T-Bank MCP

T-Bank (Т-Банк) MCP — mobile banking API server for Claude Code, Codex, ChatGPT, and other MCP-capable agents.

The buttons register the tbank-mcp command — run pip install tbank-mcp first.

Features

• 90 tools: accounts, cards, documents, operations, grocery ordering, cinema and concert tickets, train and flight booking, hotel search, orders, transfers (including payment by bank requisites, from a scanned invoice QR), messenger, investments
• A skill for every vertical, entered through the tbank router skill: grocery order, tickets, travel, transfer, bill pay, cards & documents, messenger, budget analysis, invest advisor, login
• Pinned CA trust: system store + the Russian Trusted Root CA (Минцифры), which no OS ships and every *.t-bank-app.ru host needs — that is most of the hosts this MCP talks to. Shipped in tbank_mcp/ca/roots/, pinned by SHA-256. Leaf/intermediate rotation needs no action; a root rotation is a PEM drop into tbank_mcp/ca/roots/ (or TBANK_EXTRA_CA). The verify bundle is (re)generated from that material into ~/.local/share/tbank-mcp/bundle.pem; TBANK_CA_BUNDLE relocates it — it is a write target, not a curated input, so extra roots go in via TBANK_EXTRA_CA. Certificates are never learned from the network — see the header of tbank_mcp/tls.py.
• Grocery checkout: search → cart → order → pay (proven end-to-end)
• Secure login: password/PIN stay OUT of the LLM context (local CLI or env var)

Quick Install

As a Claude Code plugin (server + every skill in one step)

/plugin marketplace add icyberdeveloper/tbank-mcp
/plugin install tbank@tbank-mcp
/reload-plugins

There is no store to be admitted to — a marketplace is just a git repo with a .claude-plugin/marketplace.json, and anyone can host one.

The venv and the Python dependencies are created on the server's first start by bin/tbank-mcp: a plugin manifest cannot run install steps (install is not a field in the schema), so the launcher does it once and every later start goes straight to the server. Only the grocery checkout needs a browser, and 150 MB is not something to download behind your back — install it yourself if you want that flow:

~/.claude/plugins/cache/tbank-mcp/tbank/*/.venv/bin/python -m playwright install chromium

From PyPI

pip install tbank-mcp
tbank-mcp-login +7XXXXXXXXXX   # first login: SMS code + password in YOUR terminal, not the LLM
claude mcp add tbank -- tbank-mcp
tbank-mcp-skills               # optional: the Claude Code skills, into ~/.claude/skills

The tbank-mcp console script starts the stdio server; the pinned CA roots, the flows reference and the skills all ship inside the wheel, so it runs from anywhere. tbank-mcp-login is the same login CLI as the repo's login_cli.py — both write the session to the same file the server reads. tbank-mcp-skills installs the skills for Claude Code (default ~/.claude/skills, --target for elsewhere) and, unlike a plain cp, first removes stale copies under retired names — re-run it after upgrades. The grocery checkout browser stays opt-in here too: python -m playwright install chromium.

Manually (clone, no plugin)

git clone https://github.com/icyberdeveloper/tbank-mcp.git
cd tbank-mcp
python -m venv .venv && . .venv/bin/activate
pip install -e .
python -m playwright install chromium

# MCP server:
claude mcp add tbank -- ./.venv/bin/python -m tbank_mcp.server

# Skills — a COPY, so it does not follow the repo. Re-run after every pull, or
# the installed skills quietly describe an older version of these tools. The
# installer also removes stale copies under RENAMED skill names, which a plain
# cp never does — the agent would keep loading the stale one:
.venv/bin/tbank-mcp-skills

🔒 Login — the password never reaches the agent

The password and the PIN are secrets, and they are not put into the model's context. Logging in is done by a local script, or through an environment variable.

