Aurify MCP Gateway
An MCP server that lets AI agents manage Aurify financial data—wallets, transactions, budgets, investments, and profile—by translating MCP calls into the same REST API the web frontend uses.
README
Aurify MCP Gateway
An MCP server that lets an AI agent (Hermes, Claude, any MCP client) do anything a user can do in the Aurify web app.
What it is, and what it deliberately is not
It is an adapter, not a microservice. It owns no database, no aggregate, no
events, and no migrations. Every tool is a translation of an MCP call into a
request against refina-web-bff — the same REST surface the web frontend uses —
so there is exactly one implementation of every business rule, and it lives in
the domain services.
Hermes ──MCP (Streamable HTTP)──► aurify-mcp ──REST──► refina-web-bff ──gRPC──► wallet / transaction / …
Bearer aur_pat_… │ Bearer <5-min JWT>
└──► refina-auth (PAT → JWT exchange)
It is intentionally not folded into the BFF: the BFF is the hot path for the browser, tuned with a Redis read cache and a rate limiter for that traffic, whereas this needs long-lived protocol sessions, tool schema discovery, scope enforcement, and an audit trail. Different lifecycle, different blast radius.
Authentication
The agent presents an Aurify personal access token; the gateway exchanges it for a five-minute JWT and calls the BFF with that.
The exchanged JWT carries exactly the claims the BFF already expects, so no downstream service needed any change to support agents. The PAT is the durable credential and can be revoked at any time; the access token it yields is too short-lived to be worth stealing.
Scopes are enforced here, before any network call. The BFF still only checks the signature, exactly as it does for the browser.
Getting a token
# 1. Log in to Aurify and grab your JWT, then mint a PAT with only the scopes
# the agent actually needs.
curl -X POST http://localhost:8080/auth/pat \
-H "Authorization: Bearer <your login JWT>" \
-H "Content-Type: application/json" \
-d '{"name":"hermes","scopes":["wallet:read","transaction:read","transaction:write"]}'
# 2. Copy the `token` from the response — it is shown once and never again.
# 3. Revoke it whenever you like:
curl -X DELETE http://localhost:8080/auth/pat/<id> -H "Authorization: Bearer <JWT>"
Available scopes: wallet:read, wallet:write, transaction:read,
transaction:write, budget:read, budget:write, investment:read,
investment:write, analytics:read, profile:read, profile:write.
Tools
28 tools by default. Reads are unrestricted; writes are annotated
readOnlyHint: false so the client prompts for confirmation.
| Group | Read | Write |
|---|---|---|
| Wallets | list_wallets, get_wallet, get_wallet_summary, list_wallet_types |
create_wallet, update_wallet |
| Transactions | list_transactions, get_transaction, list_categories |
create_transaction, create_fund_transfer, update_transaction |
| Budgets | list_budgets |
create_budget, update_budget, reset_budget |
| Analytics | get_financial_summary, get_balance_history, get_net_worth, get_transaction_analytics, get_category_transactions |
— |
| Investments | list_investments, get_investment_summary, list_asset_codes |
create_investment |
| Profile | get_profile |
update_profile |
| System | — | refresh_cache |
delete_wallet, delete_transaction, delete_budget and sell_investment are
not registered at all unless MCP_ALLOW_DELETE=true. Filtering rather than
rejecting means an agent never even sees them advertised.
Two conveniences are built in, because the web UI solves them in the frontend:
create_fund_transferfills in the reserved Cash In / Cash Out category ids itself, so they are absent from the tool schema.- Tool descriptions explain the credit-line balance convention, so an agent does not mistake available credit for money owned.
Credit cards
A wallet whose wallet_type_nature is liability is a credit line. Its
balance is the credit still available: spending lowers it, and paying the
bill — a fund transfer into the wallet — restores it. Available credit is never
part of net worth. The tool descriptions state this so the model does not have
to infer it.
Running
npm install
cp .env.example .env # point BFF_BASE_URL / AUTH_BASE_URL at your services
npm run build
npm start # POST /mcp, GET /health
npm test # unit tests
npm run lint # typecheck
Configuration
| Variable | Default | Purpose |
|---|---|---|
PORT |
8080 |
Listen port |
BFF_BASE_URL |
(required) | refina-web-bff base URL |
AUTH_BASE_URL |
(required) | refina-auth base URL, for PAT exchange |
MCP_ALLOW_DELETE |
false |
Register destructive tools |
REQUEST_TIMEOUT_MS |
15000 |
Upstream request timeout |
No JWT signing secret is needed: the gateway never mints tokens, it only asks the auth service for them.
Wiring up Hermes
Add to ~/.hermes/config.yaml and run /reload-mcp:
mcp_servers:
aurify:
url: http://127.0.0.1:8090/mcp
headers:
Authorization: Bearer aur_pat_xxxxxxxx
enabled: true
tools:
include: [list_*, get_*, create_transaction, create_fund_transfer,
update_transaction, create_budget, update_budget, refresh_cache]
resources: false
prompts: false
Verify with hermes mcp test aurify.
Use tools.include (an allowlist) rather than exclude — this is financial
data, and an allowlist fails closed when new tools are added.
Operational notes
- Stateless. Each request builds its own server and transport, both torn down when the response ends. One caller's token can never leak into another's session, and the process can be restarted or scaled freely.
- Session cache. Exchanged JWTs are cached per token until 30s before expiry, and concurrent misses collapse into one exchange. In practice the first tool call in a turn costs a round trip to auth and the rest do not.
- Audit log. One JSON line per tool call on stdout, recording the tool, scope, outcome and duration. Argument keys are logged, never values — enough to audit without copying the user's financial data into the logs.
- Bind to localhost. Hermes runs as a host process, so publishing
127.0.0.1:8090is sufficient. This gateway should not be exposed publicly. - Shared rate limiter. The BFF's Redis rate limiter is shared with browser
traffic, and its read cache can serve stale figures right after a write —
hence the
refresh_cachetool.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.