
Cold Open
io.github.TheYungStump由 GitHub 帳號 TheYungStump 發布v0.1.0-alpha.3更新於 Nov 7, 2026
Screenplay editing for agents: Fountain and .fdx; edits arrive as tracked changes in Cold Open.
概覽
Cold Open 編劇軟體的 MCP 伺服器,讓助理以修訂模式編輯 Fountain 與 Final Draft 劇本。
- 可以做什麼
- 開啟、讀取與編輯劇本檔案,劇本、區塊、頁面與場景以資源形式提供
- 插入區塊、取代文字,每處修改在桌面應用程式中顯示為修訂
- 列出場景並檢視文件結構
- 復原操作,且代理只能復原自己的變更
- 可連上桌面應用程式中已開啟的劇本,應用程式未執行時則直接開啟檔案
- 適用情境
- 適合編劇想讓 AI 助理在 Fountain 或 .fdx 劇本上修改,並保留可見、可復原的修訂紀錄,而不是整份重寫檔案的情況。
- 執行需求
- 以 coldopen mcp 指令在本機執行,隨 Cold Open 桌面應用程式安裝(Windows、macOS、Linux 的 alpha 版本),或以 Rust 1.89 以上從原始碼建置。未宣告需要帳號、API 金鑰或環境變數。Linux 上應用程式需要 X11 或 Wayland 以及 libxkbcommon。
安裝
在 SourceWeft 中
- 開啟 儀表板中的 Cold Open,將其新增到工作區。
- 為需要使用其工具的對話啟用該服務。
Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。
其他 MCP 客戶端
參照 儲存庫 中的啟動說明。
README
Cold Open
Cold Open is a free, open-source screenwriting app. It reads and writes Fountain and Final Draft (.fdx) files and never loses your work. Every action is a command, and you can drive it from the desktop app, the command line, JSON over stdio, or an MCP server.
Status: alpha, and looking for testers. Phases 1 to 5 built the editor, the command line, the MCP server, .fdx exchange, track changes, revisions, script diff and reports; phase 6a made the builds a writer installs, and 6b the user guide, feedback and crash reports. The website, https://theyungstump.github.io/cold-open/, has the user guide and how to join the alpha. See docs/spec.md for the plan, docs/progress/ for each phase and CHANGELOG.md for each release.
Install
Alpha builds for Windows (x64 and ARM64), macOS (Apple Silicon and Intel) and Linux (x86-64 and ARM64; .deb, AppImage and Flatpak) are on the releases page. Each download is under 15 MB. They aren't signed by Microsoft or notarized by Apple yet (see the code signing policy); the release notes say how to get past the first-start warning on each system. The command-line tool coldopen comes with the app on every system (docs/specs/release.md §4).
To help us measure Cold Open on your computer, choose Measure performance in the palette, or run coldopen bench, and send us the two small files it writes to your Downloads folder (docs/specs/bench.md). They hold model names and times, and nothing personal.
Code signing policy
Free code signing provided by SignPath.io, certificate by SignPath Foundation.
Cold Open is applying to SignPath Foundation's program for open-source projects. Until the project is accepted, the Windows builds are unsigned, and Windows warns about them the first time they run.
Team roles:
- Authors (committers), who may change the source without another review: TheYungStump.
- Reviewers, who review every change from anyone else, such as a pull request, before it is merged: TheYungStump.
- Approvers, who approve each signing request by hand: TheYungStump.
How a release is signed:
- Each release is built from source by GitHub Actions, from a tagged commit of this repository (
.github/workflows/release.yml). - Only Cold Open's own programs are signed: the desktop app, the command-line tool and the Windows installer.
- An approver approves each signing request by hand.
- Every team member uses multi-factor authentication for GitHub and for SignPath.
Much of Cold Open's code is written with AI coding agents, working under the author's direction and GitHub account. The author reviews every change before it is merged.
Privacy
This program will not transfer any information to other networked systems unless specifically requested by the user or the person installing or operating it.
In short: the website counts visits anonymously with Cloudflare Web Analytics, without cookies; the app sends nothing.
In practice:
- Cold Open makes no network connections. It doesn't check for updates or contact any server.
- A usage summary, on your computer only (from 0.1.0-alpha.3). Cold Open counts which features you use, on this computer only, and it's never sent unless you include it in feedback. It holds counts, never your text, names or file paths. Settings (Ctrl+Comma, or Cmd+Comma on a Mac) shows it in full, resets it, or turns counting off. (
docs/specs/usage-summary.md) - The agent link stays on your computer. The app and
coldopen mcptalk through a socket or a pipe that only your user account can open (docs/specs/mcp.md). An AI agent you connect is a separate program: what it reads from your script, and where it sends it, is up to that program and its own privacy policy. - Links open in your browser, and only when you click them.
