How canli-mcp works, and how it is tested

On a question

What happens when you ask

1. The model sees a short list. 5 tools instead of 388, which costs 2,524 tokens of context on each request. All 388 listed side by side across the seven servers would cost 70,862.
2. It searches. find_tool takes plain words, such as "who owns this company", and returns the tools that fit. describe_tool gives one tool's arguments.
3. It runs them. run_tool runs one tool or up to 25 at once, on your machine. Calculations stay there. A data tool sends only the company, series or symbol you asked about, to the source named below.
4. It gets an answer it can check. Figures come with the filing, sentence or table row they were read from, and a hash of the document received, so anyone can fetch it again and compare.
5. You can see it. Where the app supports it, the result opens as a chart or table you can use in the chat, and the model reads a summary of it.

The data

Where the numbers come from

From the body that publishes them, with no key and no account: the SEC for filings, insiders and funds; the Treasury, the Fed, FRED, the ECB, the Bank of England, Japan's Ministry of Finance and the Bank of Canada for rates; the World Bank, the OECD and the BIS for the world economy; the FDIC for banks; the CFPB, the FDA, NHTSA, the Treasury's sanctions office and the Federal Register for public records; Cboe for delayed options data; Yahoo Finance for prices. Prices and options are not the original exchange feed.

The local server fetches from these sources directly. The exceptions are point-in-time SEC fundamentals and the research record, which come from copies published on canlicapital.com; those requests carry the company, measure or search words, never your own data. The hosted connector at https://canlicapital.com/mcp/canli runs the public-data and computing tools for people who cannot install anything; the privacy policy says what it sees.

Being wrong out loud

It refuses instead of guessing

• In the public-record tools, a name matches only when every word of it is in the record, so "Newell Brands" never returns a different Newell.
• An ISIN, CUSIP or LEI with a wrong check digit is refused before any request is sent.
• A question with a false premise, such as a year a company did not report, gets a refusal and the reason, not a figure.
• verify_claims sets a draft's numbers against the filing and says which are confirmed, which are contradicted and what the wrong one matches.

Testing

How it is tested, failures included

The test that finds real problems is a new one. So each round takes releases the tools have never seen, writes down the right answers from the text first, and only then runs the tools. The first score stays on the record. As of 2026-10-30, on canli-markets-mcp 0.8.2, the 119 analyst questions all pass. Run live against SEC EDGAR. Tool level: no model reads the answers. The expected figures come from sources the tools do not read: SEC's XBRL API, a separate table parser, and the releases' own words.

A score on a set that has already led to fixes no longer says much about new questions, which is why the next set is written first. Other checks: a cross-check of 21 figures. European filings, central-bank rates, odds and crypto set against sources those tools do not read. A live check of the data tools runs every night, so a source that changes its format shows up the next morning. Before a release, a sweep sends hostile arguments to the tools. Against other servers, the head-to-head benchmark gives the same model the same questions; we wrote those questions, so its page lists the caveats, and an outside benchmark is still to come.

What testing has caught

• earnings_release could read a company's preliminary 8-K instead of its results (FMC, October 2026). Fixed in 0.8.0.
• A guidance of "negative $350 million" came back positive, and interest expense of $230 million in an assumptions column was read as earnings per share. Fixed in 0.8.0.
• "Silver price" once returned US inflation, because a plain-word search matched part of the words. Now every specific word must match.
• An earlier loose scoring rule passed two questions the strict rule fails, because it accepted the numbers in the quote. The first-run score above uses the strict rule.

Limits

What it cannot do

• It gives no analyst estimates or price targets, and no corporate bond prices: no public source offers them without a key.
• Earnings dates ahead are estimates from a company's past dates, and are labelled so.
• It reads what companies and agencies published. If a filing is wrong, so is the figure, and a parser can misread an unusual table, so the quote beside a figure is the authority.
• A sanctions search is not a compliance screening, and complaint counts are not adjusted for a company's size.
• It places no real orders. Trading tools work on a paper account only, and it does not give investment advice.

See every tool or the questions it answers.