Agent CLI alone repays the install.
长期重度 agent 档,或「免费送模型额度」。A long-term heavy-agent plan, or free bundled model quota.
BuildBuild
月付 $20 / 年付折算 $18;1,500 credits(约 $20 的 API 价 agent 用量);frontier + 开源模型;加购有量折扣、auto-reload、团队 spend cap;无限 Drive objects。$20 monthly / $18 annualized; 1,500 credits (about $20 of agent usage at API rates); frontier + open-source models; reloads with volume discounts, auto-reload, team spend cap; unlimited Drive objects.
多数个人用户。你缺的是稳定额度,不是组织治理。Most individual users. You need steady quota, not organizational governance.
AI 开发环境已进入采购、合规、平台工程的公司。Companies where the AI development environment is now a procurement, compliance, and platform-engineering concern.
一个人想省钱的捷径。A shortcut for a solo user trying to save money.
— II
模型名单与三条账本The Model Roster and Three Ledgers.
截至 2026-08-06,Model Choice 的名单:OpenAI 到 GPT-5.6(Sol / Terra / Luna)与 Codex 系;Anthropic 到 Claude Fable 5、Sonnet 5、Opus 4.8;Google 到 Gemini 3.1 Pro;xAI 到 Grok 4.5;Fireworks 托管的开源模型到 GLM 5.2、Kimi K2.7 Code、DeepSeek V4 Pro 等。不想挑就用 Auto 四档(Responsive / Cost-efficient / Genius / Open-weights)。两个心眼:pricing 页的营销名单跑在文档前面(Claude Opus 5、Kimi K3 之类),以文档名单为准;Fable 5 要求 Anthropic 数据保留、无 ZDR,在意数据策略的先看这条。3注 3Note 3Warp Docs · Agent model choice —— 截至 2026-08-06 的名单:OpenAI GPT-5.6 Sol / Terra / Luna、GPT-5.5 / 5.4、GPT-5.3 / 5.2 Codex、GPT-5.2;Anthropic Claude Fable 5、Sonnet 5、Opus 4.8 / 4.7 / 4.6 / 4.5、Sonnet 4.6 / 4.5、Haiku 4.5;Google Gemini 3.1 Pro、Gemini 3.5 Flash;xAI Grok 4.5 / 4.3、Grok Build 0.1;Fireworks-hosted GLM 5.2、Kimi K2.7 Code、Kimi K2.6、Minimax 3 / 2.7、Qwen 3.7 / 3.6 Plus、DeepSeek V4 Pro。Auto 四档:Responsive / Cost-efficient / Genius / Open-weights。Fable 5 要求 Anthropic 数据保留(无 ZDR,Enterprise 默认关)。注意:pricing 页的营销名单(Claude Opus 5、Kimi K3 等)跑在文档名单前面 —— 以文档名单为准,标日期。Warp Docs · Agent model choice — the roster as of 2026-08-06: OpenAI GPT-5.6 Sol / Terra / Luna, GPT-5.5 / 5.4, GPT-5.3 / 5.2 Codex, GPT-5.2; Anthropic Claude Fable 5, Sonnet 5, Opus 4.8 / 4.7 / 4.6 / 4.5, Sonnet 4.6 / 4.5, Haiku 4.5; Google Gemini 3.1 Pro, Gemini 3.5 Flash; xAI Grok 4.5 / 4.3, Grok Build 0.1; Fireworks-hosted GLM 5.2, Kimi K2.7 Code, Kimi K2.6, Minimax 3 / 2.7, Qwen 3.7 / 3.6 Plus, DeepSeek V4 Pro. Four Auto modes: Responsive / Cost-efficient / Genius / Open-weights. Fable 5 requires Anthropic data retention (no ZDR; Enterprise default off). Note: the pricing page's marketing roster (Claude Opus 5, Kimi K3, etc.) runs ahead of the docs list — the docs list is the dated source of truth.As of 2026-08-06, the Model Choice roster: OpenAI up to GPT-5.6 (Sol / Terra / Luna) and the Codex line; Anthropic up to Claude Fable 5, Sonnet 5, Opus 4.8; Google up to Gemini 3.1 Pro; xAI up to Grok 4.5; Fireworks-hosted open-weights up to GLM 5.2, Kimi K2.7 Code, DeepSeek V4 Pro, and more. If you don't want to pick, use the four Auto modes (Responsive / Cost-efficient / Genius / Open-weights). Two cautions: the pricing page's marketing roster runs