bd-finance-mcp

Bangladesh's financial markets, available to your AI assistant. Both stock exchanges, every listed share, and what the taka is worth today.

An MCP server built with FastMCP.

⚠️ Not investment advice. This reports publicly published prices. It does not analyse, recommend, or predict. Figures are last-traded values and may lag the live market.

Unofficial. Not affiliated with DSE, CSE, or Bangladesh Bank.

Tools

company_profile turns a bare price into something you can reason about — GP trades at a P/E of 11.44 in Telecommunication with a 326,907mn taka market cap, BRAC Bank at 6.14 in Banking. Prices alone don't tell you that.

compare_exchanges is the one you can't easily do yourself — most Bangladeshi companies list on both markets, and the prices genuinely differ. GP was trading at 240.60 on Dhaka and 241.50 on Chittagong while this was being written.

sector_performance answers the question a price list can't: how did banks do today? It aggregates every share in a sector — advancers, decliners, average move, best and worst — from one query. Partial names work, so "pharma" finds "Pharmaceuticals & Chemicals".

Prices tell you a share moved; finance_news tells you why.

Every market response carries a market block saying whether DSE is currently trading (Sunday–Thursday, 10:00–14:30 Asia/Dhaka) and whether the figures are live or the last completed session. Outside session hours, "today's gainers" means the previous close — the response says so rather than leaving you to assume.

Related MCP server: Financial Modeling Prep MCP Server

Sources

Dhaka Tribune publishes no business-only feed, so its articles are general newsroom output. finance_news leaves it out by default (business_only=True) rather than passing student politics off as market coverage — name it in sources, or set business_only=False, to include it.

Bonik Barta is not available. It is a single-page app: every path, including /api/v1/, returns the same HTML shell with HTTP 200. There is no feed or API behind it. A status code alone is not proof a source works.

New Age is not available. No working RSS feed exists at any conventional path.

Financial Times links are paywalled. Headlines and summaries come through the public feed; the articles themselves require a subscription.

No API key, login, or paid data feed.

Not included: Bangladesh Bank. Its econdata pages sit behind a CAPTCHA challenge. That is an explicit request not to automate, and this project respects it rather than working around it. Use exchange_rate for reference rates instead — but note those are mid-market, not BB's official rate, and not what a bank will actually pay you.

Not included: gold prices. BAJUS renders them client-side; nothing is in the HTML to parse.

Not included: the DSEX index. It appears only in DSE's navigation menus — the live value is injected by JavaScript, so there is nothing to read from the served HTML. market_summary gives you breadth and turnover instead, which describes the session without pretending to quote an index it cannot see.

Install

Requires uv and Python 3.12+.

git clone https://github.com/Claudefarid/bd-finance-mcp.git
cd bd-finance-mcp
uv sync
uv run fastmcp install claude-code server.py:mcp

Restart your client, then ask "Compare GP's price on both Bangladeshi exchanges" or "What's the taka doing against the riyal?"

Verify with uv run python test_server.py.

Six things that will bite you if you build this yourself

Each of these produces wrong answers rather than errors — the dangerous kind of bug.

1. The two exchanges disagree about untraded shares.

DSE reports a share that did not trade with price 0. CSE carries yesterday's price forward. Compute percent change naively and DSE hands you a list of −100% "crashes" that are really just shares nobody bought. Trade count is the only signal that works on both, so it decides whether a percentage is meaningful here.

This matters more on CSE than you would guess: 187 of its 387 listed shares did not trade on the day this was written. Nearly half the market. Get this wrong and CSE data is worthless.

2. dsebd.org serves an incomplete TLS certificate chain.

Its leaf certificate is signed by a Sectigo intermediate the server never sends. Browsers and curl recover by fetching it via AIA; Python's certifi bundle cannot, so httpx fails with CERTIFICATE_VERIFY_FAILED.

The fix is not verify=False — that disables verification altogether, which is a bad trade anywhere and a worse one for financial data. Verify against the OS trust store instead:

import ssl, truststore
ctx = truststore.SSLContext(ssl.PROTOCOL_TLS_CLIENT)
httpx.AsyncClient(verify=ctx)

3. DSE's P/E table is one column per trading day, and the newest is last.

The ratio table sits under a Particulars header whose columns are dates. Read the first column — the obvious choice — and you report a P/E from a week ago as today's. company_profile takes the last populated column and returns the date alongside it, so the figure is always checkable.

There is a second trap in the same table: the page carries several Particulars headers, and the first one reads "Unaudited / Audited" rather than dates. Matching on width picks the right one.

A missing P/E is not a parsing failure, either. DSE prints - across the row for companies without meaningful earnings — BEXIMCO, for one — so None there is the honest answer.

4. The Daily Star's feed puts HTML inside its <title> tags.

Its headlines arrive wrapped in an unescaped anchor:

<title><a href="/business/news/...">7,896 customers withdraw Tk 327 crore</a></title>

ElementTree parses that anchor as a child element, so findtext("title") returns an empty string — not an error, just nothing. Every Daily Star headline came back blank, and because an empty headline matches no keyword, the paper looked like it was ignoring every topic when it was simply never being read. "".join(element.itertext()) collects descendant text and fixes it.

5. Substring matching puts the IT sector inside "Financial Institutions".

The obvious way to resolve a sector name is query in name.lower(). Ask for "IT" and you match both IT Sector and Financial Inst**it**utions — so the most natural query for the sector fails as ambiguous. Matching runs in tiers instead: exact, then prefix, then word-boundary, then substring, taking the first tier that resolves to one sector.

6. Mid-market rates are not remittance rates.

exchange_rate returns reference rates. Banks and exchange houses apply their own spread, so what actually lands in a recipient's account is lower. The tool says so in its own description, so a model relying on it does not quietly present these as the rate you'd receive.

Please be considerate

Each call is a single page fetch, and requests identify themselves by User-Agent. Don't poll in a tight loop.

License

MIT — see LICENSE. Covers this code only, not the exchanges' data.