Universal Firefly III AI Bridge
# Universal Firefly III AI Bridge (v3.0.0)
A professional-grade, AI-agnostic bridge providing exhaustive 100% API coverage for connecting AI assistants to your [Firefly III](https://github.com/firefly-iii/firefly-iii) personal finance instance.
**66 tools** covering every major Firefly III capability: accounts, transactions, budgets, bills, recurring rules, automation, insights, attachments, currencies, and more.
---
## Compatibility
| AI Platform | Protocol | Connection | Setup Guide |
| :--- | :--- | :--- | :--- |
| **Claude Code** | MCP (Native) | stdio | [CLAUDE.md](CLAUDE.md) |
| **Claude Desktop** | MCP (Native) | stdio | [CLAUDE.md](CLAUDE.md) |
| **Gemini CLI** | MCP Extension | stdio | [gemini.md](gemini.md) |
| **Cursor / VS Code** | MCP | stdio or SSE | Manual setup below |
| **ChatGPT** | OpenAPI Actions | REST / JSON | `/openapi.json` endpoint |
| **Custom Apps** | REST API | HTTP | `/api/<tool>` endpoints |
---
## Installation
### Prerequisites
- [Node.js](https://nodejs.org/) v18 or later
- A running [Firefly III](https://github.com/firefly-iii/firefly-iii) instance
- A **Personal Access Token** from Firefly III
*(Profile → OAuth → Personal Access Tokens → Create new token)*
### Clone and install
```bash
git clone https://github.com/fabianonetto/mcp-server-firefly-iii.git
cd mcp-server-firefly-iii
npm install
```
---
## Setup by Platform
### Claude Code (Recommended)
Claude Code uses `.mcp.json` in the project directory for MCP server configuration. Credentials go in a separate `.env` file so no secrets are ever in config files.
**1. Create `.env`** in the repo root (gitignored — never committed):
```env
FIREFLY_URL=http://your-host:PORT
FIREFLY_TOKEN=your_personal_access_token
```
Get your token: Firefly III → **Profile → OAuth → Personal Access Tokens → Create new token**
**2. Create `.mcp.json`** in the repo root (gitignored — never committed):
```json
{
"mcpServers": {
"firefly-iii": {
"command": "node",
"args": ["./index.js"]
}
}
}
```
The server reads credentials from `.env` automatically. No secrets in `.mcp.json`.
**3. Start Claude Code** from the repo directory:
```bash
claude
```
The server starts automatically. Claude Code will prompt you to approve it on first launch (once only).
**4. Verify** the connection by asking:
```
Use the get_about tool
```
You should receive your Firefly III version and API information.
**Using from other projects:**
Copy both `.env` and `.mcp.json` to any other project directory, changing `./index.js` to the absolute path:
```json
"args": ["/absolute/path/to/mcp-server-firefly-iii/index.js"]
```
See [CLAUDE.md](CLAUDE.md) for the full guide including auto-approval configuration.
---
### Claude Desktop
Add the server to your Claude Desktop config file:
- **macOS**: `~/Library/Application Support/Claude/claude_desktop_config.json`
- **Windows**: `%APPDATA%\Claude\claude_desktop_config.json`
```json
{
"mcpServers": {
"firefly-iii": {
"command": "node",
"args": ["/absolute/path/to/mcp-server-firefly-iii/index.js"],
"env": {
"FIREFLY_URL": "http://your-host:PORT",
"FIREFLY_TOKEN": "your_personal_access_token"
}
}
}
}
```
Restart Claude Desktop after saving.
---
### Gemini CLI Extension (One-command install)
```bash
gemini extensions install https://github.com/fabianonetto/mcp-server-firefly-iii
```
Then configure your instance:
```bash
gemini config set extensions.firefly-iii-universal-bridge.settings.FIREFLY_URL "http://your-host:PORT"
gemini config set extensions.firefly-iii-universal-bridge.settings.FIREFLY_TOKEN "your_token"
```
See [gemini.md](gemini.md) for the full guide.
---
### Docker (Official Image)
The official image is available on GitHub Packages: `ghcr.io/fabianonetto/mcp-server-firefly-iii`.
#### Run as a service (SSE Mode)
Ideal for ChatGPT Actions, Cursor (SSE), or custom integrations.