ahead of the docs (Claude Opus 5, Kimi K3, and friends) — the docs list is the source of truth; and Fable 5 requires Anthropic data retention with no ZDR — check that first if data policy matters to you.3注 3Note 3Warp Docs · Agent model choice —— 截至 2026-08-06 的名单:OpenAI GPT-5.6 Sol / Terra / Luna、GPT-5.5 / 5.4、GPT-5.3 / 5.2 Codex、GPT-5.2;Anthropic Claude Fable 5、Sonnet 5、Opus 4.8 / 4.7 / 4.6 / 4.5、Sonnet 4.6 / 4.5、Haiku 4.5;Google Gemini 3.1 Pro、Gemini 3.5 Flash;xAI Grok 4.5 / 4.3、Grok Build 0.1;Fireworks-hosted GLM 5.2、Kimi K2.7 Code、Kimi K2.6、Minimax 3 / 2.7、Qwen 3.7 / 3.6 Plus、DeepSeek V4 Pro。Auto 四档:Responsive / Cost-efficient / Genius / Open-weights。Fable 5 要求 Anthropic 数据保留(无 ZDR,Enterprise 默认关)。注意:pricing 页的营销名单(Claude Opus 5、Kimi K3 等)跑在文档名单前面 —— 以文档名单为准,标日期。Warp Docs · Agent model choice — the roster as of 2026-08-06: OpenAI GPT-5.6 Sol / Terra / Luna, GPT-5.5 / 5.4, GPT-5.3 / 5.2 Codex, GPT-5.2; Anthropic Claude Fable 5, Sonnet 5, Opus 4.8 / 4.7 / 4.6 / 4.5, Sonnet 4.6 / 4.5, Haiku 4.5; Google Gemini 3.1 Pro, Gemini 3.5 Flash; xAI Grok 4.5 / 4.3, Grok Build 0.1; Fireworks-hosted GLM 5.2, Kimi K2.7 Code, Kimi K2.6, Minimax 3 / 2.7, Qwen 3.7 / 3.6 Plus, DeepSeek V4 Pro. Four Auto modes: Responsive / Cost-efficient / Genius / Open-weights. Fable 5 requires Anthropic data retention (no ZDR; Enterprise default off). Note: the pricing page's marketing roster (Claude Opus 5, Kimi K3, etc.) runs ahead of the docs list — the docs list is the dated source of truth.同一件 agent 任务,钱可以走三条路:用 Warp 自带 credits,用你的模型商 key,或走一个 OpenAI-compatible endpoint。别把它们混在一起。The same agent task can put spend on three ledgers: Warp's bundled credits, your provider key, or an OpenAI-compatible endpoint. Do not blur them together.
账本Ledger
省什么What it saves
硬边界Hard boundary
Warp creditsWarp credits
省配置:Auto 模型、custom routers、fallback、cloud/local agents 都走同一个账。Configuration: Auto models, custom routers, fallback, and cloud/local agents share one account.
好模型更费:文档给的相对胃口 Opus 4.7 > Sonnet 4.6 > GPT-5.5 > Gemini 3.1 Pro;月度配额不结转。Stronger models cost more: the docs' relative appetite is Opus 4.7 > Sonnet 4.6 > GPT-5.5 > Gemini 3.1 Pro; monthly quota doesn't roll over.
BYOKBYOK
省 Warp AI credits:OpenAI / Anthropic / Google 由模型商直接计费,key 存本机,选中时优先于 credits。Warp AI credits: OpenAI / Anthropic / Google bill through your provider account, keys stay local, a selected BYOK model takes priority.
不适用于 Cloud Agents;失败默认不 fallback(opt-in);ChatGPT / Claude 消费级订阅接不进;Warp 不能替你的模型商 enforce ZDR。No Cloud Agents; no default fallback on failure (opt-in); consumer ChatGPT / Claude subscriptions can't connect; Warp can't enforce provider-side ZDR for your key.
Custom endpointCustom endpoint
省供应商锁定:OpenRouter、LiteLLM、z.ai、公司内网关,三种 schema 可选(Chat Completions / Responses / Anthropic Messages)。Provider lock-in: OpenRouter, LiteLLM, z.ai, a company gateway, with three selectable schemas (Chat Completions / Responses / Anthropic Messages).
