← Coding

AmpAmp

Amp Frontier Corporation（原 Sourcegraph，2025-12 分拆独立）Amp Frontier Corporation (spun out of Sourcegraph, Dec 2025)

编程Coding 终端Terminal 云端Cloud 混合Hybrid 多模型Multiple models partial lifecycle: active 核验Checked 2026-10-23

Amp 原是 Sourcegraph 出的前沿编码 agent，2025-12 官方公告分拆为独立公司 Amp Frontier Corporation （Sourcegraph 官方博客与 ampcode.com 公告同日互证）；CLI + 云端 orb + runner 三执行面。

发一句提示就能合上笔记本继续跑。

它多模型前沿路由，Oracle 提供「第二意见」模型。

Amp is Sourcegraph's frontier coding agent, spanning CLI + cloud orbs + runners. Send a prompt and you can close your laptop while it keeps working. It routes across frontier models, with the Oracle as a second-opinion model.

适合与不适合When to use it, when not to

想让 agent 在云端机器（orb）里长跑、合上笔记本仍继续的人
需要多模型前沿路由与 Oracle 第二意见的团队
已在终端或 VS Code 系编辑器里工作、想顺带接 IDE 的人
需要企业级管控（SSO、审计日志、MCP allowlist、零数据保留）的组织

不适合假定「执行动作前一定会先问」的场景。

官方原文明确：

官方原文："By default, Amp does not ask for approval before running tools"

审批要靠自定义 policy plugin 自己加。

Four kinds of use fit:

People who want an agent that keeps working in a cloud orb after you close your laptop
Teams that need multi-model frontier routing plus the Oracle second-opinion model
People already working in a terminal or a VS Code-family editor who want IDE integration
Organisations needing enterprise controls (SSO, audit logs, MCP allowlists, zero data retention)

Not a fit for anyone who assumes tools are gated by approval — officially, "By default, Amp does not ask for approval before running tools". Approval takes a custom policy plugin that you write yourself.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-20

出品属地Origin: 国外出品foreign-made　国外主体（网信办公示名单未见对应备案主体）。foreign entity (no matching filing entity in the CAC registry).
国内使用Direct access: 大陆可直连direct OK　官网直连实测 200。official site measured 200 direct.
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-20 检索 13 批附件，https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm）；该产品未面向大陆公众提供服务，备案不适用。not in the CAC public registry (13 batch attachments searched 2026-10-20, https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm); the product does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 多模型前沿路由，按任务分档而非锁死单模型。

官方 Why Amp 原文：

"Multi-Model: GPT-5.6, Claude Fable 5, fast models—Amp uses them all, for what each model is best at."

四个 Agent Modes 是能力预设、不是固定模型选择器： low / medium / high / ultra。

官方原文："Modes are capability presets, not fixed model selectors"

路由可以自定义。它看已连接的 provider 订阅、workspace 限制与模型可用性，来决定主 agent 与 Oracle 怎么走。

Oracle 是「第二意见」模型。

high 档主 agent 用 GPT-6 Astra（medium），Oracle 用 Claude Fable 5.1。低 / 中 / ultra 档 Oracle 默认 GPT-6 Astra。

BYOK 已全线开放（官方 news 注明 2026-09-30 起对所有档）。

支持 OpenRouter、Bedrock、Google Cloud Agent Platform、 Azure Foundry、Vercel / Cloudflare AI Gateway、Ollama Cloud、OpenCode Go 与自定义端点 URL，并可用 ChatGPT 订阅。

⚠ 官方源之间存在不一致。 安全参考页仍写：

官方原文："Amp doesn't support Bring Your Own Key or self-hosted deployments"

这与定价页/news 的 BYOK 口径相反，引用时以定价页为准。
Multi-model frontier routing — tiered per task rather than locked to a single model.

Official "Why Amp" wording:

"Multi-Model: GPT-5.6, Claude Fable 5, fast models—Amp uses them all, for what each model is best at."

The four Agent Modes are capability presets, not fixed model selectors: low / medium / high / ultra.

Official wording: "Modes are capability presets, not fixed model selectors"

Routing is customizable. It looks at connected provider subscriptions, workspace limits and model availability to decide how the main agent and the Oracle run.