- Measure performance writes two report files to your Downloads folder. They leave your computer only if you send them.
From 0.1.0-alpha.2, there are two ways to tell us about a problem from inside the app. Neither sends anything by itself:
- Report a problem… opens a page in your browser, ready to fill in: a GitHub issue, or a Google Form for people without a GitHub account. The app fills in Cold Open's version and commit, the build's target, your operating system's name and version, and the renderer drawing the window. It adds no script text, no file names or paths, and no user or computer names. Nothing is sent until you submit the page yourself, and you can change it first.
- Crash reports. If Cold Open crashes, it saves a report on your computer: the error's message with any quoted text taken out, where in Cold Open's code it happened, a backtrace as code addresses with no file paths, the version, the operating system and the renderer. It holds no script text, file names or user names. At the next start, Cold Open asks whether you want to look at it. It shows you the whole report before anything leaves your computer. Then it can open a ready-filled issue or form, or copy the report for you to paste. You can choose Never ask again.
The website counts visits anonymously with Cloudflare Web Analytics, without cookies, in aggregate; nothing else measures visitors (the site's privacy page).
GitHub issues are public, under GitHub's privacy statement. Form answers go privately to the project, through Google Forms, under Google's privacy policy.
Build
You need Rust 1.89 or newer. Install it from https://rustup.rs, then build:
This builds two programs in target/release: coldopen-app, the desktop app, and coldopen, the command-line tool, which also runs the MCP server. Both run on Windows, macOS and Linux (.exe on Windows). On Linux the app needs X11 or Wayland, and libxkbcommon, at run time.
Packages for every system are built by CI (.github/workflows/build.yml, .github/packaging/package.sh); every pull request's run has them under Summary ▸ Artifacts.
Use the desktop app
- Writing. A new script starts on a scene heading. Enter and Tab move between elements as screenwriting expects (heading, action, character, dialogue, parenthetical, transition); the bar at the top says what they will give next. Typing in headings, character names and transitions comes out in capitals. Autocomplete offers character names,
INT./EXT., locations, times of day and extensions; Tab or Enter takes a suggestion. - Everything from the keyboard. Ctrl+K (Cmd+K on macOS) opens the command palette, which lists every command, every app action and their shortcuts, and recent files.
- Pages. The script is shown on pages exactly as it prints, or in a continuous draft view (Ctrl+Shift+D). The navigator on the left lists scenes; drag one, or press Alt+Up/Down, to move it.
- Files. Ctrl+S saves (in the background, without pausing your typing), Ctrl+Shift+E exports a PDF, and closing asks about unsaved changes.
- Your own keys and element flow.
keymap.jsonandflow.jsonin the settings folder change shortcuts and what Enter and Tab do; seedocs/specs/editor.md§5.2 and §10. The app never rewrites them. - Drawing. The app draws with the CPU on Windows and Linux and with the GPU on macOS. The GPU is smoother on 4K and Retina screens; the CPU uses far less memory. The palette switches between them.
The full behaviour is in docs/specs/editor.md.
Let an AI agent edit your script
coldopen mcp is an MCP server: every command is a tool (block_insert, text_replace, scene_list and so on), and the script, its blocks, pages and scenes are resources. Any MCP client works the same way.
When the script is open in the desktop app, the agent's doc_open attaches to the app instead of opening the file: you see each edit appear, a violet bar marks the blocks it changed, and Ctrl+Z undoes it like your own. Each agent step is labelled with the agent's name and the tool it used, and an agent can only undo its own steps. The status bar shows a connected agent, with Disconnect and a switch that turns agent access off.
The link between them is private to your user account on your machine: a socket in a folder only you can open, or a Windows pipe only you can reach, plus a fresh 256-bit token. Without the app, coldopen mcp opens the file itself, with the same crash journal. Details: docs/specs/mcp.md.
Use the command line
Convert a script
This rewrites a script as canonical Fountain. The meaning is unchanged; spacing and markup are tidied.
To print the document's blocks and ids as JSON instead:
Use - as the output to write to standard output. If the target format cannot hold everything in the script, nothing is written unless you add --allow-lossy.
List scenes
The output looks like this:
Each line shows the page the scene starts on and its length in eighths of a page. Add --json for machine-readable output.
