TradingView MCP系统 ·由Ferrox实验室制作

告诉你的AI你想从TradingView图表中得到什么。在屏幕上观看。

TVControl 将你的 TradingView 桌面变成你可以对话的平台。你输入一句话（“总结这张图表”，“在 SPY、QQQ 和 IWM 上 5 万和 15 百万扫描此策略”，“逐条逐条分析去年三月的突破点”），AI 会在你机器上的实际 TradingView 应用内读取、点击、输入、编译并截图。不复制粘贴，也不使用 TVControl 运营的云后台。TradingView 桌面版和明确选定的公共 API 助手仍能与 TradingView 通信，如文档所示。

它之所以有效，是因为每个 Chromium 应用，包括 TradingView Desktop，都内置了调试界面（Chrome 用来调试的同一个接口）。TVControl 代表你使用该接口，向任何使用 Model Context 协议的代理开放**113 个图表控制和诊断工具*。它是客户端中立的：任何能通过标准节点启动 MCP 服务器的设备都能正常工作，并且在多个平台上每天都在使用。用 Pine Script 配对编程。优化参数网格。快照并恢复整个图表设置。

TVControl

TradingView MCP System · by Ferrox Labs

Tell your AI what you want from your TradingView chart. Watch it happen on screen.

TVControl turns your TradingView Desktop into something you can talk to. You type a sentence (“summarise this chart”, “sweep this strategy across SPY, QQQ and IWM on 5m and 15m”, “step through last March bar by bar and call out the breakout”) and the AI reads, clicks, types, compiles and screenshots inside the actual TradingView app on your machine. No copy-paste and no TVControl-operated cloud backend. TradingView Desktop and explicitly selected public-API helpers still communicate with TradingView as documented.

It works because every Chromium app, TradingView Desktop included, ships with a built-in debugging interface (the same one Chrome uses to debug itself). TVControl speaks that interface on your behalf, exposing 113 chart-control and diagnostic tools to any agent that speaks the Model Context Protocol. It is client-neutral: anything that can launch an MCP server over stdio works, and it is in daily use across several. Pair-program in Pine Script. Optimize parameter grids. Snapshot and restore whole chart setups. Drive 4-pane layouts. Step through replay. Scan a watchlist. All by speech-to-action.

113 MCP tools · 938 deterministic offline tests · 10 verify scripts · 9 prompt-library workflows · no TVControl cloud backend. Everything in this repo is real, tested, and used daily.

What is new in 2.5

The 2.5 line is about one thing: a tool must build its answer from an independent read, never from the request or from the response of the thing it just did. Every fix below is a place where that was not true.

• A result you can act on when a call times out. chart_set_symbol and chart_set_timeframe had already applied the change by the time their readiness wait expired, then threw as though nothing had happened, so a caller that retried applied it twice. They now return chart_ready: false and say not to retry. indicator_add_from_search polls for the study instead of sleeping 1500ms and guessing. The contract is now stated in the tool descriptions: a thrown error means nothing changed.
• pine_open no longer risks the wrong script. It fetched a script’s source and pasted it into whatever buffer was open, leaving the editor bound to the previous script while reporting success. A following save wrote over the wrong file. It now reads the editor’s own title back and refuses on a mismatch. New in this release: pine_get_script_source reads a saved script over the REST API without touching the editor at all, which is what you want for comparing a saved script against a local file.