Option 1 (recommended): the local CLI

The script asks for the password itself, via getpass, so it is never echoed to the terminal and never passes through the agent. Its prompts are in Russian, as shown:

cd tbank-mcp

.venv/bin/python login_cli.py +7XXXXXXXXXX
# [1/3] login(+7XXXXXXXXXX) ...
#     SMS отправлена
# [2/3] SMS-код: ****                     ← the code from the SMS (hidden input)
# [3/3] Пароль (не отображается): ****    ← your password (hidden input)
#
# ✓ ГОТОВО! Сессия сохранена: ~/.local/share/tbank-mcp/session.json (права 0600).
#   MCP читает этот же файл — путь совпадает без ручной настройки.

Or with the password in the environment, for CI and scripts:

TBANK_PASSWORD="your-password" .venv/bin/python login_cli.py +7XXXXXXXXXX

Then start Claude Code. The agent picks up the saved session and works without the password, which never enters the LLM context.

Option 2: through the agent (convenient, but the LLM sees the password)

If you are content to hand the password to the agent:

> login(+7XXXXXXXXXX)
> [SMS code] 1234
> confirm_otp("1234")
> [bank asks password]
> confirm_password("YourPassword")

⚠️ Note: the password ends up in the model's context and in call logs. For an account you care about, use Option 1.

Both options need the SMS code typed in either way, so there is no unattended login. TBANK_PASSWORD (and TBANK_PIN) are read only by the login CLI (tbank-mcp-login / login_cli.py) — the env example above — and never by the MCP server or the LLM. TBANK_PHONE is not read anywhere: the phone is always a command-line argument.

Работа с MyT (рабочий календарь и парковка) переехала в отдельный MCP: tbank-myt. Другой аккаунт, другая сессия, свой login_cli.py — здесь их больше нет.

Other agents (Codex, ChatGPT, Hermes, OpenClaw)

{
  "mcpServers": {
    "tbank": {
      "command": "/path/to/tbank-mcp/.venv/bin/python",
      "args": ["-m", "tbank_mcp.server"],
      "cwd": "/path/to/tbank-mcp"
    }
  }
}

With a pip install tbank-mcp the whole entry shrinks to "command": "tbank-mcp" — no paths, no cwd.

Use with Cursor

Click the Install MCP Server button above, or open the Customize page from Cursor's sidebar to add the server, or create ~/.cursor/mcp.json (global) / .cursor/mcp.json (per-project):

{
  "mcpServers": {
    "tbank": { "command": "tbank-mcp" }
  }
}

Cursor renders MCP elicitation, so the payment-confirmation buttons work in chat.

Use with Cherry Studio

Settings → MCP → MCP Servers → Add: type stdio, command tbank-mcp, no arguments. Save, enable the server, wait for the healthy status.

Cherry Studio does not render MCP elicitation yet (open request CherryHQ/cherry-studio#9145), so reading works but the money tools refuse to execute there — by design, not by accident.

Use with Goose

CLI: goose configure → Add Extension → Command-line Extension, name tbank, command tbank-mcp. Desktop: sidebar → Extensions → Add custom extension, same values.

Goose renders elicitation in both Desktop and CLI; its confirmation forms time out after 5 minutes, so answer payment confirmations promptly.

Reading works in any MCP client. Paying needs a client that renders MCP elicitation — the money tools confirm the sum with a button the user presses («Перевести/Отмена», «Оплатить …?»), and a client without that capability is refused before anything is sent (grocery_checkout refuses there at any threshold; its dry_run=True preview, which creates nothing, still works). Hermes/Telegram and Claude Code (≥ 2.1.76) render it; Claude Desktop does not. See TBANK_CONFIRM_ABOVE under Security.

Tools

Each tool's docstring is the reference — this table is only a map of the surface. The docstrings, the skills and everything the tools print are in Russian: the bank is Russian and so is the person reading the answer.

get_data(section) covers dozens of read sections: subscriptions, credit_schedule, statements, loans, invest_accounts, pension, etc. (invest_portfolio is a tool of its own, not a section — see the docstring for the full list.)

