--- title: "Bank Data for Agents You Run Yourself | BankBridge" description: "Connect your bank to Hermes, Open WebUI, Qwen Code, Crush, Kilo Code and other agents you run yourself, with a local model if you want. Read-only." url: "https://bankbridge.money/local-agents" --- Agents you run yourself # Your bank, in the agent you run. BankBridge is a remote MCP server. Anything that can add one with an Authorization header can read your accounts, transactions and holdings: a chat app on your own server, an agent in your terminal, a model on your laptop. 12 tools, all read-only. Nothing can move money. ## Where your data goes With a model on your machine, the conversation stays on your machine. The numbers take this path: 1. 1. ### You ask. Your question goes to your model. With a local model, it never leaves your machine or your server. 2. 2. ### The agent calls a tool. The tool name and its arguments (a date range, a merchant name) go to bankbridge.money over HTTPS. 3. 3. ### BankBridge asks the bank-data provider. The provider reads your bank with the read-only access you approved when you linked it. Your bank password never reaches BankBridge. 4. 4. ### The result comes back. BankBridge passes it to your agent and keeps no copy: no balances, transactions or holdings are stored. The usage log keeps the tool name, time and your IP address for 30 days, never the inputs or the result. 5. 5. ### Your model reads it. Locally, if your model is local. If you point the agent at a hosted model, the result goes to that model's provider like any other text in the chat. So a local model takes the conversation off someone else’s servers. It doesn’t make the bank read local: that still needs the internet, the bank-data provider and BankBridge. The [security page](https://bankbridge.money/security) has the rest. ## Setup guides We ran each of these against the demo key before publishing it. Running Ollama? It isn’t an MCP client itself, so it goes behind one of these: [Ollama setup](https://bankbridge.money/guides/how-to-give-ollama-access-to-your-bank-account). ### Chat apps you host A chat window on your computer or your own server. Each one can run a model through Ollama or its own local runtime. [Open WebUIOllama or any OpenAI-compatible endpoint. Admin → Integrations → MCP (Streamable HTTP), Bearer auth.](https://bankbridge.money/docs/open-webui)[AnythingLLMOllama, LM Studio, or its built-in local runtime. One streamable entry in anythingllm\_mcp\_servers.json. Desktop or Docker.](https://bankbridge.money/docs/anythingllm)[LibreChatOllama as a custom endpoint, or hosted providers. One streamable-http block in librechat.yaml, then restart.](https://bankbridge.money/docs/librechat)[Msty StudioLocal models it downloads, or hosted providers. Toolbox → Add New Tool → HTTP with a Bearer Token.](https://bankbridge.money/docs/msty-studio)[LM StudioModels it downloads and runs itself. One-click deeplink install for local models.](https://bankbridge.money/docs/lm-studio) ### Terminal agents Agents that run in your shell. Each of these can point at a model on your machine. [Hermes AgentAny OpenAI-compatible endpoint, including Ollama. One mcp\_servers block in ~/.hermes/config.yaml.](https://bankbridge.money/docs/hermes)[Qwen CodeAny OpenAI-compatible endpoint. qwen mcp add. The settings field is httpUrl, not url.](https://bankbridge.money/docs/qwen-code)[CrushOllama, or any OpenAI-compatible endpoint. One mcp add line in ~/.config/crush/crushrc.](https://bankbridge.money/docs/crush)[Kilo CodeOllama, LM Studio, or hosted providers. One remote server in ~/.config/kilo/kilo.jsonc. CLI and VS Code.](https://bankbridge.money/docs/kilo-code)[GooseOllama, or hosted providers. goose configure, or one extension block in config.yaml.](https://bankbridge.money/docs/goose)[opencodeOllama, LM Studio, or hosted providers. Add to ~/.config/opencode/opencode.json and restart.](https://bankbridge.money/docs/opencode)[OpenClawOllama, or hosted providers. openclaw mcp set, then openclaw mcp probe counts the tools.](https://bankbridge.money/docs/openclaw) ### Run on your machine, model hosted The agent runs locally. By default the model runs on the vendor's servers, so tool results go there. [AuggieAugment Code's models. auggie mcp add. Augment Code's CLI.](https://bankbridge.money/docs/auggie)[Factory DroidFactory's models. droid mcp add with --no-oauth. droid mcp list counts the tools.](https://bankbridge.money/docs/factory-droid) Something else? Any client that speaks MCP over streamable HTTP works: [raw MCP docs](https://bankbridge.money/docs/api). Agents with no MCP at all can use the [command line tool](https://bankbridge.money/cli). ## Try it before you sign up The key `bbk_demo` returns sample accounts and transactions. Put it in any config on these pages, or list the tools from a terminal: terminal ``` curl -s https://bankbridge.money/api/mcp \ -H "Authorization: Bearer bbk_demo" \ -H "Content-Type: application/json" \ -H "Accept: application/json, text/event-stream" \ -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' ``` ## Questions ### If my model runs locally, does my bank data stay on my machine? The conversation does: your question, the model's reasoning and its answer never leave your machine. The numbers don't start there, though. When the agent calls a tool, BankBridge asks the bank-data provider, the provider reads your bank, and the result comes back to your agent over HTTPS. BankBridge keeps no copy of it. ### Does it work offline? No. The model can run offline, but reading your bank needs an internet connection to BankBridge and from there to your bank. ### What does BankBridge log when my agent calls a tool? The tool name, time, calling app and version, user agent, IP address, status and response time, kept for 30 days. The log never holds your question, the tool's inputs, or what the tool returned. ### Which local models work? Any model your agent can call tools with. Small models miss tool calls more often than large ones. In our test, a 0.6B Qwen 3 model on Ollama, driven by Hermes Agent, called list\_accounts and read back the first account name from the demo data. Hermes needs a context window of at least 64,000 tokens. ### What does it cost? $5 a month per connected bank, or $50 a year. The same subscription works in every agent. Your agent and your model cost whatever they cost on their own.