wise-mcp
Provides secure, scope-gated access to the Wise API, enabling read-only retrieval of profiles, balances and jars, bank details, activity feeds, statements, transfers, recipients, exchange rates, invoices, and webhooks, with opt-in write capabilities for invoices, webhooks, recipients, quotes, transfers, batch transfers, and balance moves, including confirmation flows, amount caps, allowlists, and audit logging.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
wise-mcp
A secure, scope-gated Model Context Protocol server for the Wise API. It gives an AI assistant (Claude Desktop, Claude Code, Cursor, Hermes...) access to your Wise account through a deny-by-default permission model:
Read-only by default. With no configuration, the server can only read: profiles, balances and jars, bank details (IBAN / sort code / ACH), the activity feed (card payments included), statements, transfers, recipients, exchange rates, invoices and webhooks.
Writes are opt-in, per capability (
WISE_ENABLE=invoices,...). A hard deny list (WISE_DENY=transfers,...) always wins, andWISE_READ_ONLY=trueturns every write off.Disabled tools are not registered at all, so the model never sees them.
Defense in depth: the HTTP client enforces the same policy on every request (method + path allowlist) before any network call, so even a tool registered by mistake cannot reach a denied endpoint.
Money-moving operations are two-step (
prepare_*thenexecute_*) with a single-use, 5-minute, HMAC-bound confirmation token, optional amount caps and recipient/profile allowlists.Every write attempt is audited (allowed or blocked) to a local JSONL file, without secrets or full account numbers.
This is an unofficial project, not affiliated with Wise. Enabling write scopes lets an AI agent act on a real financial account. Read thesecurity model before enabling anything.
Contents
Related MCP server: ictexam-mcp
Quickstart
Requires Python 3.12+ and a Wise personal API token.
# Run straight from the repository with uv (no install)
uvx --from git+https://github.com/NathanJ60/wise-mcp wise-mcp --print-policy
# Or from a local clone
git clone https://github.com/NathanJ60/wise-mcp && cd wise-mcp
uv sync
WISE_API_TOKEN=... uv run wise-mcp --print-policy # show effective permissions, then exit
WISE_API_TOKEN=... uv run wise-mcp # start the stdio server
# Or with pip
pip install git+https://github.com/NathanJ60/wise-mcp
wise-mcp --versionThe server speaks MCP over stdio only. It never opens a network port. Logs go to stderr.
Client configuration
All examples below are read-only. Add WISE_ENABLE / WISE_DENY to grant more
(see recommended configurations).
Claude Desktop
claude_desktop_config.json (Settings > Developer > Edit config):
{
"mcpServers": {
"wise": {
"command": "uvx",
"args": ["--from", "git+https://github.com/NathanJ60/wise-mcp", "wise-mcp"],
"env": {
"WISE_API_TOKEN": "your-token",
"WISE_READ_ONLY": "true"
}
}
}
}Claude Code
claude mcp add wise --scope user \
--env WISE_API_TOKEN=your-token --env WISE_READ_ONLY=true \
-- uvx --from git+https://github.com/NathanJ60/wise-mcp wise-mcpCursor
~/.cursor/mcp.json (or .cursor/mcp.json in a project):
{
"mcpServers": {
"wise": {
"command": "uvx",
"args": ["--from", "git+https://github.com/NathanJ60/wise-mcp", "wise-mcp"],
"env": { "WISE_API_TOKEN": "your-token", "WISE_READ_ONLY": "true" }
}
}
}Hermes Agent
hermes mcp add wise --command uvx \
--env WISE_API_TOKEN=your-token WISE_READ_ONLY=true \
--args --from git+https://github.com/NathanJ60/wise-mcp wise-mcpPrefer referencing the token from your secret store or .env rather than pasting it into a config
file that might be shared or committed.
Configuration reference
All configuration comes from environment variables. Invalid values (e.g. an unknown scope name) stop the server at startup instead of being ignored.
