J
juxt-lang/juxt
Juxt-lang: a language for directing AI agents
在 GitHub 查看 ↗链接已于 2026-11-12 核查 · HTTP 200 · 打开的是厂商自己的官网
官方说明Juxt-lang: a language for directing AI agents. Many minds, one intent. Created by Amit N.
安装与配置
客户端配置（取自该仓库的安装命令）
{
"mcpServers": {
"juxt-lang-juxt": {
"command": "npx",
"args": ["juxt-lang"]
}
}
}解析自我们于 2026-10-26 抓取的该仓库 README；我们未执行过。
command/args 来自仓库写明的安装命令；不确定时请以该项目 README 为准。
VS Code（.vscode/mcp.json —— 键名是 servers）
{
"servers": {
"juxt-lang-juxt": {
"command": "npx",
"args": ["juxt-lang"]
}
}
}VS Code 顶层键是 <code>servers</code>，其余与上面那份相同。
Claude Code（命令行）
claude mcp add --transport stdio juxt-lang-juxt -- npx juxt-lang
Claude Code 不收 JSON 文件，用这条命令直接加。
配置项的键名以你的客户端为准：客户端与 CLI
- ★ 数取值
- 1 ★ · juxt-lang/juxt · 取自 GitHub 搜索 API · （2026-11-13 20:00:50 核对）
- 来源明细
- 从该仓库的 GitHub 页面直接读取 · 条目类型：开源项目
- 收录与活跃
- 收录于 2026-10-24 · 最近见到 2026-11-14 · 共见到 25 次 · 被 2 个来源列出
- 评分构成
- 综合分 203 · 1 ★ · 2 个来源 · repo
- 安装
npx juxt-lang
常见问题（答案都来自我们的记录）
怎么安装？
该仓库给出的安装命令是
npx juxt-lang。我们的记录里它是仓库自己提供的；不保证在你的客户端上直接可用。它适合谁 / 用来做什么？
我们把它归在这些标签下: Agent 技能。标签来自条目自身，不是我们强加的分类。
谁维护它、还在更新吗？
仓库
juxt-lang/juxt；我们收录于 2026-10-24；最近见到 2026-11-14。这里没有"官方认证"或"质量评级"——我们只记录来源与时间。Created by Amit N · Contact
以下摘录自该仓库的 README（片段） · 在 GitHub 读完整 README →
Website ·
Quick start ·
Language ·
Examples ·
Roadmap ·
Discussions
▶ Watch the films · a narrated guide and the hand-drawn launch film, both animated in code
Juxt-lang is a language for directing AI agents. You put many models side by side, give them one intent, and get the full strength of each, without the guessing.
You never have to learn it. Add two words to a normal request and your coding agent compiles it into a short, reviewable program before it writes any code. The program holds the objective, the rules, the questions it would otherwise guess, and the checks that prove the work is done.
text
you › use juxt: add a cancel-order endpoint, refunds go back to the card
Understood as
objective POST /orders/{id}/cancel, refund to the original card
MUST idempotent writes (a retried cancel never refunds twice)
verify cancel twice -> 1 refund · shipped order -> 409
OPEN cancel after shipping? partial refunds?
## Why Juxt
| The problem with agents today | What Juxt does |
| Agents guess at unclear requirements and you find out in review | Unclear points become OPEN questions, asked before any code is written |
| House rules live in people's heads or in long prompts | Rules are MUST / MUST NOT lines that outrank the objective and are checked by a fresh reviewer |
| "Done" means "it compiled" | Every program carries verify checks that must pass |
| Each tool (Claude Code, Codex, CI) needs its own prompting | One program, compiled to each harness's native skills, commands and subagents |
| The same status questions are re-asked of a model every day | Digests and status updates come from git, with citations; repeat asks are served from cache |
| Nobody can audit why an agent did what it did | Programs are plain-text .juxt files in your repo, versioned and reviewable like code |
## How Juxt compares
| | Juxt | GitHub Spec Kit | Kiro | AGENTS.md |
| Format | Formal grammar, typed IR, JSON Schema | Markdown per phase | Markdown: EARS requirements, design, tasks | Free-form Markdown |
| Checks | lint, validate, drift check in CI | Checklists; CI gates via extensions | Manual review | None defined |
| Maturity | Early (0.7), few users | Widely adopted | AWS product | De facto standard |
The others are ahead on maturity, integrations and ecosystem. Juxt's difference is a program you can check mechanically and a build that fails on drift. They work together: AGENTS.md describes the repo, a .juxt program states one change and how to prove it. Full comparison: juxt-lang.github.io/#/home/compare.