```bash
docker run -d \
--name firefly-mcp \
-p 3001:3001 \
-e FIREFLY_URL="http://your-firefly-instance" \
-e FIREFLY_TOKEN="your_personal_access_token" \
-e PORT=3001 \
ghcr.io/fabianonetto/mcp-server-firefly-iii:latest
```
#### Run with Claude Desktop (stdio Mode)
Add this to your `claude_desktop_config.json`:
```json
{
"mcpServers": {
"firefly-iii": {
"command": "docker",
"args": [
"run",
"-i",
"--rm",
"-e", "FIREFLY_URL=http://your-host:PORT",
"-e", "FIREFLY_TOKEN=your_token",
"ghcr.io/fabianonetto/mcp-server-firefly-iii:latest"
]
}
}
}
```
---
### Cursor / VS Code (MCP Extension)
Add to your MCP config (`.cursor/mcp.json` or equivalent):
```json
{
"mcpServers": {
"firefly-iii": {
"command": "node",
"args": ["/absolute/path/to/mcp-server-firefly-iii/index.js"],
"env": {
"FIREFLY_URL": "http://your-host:PORT",
"FIREFLY_TOKEN": "your_personal_access_token"
}
}
}
}
```
---
### HTTP / SSE Mode (ChatGPT Actions, Custom Apps)
Start the server with a port to enable the REST and SSE endpoints:
```bash
FIREFLY_URL=http://your-host:PORT FIREFLY_TOKEN=your_token PORT=3000 node index.js
```
Available endpoints:
| Endpoint | Description |
| :--- | :--- |
| `GET /sse` | SSE transport for MCP clients |
| `POST /messages` | MCP message handler |
| `POST /api/<tool_name>` | Direct REST call to any tool |
| `GET /openapi.json` | OpenAPI 3.0 spec (import into ChatGPT Actions) |
---
## Tool Categories
| Category | Tools | Description |
| :--- | :---: | :--- |
| Core | 1 | System info & connectivity |
| Accounts | 5 | Full CRUD for all account types |
| Transactions | 7 | CRUD, split transactions, search |
| Budgets | 8 | Budgets + monetary limits |
| Bills & Piggy Banks | 7 | Bill tracking + savings goals |
| Automation | 11 | Rules, rule groups, webhooks |
| Recurring | 5 | Recurring transaction rules |
| System | 8 | Currencies + user preferences |
| Insights | 7 | Attachments, charts, net worth, spending |
| Meta | 4 | Categories + tags |
| Object Groups | 2 | Account/piggy bank organization |
| Admin | 1 | Data export |
| **Total** | **66** | |
---
## Documentation
| Document | Description |
| :--- | :--- |
| [docs/API.md](docs/API.md) | Complete reference for all 66 tools and their input schemas |
| [docs/PROMPTS.md](docs/PROMPTS.md) | Prompt examples for common financial tasks |
| [docs/USE_CASES.md](docs/USE_CASES.md) | Strategic guides: tax assistant, subscription auditor, receipt manager |
| [docs/TESTING.md](docs/TESTING.md) | Test suite documentation (78 tests, all tools covered) |
| [CLAUDE.md](CLAUDE.md) | Claude Code & Claude Desktop setup guide |
| [gemini.md](gemini.md) | Gemini CLI extension guide |
---
## Running Tests
```bash
npm test
```
78 tests covering all 66 tools. No live Firefly III instance required — all API calls are mocked.
See [docs/TESTING.md](docs/TESTING.md) for details.
---
## Security
- Use a VPN or SSH tunnel if exposing the server to the internet.
- Keep your `FIREFLY_TOKEN` secret. Never commit `.mcp.json` or `.env` files.
- See [SECURITY.md](SECURITY.md) for the full security policy.
---
## Roadmap
- [x] v1.x — Initial connectivity
- [x] v2.x — Exhaustive API coverage (CRUD & core admin)
- [x] v3.x — Power user features (splits, insights, automation)
- [x] v3.1.0 — Official Docker image & CI/CD
TDQS
Scored across 66 tools
Every tool maps to a distinct resource and action, and similar-sounding tools like get_account vs get_account_overview_chart are clearly separated by descriptions. Even create_transaction and create_split_transaction are differentiated as simple vs split creation.
Tool names consistently follow a lowercase snake_case verb_noun pattern such as list_accounts, get_account, create_account, update_account, and delete_account. Minor exceptions like get_about and create_split_transaction do not undermine the overall predictable convention.
With 66 tools, the surface is far too large for an AI-facing server; even though Firefly III has a broad API, this feels like an uncurated endpoint dump rather than a focused bridge. The volume creates significant context overhead and tool-selection burden for agents.
Core resources like accounts, transactions, budgets, rules, and recurring transactions have full CRUD coverage, but many secondary resources are notably incomplete: categories only have list/create, bills lack get/update, rule groups have no create/update/delete, and webhooks/object groups are partial. These gaps will cause frequent agent dead ends.