The Oracle is a "second opinion" model.

At high, the main agent uses GPT-6 Astra (medium) and the Oracle uses Claude Fable 5.1. At low / medium / ultra, the Oracle defaults to GPT-6 Astra.

BYOK is now open across all tiers (official news notes: for all tiers since 2026-09-30).

Supported: OpenRouter, Bedrock, Google Cloud Agent Platform, Azure Foundry, Vercel / Cloudflare AI Gateway, Ollama Cloud, OpenCode Go and custom endpoint URLs, and ChatGPT subscriptions can be used too.

⚠ The official sources disagree. The security reference page still says:

Official wording: "Amp doesn't support Bring Your Own Key or self-hosted deployments"

This contradicts the BYOK stance on the pricing page / news; when citing, treat the pricing page as authoritative.
运行位置Where it runs: 混合：本地 Client + 云端 Server。 线程还可跑在三种执行面。

官方安全参考把 Amp 拆为两个组件。

Amp Client 在本地：管代码与上下文管理、本地设置、本地线程历史。

Amp Server 是 ampcode.com 上的多租户云服务，GCP 托管。它做认证、线程同步与 LLM 代理。

线程执行面三分：本地、orb（Amp 在云端为你创建的机器）、 runner（你自己的机器，amp --no-tui --runner-id <id>）。

CLI 用 --executor local|orb|runner:<id> 选执行位置。

另有 Web 与 macOS / iOS 应用。同一 agent、同一线程到处可用。
Hybrid: local Client + cloud Server. Threads can additionally run on three execution surfaces.

The official security reference splits Amp into two components.

Amp Client runs locally: code and context management, local settings, local thread history.

Amp Server is a multi-tenant cloud service on ampcode.com, hosted on GCP. It handles authentication, thread sync and LLM proxying.

Thread execution surfaces come in three: local, orb (a machine Amp creates for you in the cloud), runner (your own machine, amp --no-tui --runner-id <id>).

The CLI picks the execution location with --executor local|orb|runner:<id>.

There are also Web and macOS / iOS apps. The same agent, the same thread — usable everywhere.
本地文件Local files: Client 只带局部上下文，官方明确不索引整库。

官方安全参考原文：

"The Amp Client and Amp Server do not see, store, clone, or index the entire codebase"

它只把「被选作上下文的代码片段 / 整文件」（partial code data）发给 Server。

工具侧有 Read / create_file / edit_file / glob / Grep / Bash 等。

IDE 集成（VS Code、Cursor、Windsurf、Zed、Neovim）：连上后 Amp 能看到当前打开文件与选区。它还能通过你的 IDE 编辑文件、支持完整 undo。
The Client carries only local context; the official explicitly does not index the whole repo.

Official security reference wording:

"The Amp Client and Amp Server do not see, store, clone, or index the entire codebase"

It only sends the Server the code snippets / whole files selected as context (partial code data).

Tool side: Read / create_file / edit_file / glob / Grep / Bash etc.

IDE integration (VS Code, Cursor, Windsurf, Zed, Neovim): once connected, Amp can see the currently open file and selection. It can also edit files through your IDE, with full undo support.
关机后的任务Tasks after shutdown: 有真后台形态——orbs。

官方首页原文：

"Run agents anywhere and they keep working after you close your laptop"

每个线程一台 orb（已含你的代码、工具与插件），可关机续跑、可从手机继续。

runners 在你自己机器上：--share 可把 runner 共享给 workspace，别人也能从 ampcode.com 在它上面起线程（机器需开机）。

定时与自唤醒：官方 manual 有 Schedules。

官方演示视频显示 agent：

官方原文："uses automations to wake itself up"

也就是按计划醒来查生产。
There is a real background form — orbs.

Official homepage wording:

"Run agents anywhere and they keep working after you close your laptop"

Each thread gets an orb (already carrying your code, tools and plugins); it keeps running after shutdown and can be continued from your phone.

Runners live on your own machines: --share can share a runner with the workspace, so others can start threads on it from ampcode.com too (the machine must be on).

Scheduling and self-wake: the official manual has Schedules.

