TradingView + Claude via MCP, Desk-Tested: What It Can and Can't Do
We ran a TradingView MCP server with Claude: the obvious pip install grabs another author's package, and a US ETF came back tagged as a crypto exchange.
Connecting Claude to TradingView through MCP does work today, but not as one thing: the projects using that name split into a read-only market-data server you can install from PyPI in a minute, and a bridge into the TradingView desktop app that drives undocumented internal APIs. Neither is a TradingView product. The one we installed and ran returns fast live quotes, quietly defaults large parts of its surface to a crypto exchange, and — the finding worth acting on before anything else — is not the package you get from the obvious install command.
This page is about what the tooling does mechanically. It contains no strategy, no signals and no results.
Before you install anything: the name is squatted
The obvious install command for the trending repository is the wrong one. Running pip install tradingview-mcp — the name that matches the GitHub repository everyone is linking to — installs a completely different author’s package: an “Advanced AI Trading Intelligence Framework” from a different repository, advertising “73 technical analysis tools” against the trending project’s “30+”. Those two strings are the packages’ own PyPI summary lines; the trending project’s long description on the same page puts its own count at 37. The trending repository’s actual package is published under a different name, tradingview-mcp-server.
We verified the provenance through PyPI’s JSON API rather than by eye, reading each package’s declared project_urls back to its source repository. We did not execute the squatter’s code, so we are making no claim about what it does — it may be entirely benign. The claim is narrower and sufficient: the package you get from the obvious command is not the package from the repository you read the stars on.
The check takes about five seconds and generalises past MCP entirely:
# Does this PyPI name actually belong to the repo you evaluated?
curl -s https://pypi.org/pypi/tradingview-mcp/json \
| python3 -c "import sys,json; print(json.load(sys.stdin)['info']['project_urls'])"If the URLs that come back do not point at the repository you read, stop. A star count belongs to a repository; a name belongs to whoever registered it first.
What “TradingView MCP” actually refers to
When we sat down to evaluate this category on 2026-08-29, the name covered two projects with almost nothing in common:
- A data server (
atilaahmettaner/tradingview-mcp), shipping on PyPI astradingview-mcp-server, version 0.8.1 at the time of our session. It exposes quote, indicator and scanner tools. This is the one we installed and ran. - A desktop bridge (
tradesdontlie/tradingview-mcp) that attaches to the TradingView desktop application over Chrome DevTools Protocol to drive the real product — your charts, your layouts, your account. It needs the desktop app installed and a subscription behind it.
Our session log records the data server at roughly 3,500 GitHub stars on 2026-08-29. We recorded no reading for the desktop bridge, so this page makes no popularity comparison, and none is needed: star counts move daily and are nobody’s quality signal. The reason to name both projects is narrower — write-ups about “the TradingView MCP” are not all pointing at the same software.
Both are third-party community projects. Neither is published by TradingView, and that distinction carries all the way through: nothing either one does is covered by a vendor support commitment, and nothing either one returns is a vendor-warranted figure.
We did not test the desktop bridge. There is no TradingView desktop app on this machine, so every capability it advertises is untested here and this page will not grade it. Worth repeating is the caution its own README carries: it drives “undocumented internal TradingView APIs” that “can change or break without notice.” That is the author being honest, and the correct mental model for the whole category.
What we ran, and what came back
We installed the correct package (tradingview-mcp-server 0.8.1) into a fresh Python 3.12 virtual environment and ran four representative tools on 2026-08-29. Verbatim from our session record:
| Tool called | Time | What came back | Defect |
|---|---|---|---|
coin_analysis("AMEX:SPY") |
0.21s | Live quote: price 769.34, change −0.034%, volume — no API key required | "exchange": "kucoin" on an AMEX-listed US ETF |
analyze_coin("BINANCE:BTCUSDT") |
— | TypeError: missing required positional args exchange, timeframe |
Sibling tool requires arguments its neighbour does not |
bollinger_scan() |
0.46s | KuCoin crypto pairs (e.g. KUCOIN:ARPAUSDT) with SMA20, Bollinger bands and RSI computed |
Called bare, silently defaults to a KuCoin universe |
bitcoin_market_pulse() |
0.16s | "source": "CoinGecko" |
Not TradingView data; the server aggregates multiple upstreams |
Four things follow from that table, in descending order of how much they should change your behaviour.
It is genuinely fast, and free. A fifth of a second for a structured quote with no credential is the compelling part for an LLM workflow.
The metadata is wrong in a way a model cannot detect. The SPY price was plausible; the venue attribution was not. Anything consuming that field downstream has no way to know.
The tool surface is inconsistent between siblings. A TypeError is a papercut for a human and a genuine failure mode for an agent, which calls a tool the way the schema reads and cannot reason about the inconsistency.