Grocery tools (grocery_search, grocery_plan_order, grocery_add_to_cart, grocery_set_cart, grocery_cart, grocery_checkout) require app_id + point_id taken from grocery_stores() — there's no silent default store, so add/cart/checkout always operate on the same cart, instead of reporting an empty one right after something was added to a different store's.

Skills

Example requests

Ask in Russian — the tools answer in Russian. Everything below was run against the live bank.

Кино и афиша

Что идёт в кино сегодня?
Купи два билета на «Майкла» на завтра в Каро 11 около 20:00 в центре зала
Отмени заказ
Покажи последние 5 моих заказов

Деньги

Покажи мои счета
Переведи 10 рублей Алёне на +79991234567
Какие последние 5 операций?
Покажи реквизиты счёта

Продукты

Хочу оливье, собери корзину с минимальным КБЖУ
Хочу оливье, собери корзину с минимальной ценой
Хочу оливье, собери корзину из премиум продуктов
Найди самый дешёвый картофель за килограмм
Отмени заказ

Карты и документы

Покажи реквизиты основной карты
Какие лимиты по основной карте?
Покажи реквизиты моего паспорта
Когда истекает мой загранпаспорт?

Tests

No pytest — the tests are standalone scripts. Run them all:

.venv/bin/python tests/run_all.py            # every file, under a minute, offline
.venv/bin/python tests/run_all.py transfer   # only files matching "transfer"

Each runs in its own process, and the runner redirects the attempt/event journals to a temp directory so a test run never writes to ~/.local/share/tbank-mcp/.

Everything needed is in the repo: request contracts are pinned against scrubbed fixtures in tests/fixtures/ (real structure and protocol values, synthetic personal data), so the suite is meaningful on a clean clone. Where the original Burp capture is present the tests additionally check the fixtures have not drifted from it.