必须公网 HTTPS;localhost、私网、VPN-only 不行;不适用于 Cloud Agents;router 用不上 endpoint。Must be public HTTPS; localhost, private networks, VPN-only are rejected; no Cloud Agents; routers can't use endpoints.
BYOLLMBYOLLM
省企业治理碎片:托管推理(当前 AWS Bedrock,Azure Foundry / Google Vertex 在路上)、编排、治理、观测一体接入。Enterprise governance fragmentation: managed inference (AWS Bedrock today, Azure Foundry / Google Vertex coming), orchestration, governance, observability in one setup.
Enterprise only,不是个人用户的本地模型开关。Enterprise only, not a personal local-model toggle.
最容易误判的还是 Cloud Agents:你自己配的 BYOK 与 custom endpoint 都是本地配置,云端拿不到,cloud run 照常消耗 Warp credits;用 Claude Code / Codex 当云端 harness 时,推理记你的 provider 账,但 compute 与 platform credits 照计。4注 4Note 4Warp Docs · Credits + BYOK —— 截至 2026-08-06:credits 分 AI / compute / platform 三桶同一池扣;相对胃口(文档标日期排序)Opus 4.7 > Sonnet 4.6 > GPT-5.5 > Gemini 3.1 Pro,单次 run 不确定;Free 无自带 agent 用量,需 BYOK / custom endpoint / SuperGrok 或升级;BYOK 选中时优先于 credits、失败默认不 fallback(opt-in);消费级 ChatGPT / Claude 订阅接不进;cloud runs 不支持用户级 BYOK。Warp Docs · Credits + BYOK — as of 2026-08-06: credits split into AI / compute / platform buckets drawn from one pool; relative appetite (the docs' dated ordering) Opus 4.7 > Sonnet 4.6 > GPT-5.5 > Gemini 3.1 Pro, non-deterministic per run; Free has no bundled agent usage — bring BYOK / a custom endpoint / SuperGrok, or upgrade; a selected BYOK model takes priority over credits with no default fallback (opt-in); consumer ChatGPT / Claude subscriptions can't connect; cloud runs don't support user-level BYOK.6注 6Note 6Warp Docs · Warp vs Claude Code + Harnesses in Oz —— 官方对比:Claude Code 是 CLI tool,Warp agent 内建于 Warp terminal,提供可视化 diff、Profiles、Rules、多模型、context referencing;Oz 可把 Claude Code / Codex 当云端 harness 跑,推理记你的 provider 账,Warp 计 compute 与 platform credits,pricing 页标 beta。Warp Docs · Warp vs Claude Code + Harnesses in Oz — the official comparison: Claude Code is a CLI tool, while Warp's agent is built into the Warp terminal with visual diffs, Profiles, Rules, multiple models, and context referencing; Oz can run Claude Code / Codex as cloud harnesses with inference on your provider account while Warp meters compute and platform credits, labeled beta on the pricing page.「Enterprise team-managed keys」那条旧例外已经取消 —— 集中托管推理的路是 BYOLLM,别按旧材料配。The easiest misread is still Cloud Agents: the BYOK and custom endpoint you configured are local, the cloud can't reach them, and cloud runs consume Warp credits as usual; when Claude Code / Codex runs as a cloud harness, inference bills your provider but compute and platform credits are still metered.4注 4Note 4Warp Docs · Credits + BYOK —— 截至 2026-08-06:credits 分 AI / compute / platform 三桶同一池扣;相对胃口(文档标日期排序)Opus 4.7 > Sonnet 4.6 > GPT-5.5 > Gemini 3.1 Pro,单次 run 不确定;Free 无自带 agent 用量,需 BYOK / custom endpoint / SuperGrok 或升级;BYOK 选中时优先于 credits、失败默认不 fallback(opt-in);消费级 ChatGPT / Claude 订阅接不进;cloud runs 不支持用户级 BYOK。Warp Docs · Credits + BYOK — as of 2026-08-06: credits split into AI / compute / platform buckets drawn from one pool; relative appetite (the docs' dated ordering) Opus 4.7 > Sonnet 4.6 > GPT-5.5 > Gemini 3.1 Pro, non-deterministic per run; Free has no bundled agent usage — bring BYOK / a custom endpoint / SuperGrok, or upgrade; a selected BYOK model takes priority over credits with no default fallback (opt-in); consumer ChatGPT / Claude subscriptions can't connect; cloud runs don't support user-level BYOK.6注 6Note 6Warp Docs · Warp vs Claude Code + Harnesses in Oz —— 官方对比:Claude Code 是 CLI tool,Warp agent 内建于 Warp terminal,提供可视化 diff、Profiles、Rules、多模型、context referencing;Oz 可把 Claude Code / Codex 当云端 harness 跑,推理记你的 provider 账,Warp 计 compute 与 platform credits,pricing 页标 beta。Warp Docs · Warp vs Claude Code + Harnesses in Oz — the official comparison: Claude Code is a CLI tool, while Warp's agent is built into the Warp terminal with visual diffs, Profiles, Rules, multiple models, and context referencing; Oz can run Claude Code / Codex as cloud harnesses with inference on your provider account while Warp meters compute and platform credits, labeled beta on the pricing page. The old 'Enterprise team-managed keys' exception is gone — centrally managed inference goes through BYOLLM, so don't configure against old material.
