Backtesting Arena

Backtesting Arena

Back to blog

The AI Copilot in Your Chart Is Reliable When It Measures and Invents When It Describes Itself

TradingView's chart copilot can connect to your own MCP servers. Tested on the Backtesting Arena connector: it answered market questions correctly and invented instructions about its own settings four times. Why that happens, and how to tell which answers to check.

Backtesting Arena·September 14, 2026·4 min read·0 views
The AI Copilot in Your Chart Is Reliable When It Measures and Invents When It Describes Itself

An AI assistant that reads your chart and can pull your data sounds like it knows what it is doing. The question that follows is not whether it is reliable, but where. We connected the Backtesting Arena connector to TradingView's chart copilot and watched for three hours. The answer is a dividing line, and it does not run where you would expect.

What the copilot is

TradingView put the "AI Chart Copilot" into public beta on 2 April 2026. Technically it is not a TradingView feature but a Chrome extension from the third party TradingView Remix, which TradingView promotes officially and plans to bring into the sidebar later. So "AI in the charts you already use" is a compression: the extension sits next to the charts, and only in Chromium browsers.

It reads the active chart, answers questions about symbol and timeframe, and lets you register any MCP server: a URL plus a key, and the model discovers the server's tools and calls them in conversation. The Arena connector exposes 89 such tools, from cycle score and macro regime to backtests and on-chain series. None of that can be read off a chart. If the two fit together, a Bitcoin chart answers questions that used to need a second window.

Four inventions

The connection failed at first because of a bug in the extension, which has been reported. What matters is what the copilot said while it was failing.

Twice it sent us to a page on tvremix.xyz to store the API key. The page exists, but it does not hold credentials for user-added servers. Once it named menu items that do not exist, "release tools to agent" and "expose tools to chat". And once it claimed at the same time that the server had returned a 401 error and that the tool was not available to it at all. Those cannot both be true, because you only get a 401 if a request actually went out.

All four statements had the same subject: its own state and its own interface.

What it got right

In the same session, asked for the Arena Pulse, a 0-to-100 heat score for the Bitcoin market that it could not retrieve because of the failing connection, it wrote that it would deliberately not name a value, because a guessed score would be worse than none.

That is exactly the right behaviour. And it sits directly next to four fabricated sets of instructions.

The difference is not the model but the question

Asked about the Pulse, the model has a tool. It calls the tool or it does not, and when the call fails, it notices. Asked about its own configuration, it has no tool. It cannot look up where a key is stored or which menu items exist. It knows from training what interfaces like this usually look like, and it produces a good guess in the register of a fact.

A model with a toolbox is dependable exactly as far as a tool reaches. Its own state is the one question it has no tool for, and that is precisely where the answer sounds most helpful.

In practice: check any instruction such an assistant gives about operating itself against the interface before you follow it. For market data the reverse holds. There the answer is as good as the tool behind it.

The objection

"That is a beta extension. Of course it hallucinates menus."

The menus are the symptom, not the finding. The same model, in the same session, refused to guess a number it had a tool for. The line runs between tool and no tool, not between beta and release.

Does it work now?

It does. The copilot finds the tools without being told their names; "where are we in the cycle?" is enough. When we checked the Pulse twice on 12 September, it read 44 and later 39, and the dashboard showed the same both times. Two readings that move together prove more than one that happens to match.

Two details are in no documentation. With 89 tools you have to pin the important ones, or the model will not find them. And the tool list is assembled when a conversation starts, so if you pin afterwards you need a fresh one.

Two findings concern our server, not the extension. The one-click login works in Claude because the connection there runs through a catalogue; a client that does not know the server has to find the login chain on its own, and that chain is missing on the Arena side. And the tool list can be read without an account, which may be sensible for discoverability but is better known in advance.

Beta software changes, and the extension will go native eventually. What lasts is the dividing line: tools make a model dependable, and self-report does not.


Not investment advice, not a recommendation, not a forecast — historical patterns are no guarantee.

Study the Past — Improve your Future 🥋

Try it yourself

Run the backtest with your own parameters and time ranges.

Run backtest →

More on this topic

Tools

Building for AI Agents: 10 Lessons From Running 73 MCP Tools

Backtesting Arenatradingstrategies.work

Your most thorough reader gets the least to read. Ten places where building for agents differs from building for people, with numbers from production.

AI agentsMethodology
Jul 29, 20261 min
Tools

Consensus Is Not an Edge: Why a Quorum of Correlated Simulations Isn't One

Backtesting Arenatradingstrategies.work

An AI swarm turned $1,000 into nearly a million? The number falls apart at its own source. Why consensus from correlated simulations is not an edge.

BacktestingMethodologyAI agents
Jul 17, 20261 min
★ FeaturedTools

Backtesting Arena Is Now a Claude Connector

Backtesting Arenatradingstrategies.work

Backtesting Arena can now be connected to claude.ai with a single click — 65 tools for Bitcoin cycle, on-chain data, macro regime and validated backtests. Instead of generic web-search answers, Claude pulls real, DSR-corrected and point-in-time data. Plus a new interactive tool: the Dip Decision calculator.

AI agentsBitcoinBacktesting
Jul 5, 20261 min
★ FeaturedTools

We Just Shipped an API That Charges $0.01 Per Call — In USDC, On-Chain, No Account

Backtesting Arenatradingstrategies.work

Phase 4 is live: the same Bitcoin cycle data, on-chain indicators, and aggregated strategy insights now reachable through three channels — REST, MCP, and x402 pay-per-call in USDC on Base. No account required for the third. An AI agent gets HTTP 402, signs a USDC authorization, retries, and has the data three seconds later. What's actually live, why we built it this way, and the Coinbase detour that cost us a day.

x402AI agentsStablecoins
May 23, 20261 min
📬

Don't miss new blog posts

One short email per new post — strategies, backtests, market analysis. No spam, unsubscribe with one click anytime.

By subscribing you accept our privacy policy. We use Resend for delivery. Double opt-in confirmation required.

Comments (0)

Join free to post comments.

Sign up →

No comments yet. Be the first!