Variable | Default | Meaning |
| (required) | Wise personal API token. Never logged or returned. |
| (empty) | Comma-separated write scopes to enable. |
| (empty) | Scopes that are hard-blocked. Always wins. Accepts |
|
|
|
| (all) | Comma-separated profile ids the server may touch (path and query are checked). |
| (all) | Recipient ids allowed in quotes, transfers and batch transfers. |
| (none) | Cap per money movement: |
| (none) | Same format, summed per UTC day and currency (persisted locally). |
|
| Audit log path. |
|
| Audit log and daily-cap state. |
| (none) | RSA private key (PEM) used to sign SCA challenges. |
|
|
|
|
| Per-request timeout (1 to 120). |
|
| Confirmation token lifetime (30 to 900). |
|
| Log level (stderr only). |
wise-mcp --print-policy (add --json for machine output) shows the effective scopes, the tools
that are registered or disabled, the allowlists, the caps and the audit log path. It never prints the
token. The get_permissions tool returns the same information to the assistant.
Security model
Scopes
Scope | Default | Moves money? | What it allows |
| on | no | Every |
| off | no (receives) | Create and update invoice drafts, publish them, cancel (invalidate) them. |
| off | no | Create and delete profile webhook subscriptions (HTTPS URLs only). |
| off | no | Create and deactivate recipient (payee) accounts. |
| off | no | Create and update quotes (needed for transfers and conversions). |
| off | yes | Validate, create, cancel and fund transfers. Denying it also denies |
| off | yes | Create, complete, cancel and fund batch groups. |
| off | yes (between your own balances) | Open/close balances and jars, convert currencies, move money between balances/jars. |
Effective scopes are computed once at startup:
effective = ({read} ∪ WISE_ENABLE) − WISE_DENY − (all write scopes, if WISE_READ_ONLY=true)WISE_DENY cannot be overridden by anything: not by WISE_ENABLE, not by a tool argument, not by
the conversation. The server has no tool that changes its own permissions.
Layers
Registration. Tools of a disabled scope are never registered with the MCP server: the client cannot list or call them.
Request allowlist. The single HTTP client (
WiseClient) checks every request against an explicit table of(method, path regex, scope)routes. Unknown paths, absolute URLs, traversal or encoded characters are refused. The deny list is checked first. This happens before any network I/O. The test suite proves that withWISE_DENY=transfers, no code path canPOST/PUT/PATCHany transfer, funding or batch endpoint, even with every tool force-registered.Allowlists.
WISE_ALLOWED_PROFILE_IDSis enforced at request level (path and query).WISE_ALLOWED_RECIPIENT_IDSis enforced for quotes, transfer creation and funding (the recipient of the fetched transfer is checked again at funding time).Two-step confirmation for operations that move money (
execute_fund_transfer,execute_fund_batch_group,execute_balance_conversion,execute_balance_move):prepare_*is read-only. It fetches the real amount from Wise (transfer, batch, quote or balances), checks caps and allowlists and returns a human-readable summary plus aconfirmation_token.The token is HMAC-SHA256 signed with a random per-process key, expires after 5 minutes, is single use, and is bound to the action and the exact parameters (including the amount). Changing anything, or calling a different
execute_*, invalidates it.execute_*re-fetches the data, re-checks caps and allowlists, then calls Wise with an idempotency key derived from the token.
Amount caps.
WISE_MAX_AMOUNT_PER_TXandWISE_MAX_AMOUNT_PER_DAYare enforced locally at prepare and execute time. When a cap is per currency, other currencies are refused (fail closed).Explicit
confirm=trueon other irreversible writes: publishing or cancelling an invoice, deleting a webhook, deactivating a recipient, cancelling a transfer, closing a balance, completing or cancelling a batch.Audit log. Every write attempt is appended to a JSONL file:
write_blocked(with the reason),write_attempt,write_ok/write_failed,confirmation_issued. Bodies are sanitized: tokens and secrets are removed, IBANs and long digit sequences are masked (****3000). If the audit log cannot be written, the write is not sent.Transport hygiene. Fixed base URL, no redirects, timeouts on every request, retries only for
GETon 429/5xx, token only fromWISE_API_TOKENand redacted from every error message.Tool annotations. Every tool declares
readOnlyHint,destructiveHint,idempotentHintandopenWorldHint, so clients that honour them can ask for approval on writes.