提示词PromptWarp tier self-audit
请根据我过去 30 天的真实使用,给我一个 Warp 选档建议,不要按「越贵越好」推荐。
我的任务:
- 每周大概运行 agent:<次数>
- 其中多步改代码 / 排障:<比例>
- 其中文件批处理 / 数据清理 / 运维:<比例>
- 是否需要 Cloud Agents / Slack / GitHub 触发:<是/否>
- 是否需要 SSH / 远程服务器上的 agent:<是/否>
- 是否需要 SSO、管理员数据控制、团队 usage metrics:<是/否>
- 是否已有 OpenAI / Anthropic / Google key 或 OpenAI-compatible gateway:<有/无>
请输出:
1. 推荐档位:Free / Build / Max / Business / Enterprise
2. 推荐账本:Warp credits / BYOK / custom endpoint / 混合
3. 升档前的 7 天验证动作
4. 哪些任务不该放进 WarpBased on my real usage in the past 30 days, recommend a Warp tier. Do not recommend by "more expensive is better."
My tasks:
- Agent runs per week: <count>
- Multi-step coding / debugging share: <share>
- File batch / data cleanup / ops share: <share>
- Need Cloud Agents / Slack / GitHub triggers: <yes/no>
- Need an agent on SSH / remote servers: <yes/no>
- Need SSO, admin data controls, team usage metrics: <yes/no>
- Already have OpenAI / Anthropic / Google keys or an OpenAI-compatible gateway: <yes/no>
Please output:
1. Recommended tier: Free / Build / Max / Business / Enterprise
2. Recommended ledger: Warp credits / BYOK / custom endpoint / hybrid
3. A 7-day validation step before upgrading
4. Which tasks should not go into Warp
请根据我过去 30 天的真实使用,给我一个 Warp 选档建议,不要按「越贵越好」推荐。
我的任务:
- 每周大概运行 agent:约 40 次
- 其中多步改代码 / 排障:60%
- 其中文件批处理 / 数据清理 / 运维:30%
- 是否需要 Cloud Agents / Slack / GitHub 触发:否
- 是否需要 SSH / 远程服务器上的 agent:是,两台只读排查
- 是否需要 SSO、管理员数据控制、团队 usage metrics:否
- 是否已有 OpenAI / Anthropic / Google key 或 OpenAI-compatible gateway:有 Anthropic key
请输出:
1. 推荐档位:Free / Build / Max / Business / Enterprise
2. 推荐账本:Warp credits / BYOK / custom endpoint / 混合
3. 升档前的 7 天验证动作
4. 哪些任务不该放进 WarpBased on my real usage in the past 30 days, recommend a Warp tier. Do not recommend by "more expensive is better."
