返回资讯中心
外部精选
软件工程
#开发者实时流

Show HN: Harness Studio: an open-source desktop coding agent for any model

A desktop coding agent for macOS and Linux. Any model, on your code and UI.

Hacker News 最新VigneshMurugan2026/10/20 · 26 天前30 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

A desktop coding agent for macOS and Linux. Any model, on your code and UI.

Download for macOS · Download for Linux · What's new

Harness Studio is a desktop app that works like Claude Code desktop, with whichever model provider you have an API key for. Download it, choose a model, open a project folder and ask.

Point at your UI, tell the agent what to change. Open your app's dev server in the Browser pane, pick an element, and say "make this bigger": the agent gets the element's source file and line, component, text, styles and HTML.

A browser the agent can use: it opens pages, reads them, clicks and types in the Browser pane, next to the conversation, with real mouse and keyboard input. In the desktop app it also reads what DevTools shows on your own development pages (localhost and your local network): console errors with stack traces, network requests with their status, and response bodies. It never attaches DevTools to public sites.

Documents, decks and spreadsheets: the agent writes PowerPoint, Word and Excel files, with editable charts in decks and spreadsheets, renders slides to check its own layout, and exports PDF, using a Python, Node and office converter bundled with the app, so nothing needs installing. Results open in the side pane's preview.

Design boards: ask the agent to design or mock up screens and it drafts them on one board, each screen and variant side by side at phone, tablet and desktop widths, using your project's own colours and fonts, before any app code changes. The board is a single HTML file with pan and zoom that opens in the preview pane or any browser.

Desktop and mobile apps: the agent builds Mac, Android and Linux desktop apps (and iPhone apps when Xcode is installed) inside its sandbox, keeping every tool's downloads in the project, then runs them and checks them: Mac apps through computer use, Android apps in a headless emulator with its screenshots, and web-based ones in the Browser pane.

Live dashboards: ask for a dashboard and the agent builds a page of charts with a small refresh script per data source (a spreadsheet, CSV, a database or warehouse through your own login, or a web API). The Dashboards pane fetches fresh data every time you open it, once you have approved the refresh scripts at their exact text; Keep fresh refreshes it while closed, and Save snapshot gives you one HTML file with today's data to send.

Charts, icons and diagrams: editable charts inside PowerPoint and Excel files, and SVG files for standalone charts, icon sets and flow or architecture diagrams, drawn from your data with nothing to install.

Speech to text, on this computer: the microphone button between the model picker and Send dictates into your message (a menu in the recording bar picks the language), and the agent can transcribe an audio or video file, or a part you marked in the Files tab, with the time each piece was said, in Chinese, English, Cantonese, Japanese or Korean. Recognition is Harness's local SenseVoice model (about 241 MB, downloaded once after you agree); no audio leaves the computer. When the agent needs ffmpeg and you have none, it asks once and downloads a pinned, checksum-checked build into Harness Studio's data folder.

Full-screen programs: the agent can run vim, less, top, interactive prompts and installers in a real terminal, read the screen and type into it, under the same sandbox as every other command.

Edit files yourself: press Edit on a file in the Files tab for a code editor with syntax colours and ⌘S; the agent's changes appear in place, and neither of you overwrites the other's unsaved work.

Side panes: Files, Changes, Session details, Browser and Terminal, stacked with a draggable divider.

Git history: a Git page in the sidebar shows every workspace's branch, upstream and uncommitted files, and each workspace's commits with a branch graph, search by message, author or changed code, every commit's files and diffs, and blame. Its Branch menu switches branches when nothing is uncommitted.

Connectors: add MCP servers by URL or command, with OAuth sign-in and per-session on/off.

Separate sessions: the agent can hand a piece of work to a new session in the sidebar with a written brief, optionally in its own git worktree, and gets that session's report back. For a self-contained job whose material is not needed afterwards, such as judging a batch of rendered frames, it uses a helper agent and keeps only the verdict.

Watches: the agent (or you, from the eye button in the session header) can watch a command's output, a port or files; the checks run on your computer without the model, and the agent is woken with a short notice only when something changes.

Remember: when you correct the agent or ask it to remember something, it proposes a one-line note; you approve it and choose this project or all your projects, and it's saved as a line in an instruction file you can edit.

Your own Chrome: an optional paired extension lets the agent browse in your real Chrome, where you are signed in, and lets you pick elements on any page with ⌥⇧S.

A calmer look: Claude Code–like density, text sizes, sidebar and zoom.

Secret files stay out of reach: commands the agent runs, and its file tools, cannot read your SSH keys, cloud and registry credentials, password files, .env files outside the project, or the app's own API keys, so a prompt injection in a README or web page cannot make the agent send them anywhere. What the agent can still send, and where, is spelled out in the app's About page under Safety model.

Command rules, like Claude Code's: block a tool call or ask first, for example never curl, ask before git push, never read ./.env, in Claude Code's permissions format, in ~/.harness-studio/permissions.json or a project's .harness-studio/permissions.json.

No account, no telemetry: there is no sign-in and no usage reporting, and your API keys are stored only on your machine and sent only to the provider they belong to.