Tool annotations and the MCP client's own approval prompts are useful, but they are hints. The guarantees come from the server-side layers above.
Recommended configurations
Read-only (safest):
WISE_READ_ONLY=trueNathan / safe default: invoices, webhooks, recipients, quotes and moves between your own balances (with two-step confirmation), but never sending money out:
WISE_ENABLE=invoices,webhooks,recipients,quotes,balance_moves
WISE_DENY=transfers,batch_transfers
# optional hardening
WISE_ALLOWED_PROFILE_IDS=12345678
WISE_MAX_AMOUNT_PER_TX=EUR=2000,USD=2000,GBP=2000
WISE_MAX_AMOUNT_PER_DAY=EUR=5000,USD=5000,GBP=5000Invoicing only:
WISE_ENABLE=invoices
WISE_DENY=transfers,batch_transfers,balance_moves,recipients,quotesFull access (only if you really mean it): enable transfers explicitly, and set recipient
allowlists and caps:
WISE_ENABLE=quotes,transfers
WISE_ALLOWED_RECIPIENT_IDS=111,222
WISE_MAX_AMOUNT_PER_TX=EUR=500
WISE_MAX_AMOUNT_PER_DAY=EUR=1000Also restrict the token itself (see below): a Read only Wise token makes every write fail on Wise's side regardless of this server's configuration.
Tools
56 tools in total. With the default configuration only the 30 read tools are registered.
Tool | Scope | Wise endpoint | Read-only | Destructive |
| read | local (no API call) | ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| read |
| ✅ | |
| invoices |
| ||
| invoices |
| ⚠️ | |
| invoices |
| ⚠️ | |
| invoices |
| ⚠️ | |
| webhooks |
| ||
| webhooks |
| ⚠️ | |
| recipients |
| ||
| recipients |
| ⚠️ | |
| quotes |
| ||
| quotes |
| ||
| transfers |
| ✅ | |
| transfers |
| ||
| transfers |
| ⚠️ | |
| transfers |
| ✅ | |
| transfers |
| ⚠️ | |
| batch_transfers |
| ||
| batch_transfers |
| ||
| batch_transfers |
| ⚠️ | |
| batch_transfers |
| ✅ | |
| batch_transfers |
| ⚠️ | |
| balance_moves |
| ||
| balance_moves |
| ⚠️ | |
| balance_moves |
| ✅ | |
| balance_moves |
| ⚠️ | |
| balance_moves |
| ✅ | |
| balance_moves |
| ⚠️ |
Notes:
get_account_detailsseparates ACTIVE currencies (real, usable bank coordinates) from AVAILABLE ones (can be opened, numbers masked by Wise, must not be shared).list_activitiesis the feed shown in the Wise app: card payments, transfers, conversions, deposits, fees. Filter withmonetary_resource_type,status,since,until; paginate withcursor.Invoices are a business-profile feature. Drafts use
selectedPaymentMethods["BANK_TRANSFER"]by default (PAY_WITH_WISEis also accepted;ACCOUNT_DETAILSis rejected by Wise). A draft is never sent to anyone untilpublish_invoice.Every tool returns structured JSON (MCP
structuredContent) with an output schema. List results are wrapped as{"count": n, "items": [...]}.
Creating a Wise API token
In the Wise web app: Settings > Developer tools > API tokens > Add new token. Personal tokens are available for personal and business accounts; business accounts may require an admin.
Read only: enough for every
readtool. Recommended if you do not need writes: Wise itself then refuses every write, whatever this server is configured to do.Full access: required for invoices, webhooks, recipients, quotes, transfers and balance moves. Combine it with
WISE_ENABLE/WISE_DENYso the server only exposes what you need.