My tasks:
- Agent runs per week: about 40
- Multi-step coding / debugging share: 60%
- File batch / data cleanup / ops share: 30%
- Need Cloud Agents / Slack / GitHub triggers: no
- Need an agent on SSH / remote servers: yes, two read-only boxes
- Need SSO, admin data controls, team usage metrics: no
- Already have OpenAI / Anthropic / Google keys or an OpenAI-compatible gateway: yes, an Anthropic key
Please output:
1. Recommended tier: Free / Build / Max / Business / Enterprise
2. Recommended ledger: Warp credits / BYOK / custom endpoint / hybrid
3. A 7-day validation step before upgrading
4. Which tasks should not go into Warp
Warp 官方对比 Claude Code 时,核心差别不是「谁更聪明」,而是表面:Claude Code 是 CLI tool;Warp 的 agent 内建在终端里,有可视化 diff、Profiles、Rules、多模型与 context referencing。6注 6Note 6Warp Docs · Warp vs Claude Code + Harnesses in Oz —— 官方对比:Claude Code 是 CLI tool,Warp agent 内建于 Warp terminal,提供可视化 diff、Profiles、Rules、多模型、context referencing;Oz 可把 Claude Code / Codex 当云端 harness 跑,推理记你的 provider 账,Warp 计 compute 与 platform credits,pricing 页标 beta。Warp Docs · Warp vs Claude Code + Harnesses in Oz — the official comparison: Claude Code is a CLI tool, while Warp's agent is built into the Warp terminal with visual diffs, Profiles, Rules, multiple models, and context referencing; Oz can run Claude Code / Codex as cloud harnesses with inference on your provider account while Warp meters compute and platform credits, labeled beta on the pricing page.所以别问谁赢,问这件事在哪个表面上最自然。In Warp's official comparison with Claude Code, the key difference is not 'which is smarter' but the surface: Claude Code is a CLI tool; Warp's agent is built into the terminal, with visual diffs, Profiles, Rules, multiple models, and context referencing.6注 6Note 6Warp Docs · Warp vs Claude Code + Harnesses in Oz —— 官方对比:Claude Code 是 CLI tool,Warp agent 内建于 Warp terminal,提供可视化 diff、Profiles、Rules、多模型、context referencing;Oz 可把 Claude Code / Codex 当云端 harness 跑,推理记你的 provider 账,Warp 计 compute 与 platform credits,pricing 页标 beta。Warp Docs · Warp vs Claude Code + Harnesses in Oz — the official comparison: Claude Code is a CLI tool, while Warp's agent is built into the Warp terminal with visual diffs, Profiles, Rules, multiple models, and context referencing; Oz can run Claude Code / Codex as cloud harnesses with inference on your provider account while Warp meters compute and platform credits, labeled beta on the pricing page. So don't ask which tool wins — ask which surface this task naturally belongs on.截至 2026-08-06,这道分工题有两个条件变了。其一,「用 Warp」和「用你已付费的 CLI agent」不互斥:Warp 自动识别跑在它里面的第三方 CLI agent —— 官方名单 10 个,Claude Code、Codex、OpenCode、Amp、Auggie、Copilot CLI、Cursor CLI、Gemini CLI、Droid、Pi —— 叠上 rich input、code review、上下文引用等 UI 层(注意:agent 通知只有 Claude Code / Codex / OpenCode 三家支持)。其二,Warp 自己也出了 CLI:warp 二进制把 Warp Agent 带进任何终端含 SSH,同账户同 harness,Free 档就能用 —— 「远程场景回普通终端」的旧建议已被它改写。5注 5Note 5Warp Docs · Third-party CLI agents + Warp Agent CLI —— 截至 2026-08-06:universal agent support 官方名单 10 个 —— Claude Code、Codex、OpenCode、Amp、Auggie、Copilot CLI、Cursor CLI、Gemini CLI、Droid、Pi;agent 通知只有 Claude Code / Codex / OpenCode 支持(需插件/配置);叠加能力:rich input(Ctrl-G)、code review、attach code as context、vertical tabs、Tab Configs、Remote Control。Warp 自己的 Warp Agent CLI(warp 二进制)把 Warp Agent 带进任何终端含 SSH,同账户同 harness —— Free 档就可用。Warp Docs · Third-party CLI agents + Warp Agent CLI — as of 2026-08-06: the official universal-agent-support list has 10 — Claude Code, Codex, OpenCode, Amp, Auggie, Copilot CLI, Cursor CLI, Gemini CLI, Droid, Pi; agent notifications only work for Claude Code / Codex / OpenCode (plugin/config required); layered capabilities: rich input (Ctrl-G), code review, attach code as context, vertical tabs, Tab Configs, Remote Control. Warp's own Warp Agent CLI (the warp binary) brings the Warp Agent to any terminal including SSH, same account and harness — available on the Free tier.