## Does it help? Measure it
bench/ runs the same agent on the same tasks with and without Juxt, scores them with hidden acceptance tests, and records the agent's own token usage and cost. No results are claimed yet; they will be published unedited, losses included.
## How it works
mermaid
flowchart LR
A["Plain request 'use juxt: …'"] --> B["Compile objective · rules · contract"]
B --> C{"OPEN questions?"}
C -- yes --> D["Ask you"] --> B
C -- no --> E["Build native skills and subagents"]
E --> F["Verify tests + fresh reviewer vs every MUST"]
F --> G["juxt/.juxt saved in git"]
A compiled program is short and readable (examples/cancel-order.juxt):
Quick start ·
Language ·
Examples ·
Roadmap ·
Discussions
▶ Watch the films · a narrated guide and the hand-drawn launch film, both animated in code
Juxt-lang is a language for directing AI agents. You put many models side by side, give them one intent, and get the full strength of each, without the guessing.
You never have to learn it. Add two words to a normal request and your coding agent compiles it into a short, reviewable program before it writes any code. The program holds the objective, the rules, the questions it would otherwise guess, and the checks that prove the work is done.
text
you › use juxt: add a cancel-order endpoint, refunds go back to the card
Understood as
objective POST /orders/{id}/cancel, refund to the original card
MUST idempotent writes (a retried cancel never refunds twice)
verify cancel twice -> 1 refund · shipped order -> 409
OPEN cancel after shipping? partial refunds?
## Why Juxt
| The problem with agents today | What Juxt does |
| Agents guess at unclear requirements and you find out in review | Unclear points become OPEN questions, asked before any code is written |
| House rules live in people's heads or in long prompts | Rules are MUST / MUST NOT lines that outrank the objective and are checked by a fresh reviewer |
| "Done" means "it compiled" | Every program carries verify checks that must pass |
| Each tool (Claude Code, Codex, CI) needs its own prompting | One program, compiled to each harness's native skills, commands and subagents |
| The same status questions are re-asked of a model every day | Digests and status updates come from git, with citations; repeat asks are served from cache |
| Nobody can audit why an agent did what it did | Programs are plain-text .juxt files in your repo, versioned and reviewable like code |
## How Juxt compares
| | Juxt | GitHub Spec Kit | Kiro | AGENTS.md |
| Format | Formal grammar, typed IR, JSON Schema | Markdown per phase | Markdown: EARS requirements, design, tasks | Free-form Markdown |
| Checks | lint, validate, drift check in CI | Checklists; CI gates via extensions | Manual review | None defined |
| Maturity | Early (0.7), few users | Widely adopted | AWS product | De facto standard |
The others are ahead on maturity, integrations and ecosystem. Juxt's difference is a program you can check mechanically and a build that fails on drift. They work together: AGENTS.md describes the repo, a .juxt program states one change and how to prove it. Full comparison: juxt-lang.github.io/#/home/compare.
## Does it help? Measure it
bench/ runs the same agent on the same tasks with and without Juxt, scores them with hidden acceptance tests, and records the agent's own token usage and cost. No results are claimed yet; they will be published unedited, losses included.
## How it works
mermaid
flowchart LR
A["Plain request 'use juxt: …'"] --> B["Compile objective · rules · contract"]
B --> C{"OPEN questions?"}
C -- yes --> D["Ask you"] --> B
C -- no --> E["Build native skills and subagents"]
E --> F["Verify tests + fresh reviewer vs every MUST"]
F --> G["juxt/.juxt saved in git"]
A compiled program is short and readable (examples/cancel-order.juxt):
收录徽章（可选）
想把收录标记放进 README 就复制这行，可选，不影响收录。
[](https://anclum.com/zh/s/juxt-lang-juxt/)
- 来源
- github · https://github.com/juxt-lang/juxt
- 按领域浏览
- Agent 技能
- Traction
- 1 ★
- 收录来源
- 2 个独立来源
- 收录于
- 今天
同类推荐
所属领域与标签
每个标签都能点进去看同类的全部条目。数字是当前库里的真实条数。