An official demo video shows the agent

official wording: "uses automations to wake itself up"

that is, waking up on schedule to check production.
工具与扩展Tools & extensions: 内置工具 + 子代理 + 插件 + MCP，四层都有官方页。

内置工具：amp tools list 可见的有 Bash、Read、create_file、edit_file。另有 glob、Grep、web_search、read_web_page、Task、todo_*、undo_edit。

子代理 / 工具：oracle（复杂推理第二意见）、 Librarian（跨仓库检索。可读全部公开 GitHub 代码，也可读你的私有仓库）、 Painter（GPT Image 生成与编辑）、Search、Read Thread。

插件系统：可写自定义 policy plugin（含 permissions 插件示例）。也可自定义 agent mode 与 subagent。

MCP：本地（amp.mcpServers，command / args / env，stdio）与远程定义（存 ampcode.com，url / headers）。

传输含 SSE 与 Streamable。

认证含自动 OAuth、bearer、Amp ID token、Workload Identity。

官方建议把 MCP server 打包进 skills（mcp.json）。这样能保持工具列表精简。
Built-in tools + subagents + plugins + MCP — all four layers have official pages.

Built-in tools: amp tools list shows Bash, Read, create_file, edit_file. Also glob, Grep, web_search, read_web_page, Task, todo_*, undo_edit.

Subagents / tools: oracle (second opinion for complex reasoning), Librarian (cross-repo retrieval. Can read all public GitHub code, and your private repos too), Painter (GPT Image generation and editing), Search, Read Thread.

Plugin system: you can write custom policy plugins (including a permissions plugin example). Custom agent modes and subagents are also possible.

MCP: local (amp.mcpServers, command / args / env, stdio) and remote definitions (stored on ampcode.com, url / headers).

Transports include SSE and Streamable.

Auth includes automatic OAuth, bearer, Amp ID token, Workload Identity.

The official recommends packaging MCP servers into skills (mcp.json) — that keeps the tool list lean.
上下文与记忆Context & memory: threads 是核心单位，可保存、检索、分享。

官方 Why Amp 原文：

"Threads: You can save and share your interactions with Amp."

可 @ 引用别的 thread（URL 或 @T-...）。也可按关键词 / 文件 / 仓库 / 作者 / 日期检索历史 thread。 workspace 支持 thread sharing 与 multiplayer。

AGENTS.md 分层：官方列 cwd / 父目录 / 子树、$HOME/.config/amp/AGENTS.md、系统级 /etc/ampcode/AGENTS.md 等层级。子树 AGENTS.md 在读到该子树文件时纳入。

它还支持 @-mention、globs 条件加载，并给出从 CLAUDE.md / .cursorrules 的迁移方式。

压缩：官方 Models 页把 Compaction 列为「Context summarization for long threads」。
Threads are the core unit — saveable, searchable, shareable.

Official "Why Amp" wording:

"Threads: You can save and share your interactions with Amp."

You can @-reference other threads (URL or @T-...). Past threads are searchable by keyword / file / repo / author / date. Workspaces support thread sharing and multiplayer.

AGENTS.md layering: the official lists cwd / parent directories / subtrees, $HOME/.config/amp/AGENTS.md, system-level /etc/ampcode/AGENTS.md and other layers. A subtree AGENTS.md is pulled in when files in that subtree are read.

It also supports @-mention and globs conditional loading, and gives migration paths from CLAUDE.md / .cursorrules.

Compaction: the official Models page lists Compaction as "Context summarization for long threads".
权限与限制Permissions & limits: 默认不审批——这是最需要点出的一条。

官方 Tools 页原文：

"By default, Amp does not ask for approval before running tools."

官方给出的替代是自定义 policy plugin （可分发为 workspace 级 plugin）。

workspace 的 MCP server 需要显式批准。 .amp/settings.json 里的 server 首次检测时提示 awaiting approval，需 amp mcp approve。

全局设置与 --mcp-config 传的不需要。

数据面较硬：官方自动检测并脱敏常见密钥。（AWS / GCP / Azure、GitHub / GitLab、OpenAI / Anthropic、Stripe / Slack / npm 等，替换为 [REDACTED:amp]），但承认是 best-effort。

线程数据存于 ampcode.com（GCP）。默认不用你的数据训练（企业档永不可开）。删除线程后数据 30 天内移除；Enterprise 提供零数据保留（ZDR）。

订阅与 BYOK 额度是两回事。自带 Key 或订阅时，无 Amp token 费用。
No approval by default — this is the point that most needs calling out.