Export a PDF
The PDF uses the bundled Courier Prime font and comes out byte for byte the same on Windows, macOS and Linux. --no-title-page leaves out the title page. Characters the font has no glyph for (for example Chinese or emoji) print as a box, and the command says which.
See the pages as text
prints every page the way it prints, as plain text on a 10-characters-per-inch grid, with a header line per page saying why it ended. It is handy for checking page breaks.
Licenses
coldopen licenses prints Cold Open's license and the SIL Open Font License of the bundled fonts. The app shows them under About in the palette.
Run a batch of edits
edits.json is a JSON array of commands, or one command per line:
Run it against a script:
runprints one JSON response per command.- It stops at the first failure unless you add
--keep-going. - It exits with status 1 if any command failed.
- Without
--save(or adoc.savein the batch), changes are discarded when the batch ends. - Unsaved work recovered from a crash journal is never thrown away. If the batch leaves it untouched, it stays in the journal for the next open; if the batch changed it, a copy of the journal is kept, and
runsays where.
Serve commands over stdio
The server reads one JSON request per line and answers each with one JSON response line. Any language can drive it this way:
The full command list, parameters and error codes are in docs/specs/commands.md.
Your work is safe
The crash journal. While a script is open, every change is written to a crash journal before the command returns. For draft.fountain, the journal is the hidden file .draft.fountain.coldopen-journal next to it.
- If Cold Open crashes, or the machine loses power, the next
doc.openreplays the journal and returns"recovered": {"commands": N}. Your undo history comes back too. The desktop app shows what came back and lets you keep it or go back to the saved file. - In the desktop app, every key is written to the journal before it shows, and flushed to disk within 250 ms. Ending the app at any moment loses nothing that was on screen; a power cut loses at most the last quarter of a second. A script you never saved has a journal too, and the app offers it at the next start.
- Saving resets the journal, and closing a saved document deletes it.
- If the script is in a git repository, add
.*.coldopen-journalto your.gitignore.
Atomic saves. Saves are atomic: a crash mid-save leaves the old file or the new one, never half of each.
The details are in docs/specs/journal.md.
Fountain support
Cold Open supports everything on the Fountain syntax page:
- the title page
- scene headings, including forced headings and scene numbers, plus the reversed
EXT/INTandE/Iprefixes that some writers use - action, characters, dialogue, parentheticals and dual dialogue
- lyrics, transitions and centered text
- emphasis
- notes, boneyard, sections and synopses
- page breaks
Every rule, and how we settle the cases the syntax page leaves open, is written down in docs/specs/fountain.md.
Test
The test suite includes:
- Fountain conformance cases. One case per rule, in
tests/fixtures/fountain, with expected results written by hand. - Property tests. Random Fountain and random edit sequences check round trips, undo, redo and journal replay.
- A crash test. It kills
coldopen serve --stdioin the middle of a stream of edits, then checks that a new process recovers every edit that had been acknowledged. - The app's gates. The app's own tests run it headless, with real key presses through the real interface. The typing gate types a short scene with smart flow, autocomplete, accents, an undo and a save, and compares the file byte for byte; the crash gate kills the app as a separate process at random points mid-typing and checks the recovery; the agent gate drives a real
coldopen mcpprocess against the app.
Private corpus
Real scripts are copyrighted, so they are never committed. Put any .fountain files under corpus/ (which is gitignored), or point COLDOPEN_CORPUS at another folder. The corpus test then checks that each script round-trips without loss. Without a corpus, that test is skipped.
Fuzzing
The fuzz/ folder holds cargo-fuzz targets. They need nightly Rust and run on Linux or macOS:
The other targets are envelope, journal_replay, fdx_read, link_line and app_config. A scheduled workflow runs all six every Sunday; you can also start it by hand from the Actions tab.
Layout
The format libraries are permissively licensed so other apps can use them. The app itself is GPL-3.0-or-later; see LICENSE. The Courier Prime and Atkinson Hyperlegible Next fonts are under the SIL Open Font License 1.1 (crates/fonts/fonts/); PDFs you export are yours and are not covered by it.
tools/refpdf is a development tool that measures reference PDFs for the layout tests; it is never shipped.
來源:README.md,提交 4de7b2f
工具
0版本歷史
1- v0.1.0-alpha.3最新Nov 7, 2026
評分
已滿足適用檢查項的 50%
該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。
- 近 30 天內可連通25 分
請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。
- 已宣告認證方式10 分
請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。
- 已列出工具15 分
請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。
- 描述必需15 分
- 圖示10 分
- 文件15 分
- 授權條款10 分