Zero dependencies: Yan needs only Node.js 22 or later. No npm packages to install; every front-end library ships with the project and nothing loads from the web.
Double-click start.cmd, or run in the project folder:
npm startYour browser opens http://127.0.0.1:8787. The first time, fill in a display name, model ID, Base URL and API key in Settings → Model to start; the project never reads or bundles any configuration from the developer's machine. You can add several APIs and switch the default model. Each model chooses its API: OpenAI compatible (chat/completions) or Anthropic (Messages API); a Base URL containing anthropic.com is recognized even if you don't choose.
Switch the interface language in Settings → General → Language (中文 / English); the page reloads after switching. If you never chose, it follows the browser language: Chinese for Chinese locales, English otherwise. The English interface only changes the words on screen; the prompts sent to the model stay the same. Settings → Docs is Chinese only for now.
Settings → General → Desktop notifications is off by default. When on and the page isn't in front of you (another window or app), a system notification pops up when an answer finishes, something awaits your decision (the approval bar, a form), or a subagent breaks off; a newer one for the same chat replaces the older, and clicking it takes you back to that chat. If a follow-up started another turn or subagents are still working, the chat isn't idle yet, so nothing pops up; nor when you pressed stop. Permission is kept per browser and asked the first time you click On; settings sync through the storage folder, so a browser that hasn't allowed it stays quiet until you click On there again.
The CMD window opened by start.cmd is the local bridge process and must stay open. If you close it, pages already loaded stay for a while, but reloading, model forwarding, web access and the agent all stop. If the browser is open but the page says the local bridge has stopped, look at that window first: when the bridge exits it stays open and prints the reason (port in use, node missing and so on) instead of flashing away.
If you'd rather not install Node first, use the portable package: it unzips to a 言 folder; double-click 言.exe inside it and everything else is as above. 言.exe opens no window: the bridge runs in the background, and the commands it starts don't pop up console windows either. Close Yan's page and the bridge exits on its own once nobody has reconnected for 30 seconds and no background command is running (enough for a reload or reopening the browser; background commands keep running and the bridge exits after they finish); double-clicking while Yan is already running just opens the page. The bridge's output goes to 桥接.log in the 言 folder (rewritten on each start), and if the bridge exits unexpectedly a dialog shows its last lines. To debug with a terminal in view, use start.cmd instead. The package bundles the official Node (node\, including npm), and Yan's own files are identical to the repository. start.cmd also uses this node\, so the two are equivalent. The bundled Node is appended to the end of PATH: the model's node and npx, and front-end tools installed through Environment, can find it, while a system that already has Node keeps using its own. Chats, Files and settings stay in ~/.yan/, so switching between the portable package and source works seamlessly.
言.exe is not code-signed. The first time you double-click it, Windows may say "Windows protected your PC"; click More info → Run anyway.
Packaging: run npm run pack in the repository to get dist/yan-<version>-win-x64.zip. It uses the system's curl.exe, tar.exe and the .NET Framework 4 csc.exe (all present on Windows 10 / 11). The Node version is set at the top of pack/pack.mjs; the archive is used only after it matches the nodejs.org checksum. After changing the version, run the whole test suite with the bundled Node (dist\言\node\node.exe test/run.mjs). The launcher's source is pack/launcher.cs; the icon pack/icon.ico is drawn from icon.svg.
- Run the task "言：启动模型桥接" (start the model bridge) via Terminal → Run Task, and keep its terminal open.
- Open
http://127.0.0.1:8787in your system browser. Simple Browser embeds the page in an iframe, which Yan'sframe-ancestors 'none'policy blocks; the bridge also doesn't open the agent API to extension webview origins. - Fill in your settings in Settings → Model; you can click Fetch list and then Test connection. Web search is tested in Settings → Tools.
Yan must be opened through the local bridge: run start.cmd (or npm start) and visit http://127.0.0.1:8787 in your system browser. Double-clicking index.html (file://) can't reach the bridge: the bridge refuses pages whose origin is null (a file:// page is one, and so is a sandboxed iframe embedded by another site; they can't be told apart). After editing src/ or styles/, just reload (the bridge concatenates on each request and writes no build output).
YAN_PORT: change the bridge port; default8787YAN_HOME: set the storage folder directly (chats, Files and settings all live in it); default~/.yan/. When set, the file that remembers the storage location,%APPDATA%\言\位置.json, is neither read nor written (used by tests)YAN_DEBUG=1: print every request and every abnormally closed connection to the console. Use it when the page reports "Failed to fetch" or that the local bridge has stopped, to see whether the request reached the bridge at all
server.js and server/ are the local bridge (see Code structure and development for what each part does). It serves the page, forwards model requests, runs web search and page reading on the model's behalf (refusing local and private addresses, with IPv6 literals and IPv4-mapped addresses judged the same way; bodies are truncated while reading, at most 3 MB), and performs the agent's commands and file operations. Static files carry ETag validation, so changes take effect on reload; the console never shows API keys.
A single request carries at most 128 tool definitions; going over is an explicit error rather than silently dropping the rest.