Official Tools page wording:

"By default, Amp does not ask for approval before running tools."

The official alternative is custom policy plugins (distributable as workspace-level plugins).

Workspace MCP servers require explicit approval. Servers in .amp/settings.json show awaiting approval on first detection; you need amp mcp approve.

Global settings and those passed via --mcp-config do not.

The data side is fairly hard-nosed: the official auto-detects and redacts common secrets. (AWS / GCP / Azure, GitHub / GitLab, OpenAI / Anthropic, Stripe / Slack / npm etc., replaced with [REDACTED:amp]), but admits it is best-effort.

Thread data is stored on ampcode.com (GCP). Your data is not used for training by default (never enableable on Enterprise). After deleting a thread, data is removed within 30 days; Enterprise offers zero data retention (ZDR).

Subscriptions and BYOK quotas are two different things. With your own key or subscription, there is no Amp token fee.
适合什么任务Best for: 适合四类用法：

想让 agent 在自己的云端机器（orb）里长跑、合上笔记本仍继续的人
需要多模型前沿路由 + Oracle 第二意见的团队
已在终端或 VS Code 系编辑器里工作、想顺带接 IDE 的人
需要企业级管控（SSO、审计日志、MCP allowlist、零数据保留）的组织
Fits four usage patterns:

people who want the agent to run long on their own cloud machine (orb), continuing after the laptop closes
teams needing multi-model frontier routing + an Oracle second opinion
people already working in the terminal or VS Code-family editors who want IDE hookup along the way
organizations needing enterprise controls (SSO, audit logs, MCP allowlist, zero data retention)

头号误解Top misconception

以为默认会在执行工具前请求批准——官方明确默认不审批Assuming tools run only after approval — officially, Amp does not ask for approval before running tools.
以为有独立 VS Code 扩展——实际靠 CLI + ide connect 接入（VS Code/Cursor/Windsurf/Zed/Neovim）Expecting a standalone VS Code extension — integration is via the CLI and ide connect (VS Code / Cursor / Windsurf / Zed / Neovim).
把订阅当成全部开销——orbs 按分钟、超量按 credits 另计Treating $20/mo as the whole cost — orbs are billed by the minute and overage is charged against credits.

未知项清单Known unknowns

Individual 档 Megawatt 与 Gigawatt 的具体额度差：官方定价页以切换形式呈现，未并列给出（orb 时长额度见官方页；报价按本站的报价记录约定不存）
官方定价页 FAQ 的问答正文：本页只抓到问题标题，具体答复内容未展开
orb 的并发上限与单次最长时长：官方未说明（已查 docs/orbs 与定价页）
自托管部署：官方安全参考称为空（明确不支持），未提供任何自托管路径

The concrete quota difference between the Individual tier's Megawatt and Gigawatt levels: the official pricing page presents them as a toggle, never side by side (orb-hour quotas on the official page; prices not stored per D9 policy)
The answer bodies of the official pricing-page FAQ: this page captured only question titles; answer content was not expanded
Orb concurrency limits and maximum single-run duration: the official does not state (checked docs/orbs and the pricing page)
Self-hosted deployment: the official security reference says it is empty (explicitly unsupported); no self-hosted path is provided

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-23，内容更新日content updated on 2026-10-08。

类型Type	名称Name	链接Link
docs	Amp · 官方首页（frontier agent / orbs，2026-10-08 取到）Amp · official homepage (frontier agent / orbs, fetched 2026-10-08)	https://ampcode.com/
pricing	Amp · 官方定价页（Hobby 免费 / Individual / Teams / Enterprise 四档；报价按本站的报价记录约定不存）Amp · official pricing page (Hobby free / Individual / Teams / Enterprise four tiers; prices not stored per D9 policy)	https://ampcode.com/pricing
changelog	Amp · Chronicle（官方更新日志，含 2026-09-24、2026-08-25 等条目）Amp · Chronicle (official changelog, with entries such as 2026-09-24, 2026-08-25)	https://ampcode.com/chronicle
docs	Amp · Owner's Manual（Agent Modes / AGENTS.md / Orbs / CLI / Pricing）Amp · Owner's Manual (Agent Modes / AGENTS.md / Orbs / CLI / Pricing)	https://ampcode.com/manual
docs	Amp Docs · CLI（安装、--executor、IDE 连接、Streaming JSON）Amp Docs · CLI (install, --executor, IDE connect, Streaming JSON)	https://ampcode.com/docs/cli
docs	Amp Docs · Tools（Oracle / Librarian / Painter、默认不审批）Amp Docs · Tools (Oracle / Librarian / Painter, no approval by default)	https://ampcode.com/docs/tools
docs	Amp Docs · MCP（本地/远程、OAuth、Amp ID token、workspace 审批）Amp Docs · MCP (local/remote, OAuth, Amp ID token, workspace approval)	https://ampcode.com/docs/customize/mcp
docs	Amp · Security Reference（Client/Server 架构、密钥脱敏、数据保留、模型训练）Amp · Security Reference (Client/Server architecture, secret redaction, data retention, model training)	https://ampcode.com/security

实测记录Measured runs

本档未跑端到端编码任务；能力结论来自官方文档，编码向测量协议见 protocols/coding/_protocol.md。

实测建议：重点验三处。

① 无审批默认下跑工具的越权边界（是否只靠 policy plugin 兜底）
② orb 的「合上笔记本继续跑」与暂停计费规则（官方称 5 / 20 分钟阈值）
③ amp --execute ... --stream-json 在脚本 / CI 里的可用性（官方称尽量兼容 Claude Code 格式）

This site has not completed measured runs. The test protocol is at protocols/coding/_protocol.md.

Suggested measurement: focus on three things.

① the over-reach boundary of running tools under the no-approval default (whether only policy plugins backstop it)
② the orb's "keeps running after you close the laptop" behavior and pause-billing rules (the official cites 5 / 20 minute thresholds)
③ amp --execute ... --stream-json usability in scripts / CI (the official aims for Claude Code format compatibility)

核验说明Verification note

confidence: partial 的原因：大部分维度都已从官方页核到。

具体是运行面、执行面、provider / 模式、MCP 接入、权限与数据面。

没核到的是两处。一是 Individual 档 Megawatt / Gigawatt 两级的具体额度差，二是官方定价页「Questions & Answers」问答正文。

前者官方以切换形式呈现、未页面并列给出；后者本次抓取未展开。故不给 verified。

两处需标注的官方源不一致：

① 安全参考页仍写「不支持 BYOK 与自托管」，与定价页 / news 的 BYOK 口径相反（以定价页与新近 news 为准）
② 任务给的 npm 包名是 @sourcegraph/amp，而官方安装文档当前给的是 @ampcode/cli（官方标注 not recommended），两者以官方文档为准

已核到的官方页：首页、定价页、Chronicle、Owner's Manual、docs/cli、docs/tools、docs/customize/mcp、Security Reference。

Why confidence: partial: most dimensions were verified against official pages.

Specifically: the runtime surface, execution surfaces, providers / modes, MCP hookup, permissions and the data side.

Two things were not verified. First, the concrete quota difference between the Individual tier's Megawatt / Gigawatt levels, second, the body text of the official pricing page's "Questions & Answers".

The first is presented by the official as a toggle, never side by side on one page; the second was not expanded in this fetch. Hence no verified.

Two official-source inconsistencies worth flagging:

① the security reference page still says "no BYOK and no self-hosted deployments", contradicting the BYOK stance on the pricing page / news (treat the pricing page and recent news as authoritative)
② the task brief gave the npm package as @sourcegraph/amp, while the official install docs currently give @ampcode/cli (officially marked not recommended) — treat the official docs as authoritative

Official pages checked the 2026-10-08 check: homepage, pricing, Chronicle, Owner's Manual, docs/cli, docs/tools, docs/customize/mcp, Security Reference.