Treat the token like a password. Store it in your OS keychain, secret manager or a .env file that
is never committed (see .env.example), and revoke it from the same Wise page if it
leaks.
Strong Customer Authentication (SCA)
Some endpoints (notably balance statements and funding) are SCA-protected for profiles registered
outside the US, AU, NZ, SG, CA and MY (EU/UK PSD2 rules). Wise then answers 403 with a one-time
token in x-2fa-approval. To let the server complete the challenge:
openssl genrsa -out wise-sca-private.pem 2048
openssl rsa -pubout -in wise-sca-private.pem -out wise-sca-public.pemUpload wise-sca-public.pem in Wise (Settings > Developer tools > API tokens > Manage public
keys) and set WISE_SCA_PRIVATE_KEY_PATH=/path/to/wise-sca-private.pem. The server signs the
one-time token (SHA256withRSA) and retries once with x-2fa-approval + X-Signature. Without a key
you get an explicit error explaining what to do. For statements the challenge only needs to be
completed once every 90 days. See
Wise: SCA and 2FA.
Sandbox
Create a sandbox account and token on sandbox.transferwise.tech,
then run with WISE_SANDBOX=true. The base URL becomes https://api.wise-sandbox.com. The base URL
can only be one of these two values; it cannot be overridden.
Limitations
What a personal token cannot do (checked against the official docs and/or production):
Cards (list, details, card transactions, card controls):
UserToken(partner OAuth) only; production returns403. Card payments are still visible throughlist_activities.Direct debit accounts:
UserTokenonly; production returns401.Funding transfers / batches from EU/UK profiles may be refused for personal tokens under PSD2 even with SCA; use the Wise app to fund in that case.
Quotes, recipients and webhooks are documented by Wise as
UserTokenendpoints. They are included because personal tokens are accepted in practice for several of them, but Wise may refuse some calls; you will then get the Wise error message.Receipts and statements in PDF/CSV/XLSX and other binary downloads are not exposed (JSON only).
Payment request list filters: Wise does not publicly document the acquiring API;
statusis passed through on a best-effort basis.Account opening, KYC, partner/OAuth flows, card ordering and spend limits are out of scope.
Development
uv sync
uv run ruff check src tests && uv run ruff format --check src tests
uv run mypy
uv run pytestTests never touch the network: Wise is mocked with respx, tools
are exercised through the official in-process MCP client, and one test spawns the real stdio server.
Dependencies are pinned in uv.lock. Runtime dependencies: mcp, httpx, pydantic and
cryptography (SCA signing). No telemetry: the only host the server talks to is the Wise API.
See SECURITY.md for vulnerability reporting and CHANGELOG.md.
License
MIT © Nathan Jacas
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
QuickBooks Online for AI assistants, with per-user, per-company permissions.
1Connect authorized ad accounts to AI assistants for audits, reporting, and approval-gated changes.
Connect AI to store orders, products and inventory with scoped access and human approvals.
Connect brokerage accounts to AI assistants: holdings, orders, asset lookup, and approved trading.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnables AI assistants to interact with Basecamp projects, messages, to-dos, and Card Tables, with read-only access by default and write operations requiring explicit approval.MIT

ictexam-mcpofficial
AlicenseAqualityBmaintenanceEnables AI assistants to read ICTExam exams, gradebooks, and per-question item analysis, and optionally parse question papers and publish or unpublish exams when write access is enabled.642 npm17MIT- AlicenseAqualityCmaintenanceEnables AI assistants to access Saxo Bank portfolio data, account balances, positions, orders, transactions, and market data such as instruments, quotes, and charts. It is read-only by default, with trading tools available only when explicitly enabled.10MIT
- AlicenseNot gradedqualityAmaintenanceEnables AI assistants to inspect, query, and diagnose an Octopus instance and, when write access is enabled, create releases, deploy releases, run runbooks, and adjust feature toggles.547 npmMIT