How the pieces fit
Three parts: Ollama runs the model, an agent turns the model’s tool calls into MCP requests, and BankBridge answers those requests from your bank. Ollama never talks to BankBridge. The agent talks to both.
You need a BankBridge key first: sign up at bankbridge.money/login, link a bank, and copy the bbk_ key from your dashboard. To try the setup before you pay, use the key bbk_demo, which returns sample data.
1. Run a model that calls tools
Pull a model with tool support and give the server a large context. Agents send a long list of tools with every request, and a short window cuts it off.
ollama pull qwen3:8b OLLAMA_CONTEXT_LENGTH=65536 ollama serve
2. Hermes Agent in front of Ollama
Hermes Agent reads both the model and the MCP servers from ~/.hermes/config.yaml. Point the model at Ollama’s OpenAI-compatible endpoint and add BankBridge:
model:
provider: custom
default: qwen3:8b
base_url: http://127.0.0.1:11434/v1
api_key: ollama
context_length: 65536
ollama_num_ctx: 65536
mcp_servers:
bankbridge:
url: "https://bankbridge.money/api/mcp"
headers:
Authorization: "Bearer bbk_your_api_key_here"Hermes refuses models with less than 64,000 tokens of context, which is why both context_length and ollama_num_ctx are set. Check the connection with hermes mcp test bankbridge (it should report 12 tools), then ask:
hermes chat -q "List my bank accounts and balances"
With the demo key and qwen3:0.6b, Hermes called mcp__bankbridge__list_accounts and answered “The first account is Chase Total Checking.” Full steps: Hermes Agent setup.
Or: Open WebUI in front of Ollama
Open WebUI is a chat window in your browser that uses Ollama for models. Add BankBridge as an MCP tool server in its admin settings, then turn it on in a chat. The steps are on the Open WebUI setup page.
What about LM Studio?
LM Studio doesn’t run behind Ollama. It downloads and runs its own models, so it replaces Ollama rather than sitting in front of it. If you want one app that runs the model and the tools, LM Studio takes BankBridge in its mcp.json.
What stays on your machine
Your question, the model’s work and its answer stay on your machine. When the model calls a tool, the tool name and arguments go to BankBridge over HTTPS, BankBridge asks the bank-data provider, and the result comes back to your agent. BankBridge keeps no copy of balances or transactions. More on this in agents you run yourself.