The scanner picks a universe for you. If you assumed a scan meant US equities because the project has “tradingview” in its name, you would be reading the right-looking output of the wrong question. Per its own README the server also draws on Yahoo Finance — a multi-source aggregator wearing a TradingView name, which is fine, and worth knowing before you cite one of its numbers as “from TradingView.”
What this means if you want an LLM near your charts
If you want a model to fetch quotes and compute standard indicators, the data server does that, quickly, for free, without credentials. Treat the metadata as unreliable, name the exchange and timeframe explicitly on every call rather than accepting defaults, and check which upstream source each tool actually uses before you rely on a figure.
If you want a model to write the strategy itself, that is a different workflow with its own failure modes — what a prompt-to-Pine-Script loop cannot measure about its own output is in the Claude trading bot playbook.
If you want a model to drive the real TradingView product — your saved layouts, your indicators, your alerts — that is the desktop-bridge project’s territory, it needs the desktop app and a paid account, and we have not tested it. If you go that way, go in knowing it depends on internal endpoints its own author says can break without notice.
If what you actually want is the platform’s own features, the built-in tools are the supported path — including the AI Screener, which turns a plain-English request into screener filters inside the product itself. That comparison is the useful one: an official feature has documented boundaries and a support channel; a community MCP server has a GitHub issues tab. The desktop-bridge route needs a paid account before any of it applies — TradingView. Whether that tier is worth it is a separate question we work through in free versus paid.
The open questions we are not answering yet
Three things we deliberately did not assert from a single session, listed because a review that pretends to certainty it does not have is worth less than one that draws the line:
- Licensing and terms-of-service posture. Unofficial access to a vendor’s endpoints raises a question about that vendor’s terms that we have not resolved. Read them yourself before building anything durable.
- Reliability across vendor changes. We ran four tools once. Nothing here says what happens the week an upstream endpoint moves.
- Indicator accuracy. The scanner computed Bollinger bands and RSI, and we have not reproduced those numbers. The obstacle is specific rather than lazy: called bare,
bollinger_scan()returned KuCoin crypto pairs, and this desk holds no independent bar history for KuCoin symbols to reconcile against — our own indicator code runs on the equity and ETF bars behind our regime read method. The closing condition is stated: re-run the scanner with an explicit US-equity symbol and timeframe, recompute SMA20, the Bollinger bands and RSI from our own bars over the same window, and publish both sets side by side. Until that runs, we are not vouching for the computed figures.
That is the discipline we apply to every tool crossing this desk, per our trial protocol: a number a tool hands you is a claim until you have reproduced it.
The general lesson, beyond this one server
An MCP server is a credential and capability surface attached to a model. The interesting question is never how many tools it advertises — the squatting package advertised 73, and the legitimate one’s own long description says 37. It is: who wrote it, what does it reach, what happens when it is wrong, and how would you know? The same first question is worth asking of TradingView’s AI Chart Copilot extension, where the name on the announcement and the name on the listing you install do not match.
A wrong exchange label is the server being wrong in a way the model cannot detect. A silently defaulted universe is the server answering a question you did not ask. Those are cheap failures when a human reads the output and expensive ones when a model acts on it — which is exactly the shape of every operational problem we have recorded on our own automated rigs.
Frequently asked questions
Is there an official TradingView MCP server?
Not among the projects we evaluated — both are third-party, community-maintained, and not published by TradingView. Treat anything in this space as unofficial unless the vendor says otherwise on its own site.
Which package should I install for the TradingView MCP server?
For the trending data-server repository, the published package name differs from the repository name: pip install tradingview-mcp fetches a different author’s project. Check the PyPI page’s project_urls and confirm it points at the repository you evaluated before running any install command.
Do I need a TradingView subscription to use an MCP server?
For the data server we tested, no — it ran with no API key and no account. For the desktop-bridge project, yes: it requires the TradingView desktop application and a subscription behind it.
Can a TradingView MCP server place trades?
Nothing we ran does. The tools we exercised are read-only market data and indicator computation. Any project that could touch an account would be a different risk category entirely — the permission bands we take apart in broker MCP servers — and would need least-privilege credentials, paper endpoints and kill conditions specified in advance.
Is the data accurate?
The quote we pulled was plausible; the exchange metadata attached to it was wrong, and the computed indicators are unverified against our own until the check described above runs. Do not cite figures from an aggregator as vendor figures.
Session details: tradingview-mcp-server 0.8.1, fresh virtual environment, Python 3.12, tools run 2026-08-29; the ~3,500 star count for that repository was read the same session, and star counts move daily. Package provenance and both packages’ declared tool counts were read from the PyPI JSON API during that same 2026-08-29 session, and both packages’ figures are quoted from it. Package metadata changes without notice — run the check below rather than trusting these figures.
Disclosure: this page contains an affiliate link. If you subscribe through it, this site may earn a commission at no additional cost to you. It does not change anything stated above — the third-party projects described here pay us nothing, and our findings on them are unflattering in places.
This is educational content about software tooling, not financial advice. No positions, signals, account values or trading results appear on this page.