Security

• session.json — canonical path ~/.local/share/tbank-mcp/session.json (override with TBANK_SESSION), mode 0600, owner-only. It holds tokens. Both the login CLI and the MCP server read the same file, so there is nothing to configure. On start-up the MCP logs the path, size and permissions only — never a token or a cookie.
• Password / PIN — not in git, not in the code, and not in the LLM context if you use the login CLI (tbank-mcp-login / login_cli.py).
• No secrets in the repo. Two kinds of committed material look secret-adjacent and are not: tbank_mcp/ca/roots/*.pem are public CA root certificates, shipped on purpose and pinned by SHA-256 in tbank_mcp/tls.py; tests/fixtures/*.json are request contracts scrubbed from a real capture — real structure and protocol values, synthetic account, phone, address and device ids. The captures themselves are gitignored and never leave the machine.
• events.jsonl + attempts.jsonl — redacted diagnostics in ~/.local/share/tbank-mcp/. They carry step, http_status, blame, amount and order id, and never tokens, cookies, addresses, phone numbers, emails or account numbers. Safe to share while debugging; the diagnostics tool reads them.
• calls.jsonl — one line per tool call, so it can be seen how an agent uses this MCP: the tool, its arguments, the duration, and the FIRST LINE of the answer, which is what the agent actually read. Held to the same promise as the files above: arguments that are free text a person wrote (a chat message, a transfer note) or a credential are measured, never stored; long digit runs — account, card, order and payment ids — are replaced in the recorded line, both to keep them out and because the report groups by that line. The debug_report tool reads it. On by default; TBANK_TRACE=0 disables it, TBANK_TRACE_FILE moves it, and it rotates at 5 MB.
• TBANK_CONFIRM_ABOVE — the ruble threshold from which the paying tools that debit on the spot (transfer, transfer_requisites, pay_bill, ticket_pay, grocery_checkout, train_pay, flight_book) show the confirmation button — an MCP elicitation dialog («Перевести/Отмена», «Оплатить …?», «Оформить заказ на N ₽?») rendered by the client (default 0: every payment asks). It is a server-side setting, not a tool argument. Clients without elicitation are NOT waved through: at or above the threshold the tool refuses («ПЛАТЁЖ НЕ ВЫПОЛНЕН…») before anything is journalled or sent — no button, no payment. Hermes/Telegram and Claude Code (≥ 2.1.76) render elicitation; Claude Desktop does not (reads work there, paying does not). Below a positive threshold nothing is asked and the payment proceeds in any client — except grocery_checkout, which refuses a client without elicitation at any threshold: it is the one paying tool that must load the checkout page to learn its sum at all, and doing that means asking the store to hold a delivery slot, so it says no before doing that work rather than after. grocery_checkout(dry_run=True) — a preview that creates nothing — still works in any client.
• Device profile. Payments carry a 3DS/anti-fraud block whose device facts — screen size, locale, timezone, hardware model — default to the device the traffic was captured from. Override them with TBANK_DEVICE_SCREEN_HEIGHT / _WIDTH / TBANK_DEVICE_LANGUAGE / TBANK_DEVICE_TIMEZONE / TBANK_DEVICE_MODEL so your payments do not describe someone else's phone.
• Request-shape switches. Two divergences from the captured app are corrected behind env vars, so a rollback is one variable and no re-login (neither touches session.json):
  • TBANK_QUERY_PROFILE=legacy — restores sending wuid to every host and injecting vendor/client_version on every read. The app sends wuid only to www.tbank.ru under /api/common/, and the other two only on the OIDC authorize call, so the default is now the scoped form.
  • TBANK_ACCEPT_PROFILE — json (default, and today's behaviour byte-for-byte) | auto | a comma-separated host list. The app does not send application/json to its native hosts; that string is the Apple URL-loading default that appears when no Accept is set. The captured responses are application/json either way, so this is fidelity rather than a fix — but 63 templates share the busiest host and there is no staging environment, so it is OFF until driven live. Roll it out one host class at a time, cheapest first: webview/shortcuts/my-home (unreachable or trivial reads) → api-invest* (invest_accounts, invest_portfolio) → api.t-bank-app.ru starting with keepalive, whose Content-Type demonstrably becomes text/html while its body stays JSON → www.tbank.ru → the three lifestyle shelf paths. A regression has one signature: _unwrap raising HTTP_200 because the body no longer parses. Compare debug_report() before and after each step.
• Money tools (transfer, transfer_requisites, grocery_checkout, ticket_pay, pay_bill, train_pay, flight_book, confirm_payment) require confirmation of a specific amount — "buy it" is not a confirmation. That confirmation is the button the tool shows itself (elicitation, see TBANK_CONFIRM_ABOVE above) with the real total — the agent shows the details beforehand (recipient, requisites, cart, seats + fee) and does not ask «да/нет» in text; grocery_checkout quotes the final sum itself and charges exactly what the button named — and if that quote comes back unpriced (empty cart, a preview the store refused, no finite positive total), it returns the preview and charges nothing. A /v1/pay the bank holds at WAITING_CONFIRMATION is resumed with confirm_payment(attempt_id, otp) and reconciled with payment_status(attempt_id) — never by repeating the transfer, which would create a second pending payment.
• Tool annotations. Every tool declares what it does, in one table — TOOL_KINDS in tbank_mcp/server.py — and a tool missing from it raises at import rather than defaulting to anything. Three kinds: 67 are readOnlyHint: true and may run without a prompt; 15 write something that costs nothing (a cart, a booking, a message, an OTP, a token, a local file) and are marked destructiveHint: false; 8 debit an account — transfer, transfer_requisites, grocery_checkout, ticket_pay, pay_bill, train_pay, flight_book, confirm_payment — and are the only ones carrying destructiveHint, which is what makes the host prompt before running them (the sum itself is then confirmed by the tool's own elicitation button, see above). The line is drawn at money on purpose: a booking expires by itself and a cart line is a rewrite away, so confirming those is friction that teaches people to click through the one dialog that matters. The 15 writers are not marked read-only, because they do modify things and that flag states the opposite — if your client still prompts on them, allow them once in the client rather than changing what the server claims.

Disclaimer

For personal use with your own T-Bank account. Not affiliated with T-Bank.