但别读成「Warp 罩住一切」:这是 Warp 想成为 agent 之上那层的产品策略,你的 CLI agent 在普通终端里跑一样成立;值不值得为这层驾驶舱多养一个 GUI 终端,仍按下面这张表判断。As of 2026-08-06, two conditions of this split changed. First, 'use Warp' and 'use the CLI agent you already pay for' aren't mutually exclusive: Warp auto-detects third-party CLI agents running inside it — the official list is 10: Claude Code, Codex, OpenCode, Amp, Auggie, Copilot CLI, Cursor CLI, Gemini CLI, Droid, Pi — layering rich input, code review, and context referencing on top (note: agent notifications only work for Claude Code / Codex / OpenCode). Second, Warp shipped its own CLI: the warp binary brings the Warp Agent to any terminal including SSH, same account and harness, available on Free — the old 'remote means a plain terminal' advice has been rewritten by it.5注 5Note 5Warp Docs · Third-party CLI agents + Warp Agent CLI —— 截至 2026-08-06:universal agent support 官方名单 10 个 —— Claude Code、Codex、OpenCode、Amp、Auggie、Copilot CLI、Cursor CLI、Gemini CLI、Droid、Pi;agent 通知只有 Claude Code / Codex / OpenCode 支持(需插件/配置);叠加能力:rich input(Ctrl-G)、code review、attach code as context、vertical tabs、Tab Configs、Remote Control。Warp 自己的 Warp Agent CLI(warp 二进制)把 Warp Agent 带进任何终端含 SSH,同账户同 harness —— Free 档就可用。Warp Docs · Third-party CLI agents + Warp Agent CLI — as of 2026-08-06: the official universal-agent-support list has 10 — Claude Code, Codex, OpenCode, Amp, Auggie, Copilot CLI, Cursor CLI, Gemini CLI, Droid, Pi; agent notifications only work for Claude Code / Codex / OpenCode (plugin/config required); layered capabilities: rich input (Ctrl-G), code review, attach code as context, vertical tabs, Tab Configs, Remote Control. Warp's own Warp Agent CLI (the warp binary) brings the Warp Agent to any terminal including SSH, same account and harness — available on the Free tier. But don't read this as 'Warp wraps everything': that's Warp's play to become the layer above other agents, your CLI agent runs fine in a plain terminal, and whether the cockpit justifies one more GUI terminal is still decided by the table below.
浏览器自动化 / 专门 Computer Use;Warp 的 cloud sandbox 只用在适配场景。Browser automation / dedicated Computer Use; Warp's cloud sandbox only when the fit is right.
Warp 的 Computer Use 是云端沙箱能力,不是接管你本机桌面。Warp's Computer Use is a cloud-sandbox capability, not control of your local desktop.
— IV
最后用五个问题做决定Make the Decision with Five Questions.
我每周是否至少有 3 件终端侧的手动活,能被 agent 接手或被 Drive 沉淀?没有,先 Free。Do I have at least 3 terminal-side manual tasks per week that an agent can take over or Drive can distill? If not, stay Free.
我撞到的是 credits 不够,还是组织控制不够?前者看 Build / Max / 加购 / BYOK / custom endpoint;后者看 Business / Enterprise。Am I hitting a credits limit or an organizational-control limit? The former points to Build / Max / reloads / BYOK / custom endpoint; the latter to Business / Enterprise.
月度配额烧完前,我是否先算了加购的单价(最低约 $0.015/credit)再决定升档?Before upgrading, did I price reloads first (as low as about $0.015/credit)?
我有没有把 SSH 场景错判成「只能普通终端」?先试试 Warp Agent CLI,它在 Free 档就有。Did I misjudge SSH as plain-terminal-only? Try the Warp Agent CLI first — it's on the Free tier.
我准备上 Cloud Agent 之前,是否确认了触发、上下文、验收、额度、日志与人工回滚?没有,就先留本地。Before sending work to a Cloud Agent, have I defined trigger, context, validation, credits, logs, and human rollback? If not, keep it local.
选档看缺口,
分工看表面.
Pick the tier by the gap,
split the work by the surface.
Aklman Library
— 讨论Discussion
讨论Discussion.
评论区初始化中…Initializing comments…
01 / 01
没有匹配结果No matches.
换个关键词，或按 Esc 回到页面Try another keyword, or press Esc to return