Harness Studio is built on DeepSeek Harness, the open-source, any-model agent app. It is an independent project, not affiliated with or endorsed by DeepSeek or Anthropic; their names are used only to describe compatibility and inspiration.

Dark mode, with a reply that cites files and code from the project:

Files pane: the project's tree beside open files with line numbers and syntax colours. Type ? in the filter to search inside files.

Changes pane: the working tree against its branch, with per-file counts and coloured diffs.

Separate sessions: the agent proposes handing a task to a new session with its own git worktree, and you start it or keep the work in this session.

Connectors: MCP servers such as GitHub and the Chrome extension, each with its status and tool count, added by URL or command.

The + menu: files, goals, plan mode, commands and per-session connectors in a compact popover above the message box.

Session pane: the session's changes and its usage at a glance.

Automation tasks: ask for a follow-up later, and the agent schedules it and replies in the same session when it fires.

Developer preview, version 0.3.2, built on Harness 0.2.1-rc.2 through Harness Studio's fork of it, harness-studio-core (pinned in docs/HARNESS_PIN.md). Tested on macOS (Apple Silicon) and Linux (x64 and arm64). Windows isn't available yet.

  1. Install Harness Studio

Harness Studio is a ready-to-run desktop app with the plugins built in, so you don't need DeepSeek Harness, Node.js or a build. Download it from the Releases page:

Windows isn't available yet. How each app is built, and what to know before installing on Linux, is in desktop/README.md.

The Mac app isn't signed or notarised by Apple, so macOS blocks it the first time. Allow it once:

After that it opens normally. If macOS instead says the app "is damaged and can't be opened", the download's quarantine flag is the cause, not a broken file; clear it and open the app again:

xattr -dr com.apple.quarantine "/Applications/Harness Studio.app"

To check the download first, compare shasum -a 256 Harness-Studio-<version>-mac-arm64.dmg with the .sha256 file published next to it.

On first launch, click Choose a model provider and add an API key for any provider in Settings → Models.

Chrome extension (optional): follow chrome-bridge/README.md to load the extension and add its connector, so the agent can use your own Chrome.

The companion script and source loader come from this repository: git clone https://github.com/vigneshv1cky/harness-studio.git, then run the commands below from its folder.

a. Companion script — copy it into your app and load it only in development.

cp plugins/devpreview-companion/companion.js /path/to/your-app/public/devpreview-companion.js

Next.js App Router, in app/layout.tsx (inside <body>):

import Script from "next/script";

{process.env.NODE_ENV === "development" && (

<Script src="/devpreview-companion.js" strategy="afterInteractive" /> )} Plain HTML: <script src="/devpreview-companion.js"></script> in development pages only.

b. Source locations (Next.js 16 / Turbopack) — install the loader and add the rule.

npm install --save-dev /path/to/harness-studio/plugins/devpreview-source-loader

// next.config.ts const nextConfig: NextConfig = { allowedDevOrigins: ["127.0.0.1"], // only needed if you open the dev server as 127.0.0.1 turbopack: { rules: { "*.{jsx,tsx}": { condition: { all: ["development", { not: "foreign" }] }, loaders: [{ loader: "devpreview-source-loader", options: { root: process.cwd() } }], // no as — it would rename modules and break relative imports }, }, }, };

Without the loader, Pick still works; the agent locates the element from its text, selector and component names instead of an exact line.

Start your dev server (npm run dev).

In Harness Studio, set the Session's workspace to your app's folder (so the agent can edit without extra approvals and paths match the stamps).

Click the Browser icon in the conversation header and open your dev server: type localhost:3000, or pick it from the Dev servers menu in the pane header.

Click Pick (the arrow) in the bar at the bottom of the browser, hover, click an element. The card shows its source, text, size, selector and styles.

Click Send to agent — a reference like [Selected <button> in CounterButton · src/app/counter-button.tsx:8:5] lands in the message box.

Type what you want ("make this bigger", "use the brand blue", "move it under the heading") and send.

✕ on the card clears the selection for the agent.

Your instructions

Harness Studio follows instruction files the way Claude Code follows CLAUDE.md. It reads, in this order:

Each folder contributes one file, so a repository that keeps both AGENTS.md and CLAUDE.md for different tools isn't charged twice. Files can include other files with an @path line, as in Claude Code; a line that includes an instruction file Harness Studio already reads, such as a CLAUDE.md containing just @AGENTS.md, is skipped. A project's files may only include files inside that project; your personal file may include any file, so to reuse your Claude Code rules put this line in ~/.harness-studio/AGENTS.md:

@~/.claude/CLAUDE.md

Instructions, in the left sidebar below Automation tasks, lists every file a workspace reads with its size, token cost and status, lets you edit or open each one, and sets the size limit. All instructions together may use up to 128 KB by default; the page offers 64 KB to 1 MB, or no limit. When they're larger, the broadest files are left out first and then the most specific one is shortened, and the instructions row at the start of the conversation names each file left out or cut, with its size. Editing an included file takes effect in the next session.

From 0.3.0, Harness Studio includes a harness-studio command for scripts, CI and one-off jobs. On macOS, choose Install Command-Line Tool… in the app menu, below Check for Updates…; it adds /usr/local/bin/harness-studio and may ask for your

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →