sure/docs/hosting/mcp.md

# MCP Server for External AI Assistants

Sure includes a Model Context Protocol (MCP) server endpoint that allows external AI assistants like Claude.ai, Claude Desktop, GPT agents, or custom AI clients to query and act on your financial data.

## What is MCP?

[Model Context Protocol](https://modelcontextprotocol.io/) is a JSON-RPC 2.0 protocol that enables AI assistants to access structured data and tools from external applications. Instead of copying and pasting financial data into a chat window, your AI assistant can directly query Sure's data through a secure API.

This is useful when:
- You want to use an external AI assistant (Claude, GPT, custom agents) to analyze your Sure financial data
- You prefer to keep your LLM provider separate from Sure
- You're building custom AI agents that need access to financial tools

## Authentication Modes

Sure supports two ways to authenticate MCP clients:

### 1. OAuth 2.0 / dynamic client registration (recommended)

This is the best option for Claude.ai and other MCP clients that support OAuth. Sure exposes:

- `/.well-known/oauth-protected-resource`
- `/.well-known/oauth-authorization-server`
- `POST /register` for dynamic client registration

These endpoints let compatible MCP clients register a public OAuth client, redirect you back to Sure for sign-in, and receive a bearer token with the `read_write` scope.

### 2. Static bearer token via environment variables

This is the simpler fallback for custom agents, scripts, and deployments where you want to pin the MCP server to a specific Sure user.

Set these environment variables:

| Variable | Description | Example |
|----------|-------------|---------|
| `MCP_API_TOKEN` | Bearer token for authentication | `your-secret-token-here` |
| `MCP_USER_EMAIL` | Email of the Sure user whose data the assistant can access | `user@example.com` |

Both variables are required for the legacy token flow. OAuth clients using the
MCP discovery and dynamic registration endpoints do not need these variables.

### Generating a secure token

Generate a random token for `MCP_API_TOKEN`:

```bash
# macOS/Linux
openssl rand -base64 32

# Or use any secure password generator
```

### Choosing the user for static-token auth

The `MCP_USER_EMAIL` must match an existing Sure user's email address. The AI assistant will have access to all financial data for that user's family.

> [!CAUTION]
> The AI assistant can call the MCP tools available to the specified user. This includes reading financial data and write-capable tools such as statement import, goal/category/tag changes, transaction updates, and budget updates. Only set this for users you trust with your AI provider.

## Configuration

### Docker Compose

Add the environment variables to your `compose.yml`:

```yaml
x-rails-env: &rails_env
  MCP_API_TOKEN: your-secret-token-here
  MCP_USER_EMAIL: user@example.com
```

Both `web` and `worker` services inherit this configuration.

### Kubernetes (Helm)

Add the variables to your `values.yaml` or set them via Secrets:

```yaml
env:
  MCP_API_TOKEN: your-secret-token-here
  MCP_USER_EMAIL: user@example.com
```

Or create a Secret and reference it:

```yaml
envFrom:
  - secretRef:
      name: sure-mcp-credentials
```

## Protocol Details

The MCP endpoint is available at:

```
POST /mcp
```

### Authentication

MCP supports OAuth authorization-code flow for clients such as Claude Code.
Clients should discover the protected-resource metadata, register dynamically,
request the advertised `read_write` scope, and send the resulting access token
as a Bearer token. Dynamically registered clients are assigned this scope so
their tokens can authenticate to MCP.

For self-hosted deployments or clients without OAuth support, requests may use
the legacy `MCP_API_TOKEN` as a Bearer token:

```
Authorization: Bearer <token>
```

That token can come from either:

- an OAuth authorization flow handled by the MCP client, or
- the static `MCP_API_TOKEN` environment variable described above.

### Supported Methods

Sure implements the following JSON-RPC 2.0 methods:

| Method | Description |
|--------|-------------|
| `initialize` | Protocol handshake, returns server info and capabilities |
| `tools/list` | Lists available financial tools with schemas |
| `tools/call` | Executes a tool with provided arguments |

### Available Tools

The MCP endpoint exposes the same tool registry used by Sure's built-in assistant. Clients should treat `tools/list` as the source of truth.

At the time of writing, `tools/list` includes:

| Tool | Description |
|------|-------------|
| `get_transactions` | Search transactions with filters (exact names or ids), sorting by date or absolute amount, and pagination |
| `get_recurring_transactions` | Detected and manual recurring transactions (subscriptions, bills, salaries) with expected dates and per-currency totals |
| `get_accounts` | Accounts with ids and current balances; pass `include_balance_series: true` for a period-bounded history series |
| `get_holdings` | Query investment holdings |
| `get_balance_sheet` | Net worth, assets and liabilities with a configurable history period and interval |
| `get_income_statement` | Income and expenses for a period, with optional monthly series, prior-period comparison and account filtering |
| `get_budget` | Budget summary for a month, with optional prior months |
| `get_merchants` | Merchants with the ids `update_transaction` accepts and the exact names `get_transactions` filters on |
| `get_tags` | Tags with pagination |
| `get_categories` | Categories with hierarchy and pagination |
| `create_goal` | Create a savings goal linked to depository accounts |
| `create_tag` / `update_tag` | Manage tags |
| `create_category` / `update_category` | Manage categories |
| `update_transaction` | Edit a transaction's metadata (name, notes, category, merchant, tags) |
| `update_budget` | Update budget allocations for a month |
| `import_bank_statement` | Import bank statement data |
| `search_family_files` | Search documents uploaded through the import flow. Note this is the vector-store document index, not the Statement Vault — statements archived via `upload_account_statement` are not searchable through it |

These are the same tools used by Sure's built-in AI assistant.

### Preview Tools

These additional tools appear only when the MCP user has opted into preview
features (Settings → Preferences). Until then they are absent from `tools/list`,
and calling one by name returns an "Unknown tool" error. The Statement Vault
tools additionally require the user to be an admin or member, matching the
permissions enforced in the web UI.

| Tool | Description |
|------|-------------|
| `upload_account_statement` | Store a statement document (PDF/CSV/XLSX) in the Statement Vault; deduplicates by SHA-256 |
| `list_account_statements` | List vault documents with their SHA-256, period, linked account and review status |
| `get_account_statement` | One statement's details and its reconciliation checks against the ledger — present only once someone has entered the statement's opening/closing balances in the web UI, since nothing extracts them from the document. Does not return the file: stored documents are served only to a signed-in browser session |
| `get_statement_coverage` | Month-by-month statement coverage for an account: `covered`, `missing`, `mismatched`, `ambiguous`, `duplicate`, `not_expected`, each with a reconciliation status |
| `record_valuation` | Record an account's value on a date, with a required source citation |
| `get_valuations` | List recorded valuations newest first, including the citation stored in each entry's notes; the read pair for `record_valuation` |
| `get_insights` | Read the proactive insights feed (spending anomalies, cash-flow warnings, subscription audits and more) without marking anything read |

They exist for agents that maintain a document-backed record of a family's
wealth over time. See
[Wealth history with an external agent harness](../llm-guides/wealth-agent-harness.md).

## Example Requests

### Initialize

Handshake to verify protocol version and capabilities:

```bash
curl -X POST https://your-sure-instance/mcp \
  -H "Authorization: Bearer your-secret-token" \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "initialize"
  }'
```

Response:

```json
{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "protocolVersion": "2025-03-26",
    "capabilities": {
      "tools": {}
    },
    "serverInfo": {
      "name": "sure",
      "version": "1.0"
    }
  }
}
```

### List Tools

Get available tools with their schemas:

```bash
curl -X POST https://your-sure-instance/mcp \
  -H "Authorization: Bearer your-secret-token" \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 2,
    "method": "tools/list"
  }'
```

Response includes tool names, descriptions, and JSON schemas for parameters.

### OAuth discovery

MCP clients that support OAuth can discover Sure's metadata automatically:

```bash
curl https://your-sure-instance/.well-known/oauth-protected-resource
curl https://your-sure-instance/.well-known/oauth-authorization-server
```

The authorization-server metadata includes:

- `authorization_endpoint`: `https://your-sure-instance/oauth/authorize`
- `token_endpoint`: `https://your-sure-instance/oauth/token`
- `registration_endpoint`: `https://your-sure-instance/register`
- `scopes_supported`: `["read_write"]`

### Call a Tool

Execute a tool to get transactions:

```bash
curl -X POST https://your-sure-instance/mcp \
  -H "Authorization: Bearer your-secret-token" \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 3,
    "method": "tools/call",
    "params": {
      "name": "get_transactions",
      "arguments": {
        "start_date": "2024-01-01",
        "end_date": "2024-01-31"
      }
    }
  }'
```

Response:

```json
{
  "jsonrpc": "2.0",
  "id": 3,
  "result": {
    "content": [
      {
        "type": "text",
        "text": "[{\"id\":\"...\",\"amount\":-45.99,\"date\":\"2024-01-15\",\"name\":\"Coffee Shop\"}]"
      }
    ]
  }
}
```

## Security Considerations

### Transient Session Isolation

The MCP controller creates a **transient session** for each request. This prevents session state leaks that could expose other users' data if the Sure instance is using impersonation features.

Each MCP request:
1. Authenticates the token
2. Loads the authorized Sure user
3. Creates a temporary session scoped to that user
4. Executes the tool call
5. Discards the session

This ensures the AI assistant can only access data for the intended user.

### Pipelock Security Scanning

For production deployments, we recommend using [Pipelock](https://github.com/luckyPipewrench/pipelock) to scan MCP traffic for security threats.

Pipelock provides:
- **DLP scanning**: Detects secrets being exfiltrated through tool calls
- **Prompt injection detection**: Identifies attempts to manipulate the AI
- **Tool poisoning detection**: Prevents malicious tool call sequences
- **Policy enforcement**: Block or warn on suspicious patterns
- **Signed receipts**: Produces verifiable evidence for mediated MCP decisions when the flight recorder is configured with storage and a signing key

See the [Pipelock documentation](pipelock.md) and the example configuration in `compose.example.ai.yml` for setup instructions.

### Network Security

The `/mcp` endpoint is exposed on the same port as the web UI (default 3000). For hardened deployments:

**Docker Compose:**
- The MCP endpoint is protected by the `MCP_API_TOKEN` but is reachable on port 3000
- For additional security, use Pipelock's MCP reverse proxy (port 8889) which adds scanning
- See `compose.example.ai.yml` for a Pipelock configuration

**Kubernetes:**
- Use NetworkPolicies to restrict access to the MCP endpoint
- Route external agents through Pipelock's MCP reverse proxy
- See the [Helm chart documentation](../../charts/sure/README.md) for Pipelock ingress setup

## Production Deployment

For a production-ready setup with security scanning:

1. **Download the example configuration:**

   ```bash
   curl -o compose.ai.yml https://raw.githubusercontent.com/we-promise/sure/main/compose.example.ai.yml
   curl -o pipelock.example.yaml https://raw.githubusercontent.com/we-promise/sure/main/pipelock.example.yaml
   ```

2. **Set your MCP credentials in `.env`:**

   ```bash
   MCP_API_TOKEN=your-secret-token
   MCP_USER_EMAIL=user@example.com
   ```

3. **Start the stack:**

   ```bash
   docker compose -f compose.ai.yml up -d
   ```

4. **Connect your AI assistant to the Pipelock MCP proxy:**

   ```
   http://your-server:8889
   ```

The Pipelock proxy (port 8889) scans all MCP traffic before forwarding to Sure's `/mcp` endpoint.

## Connecting AI Assistants

### Claude.ai

Sure's Settings UI is already geared toward Claude.ai OAuth integrations:

1. Open **Settings -> Integrations** in Claude.ai
2. Click **Add integration**
3. Paste your Sure MCP URL
4. Claude redirects you to Sure to sign in and authorize access

If you are using Pipelock, use the reverse-proxy URL on port `8889`. Otherwise use the app URL ending in `/mcp`.

### Claude Desktop

If your Claude Desktop build expects a raw MCP endpoint instead of an OAuth integration flow, point it at:

- `http://your-server:8889` when using Pipelock, or
- `http://your-server:3000/mcp` for direct access

Use either the client's OAuth support or a bearer token, depending on what that build supports.

### Custom Agents

Any AI agent that supports JSON-RPC 2.0 can connect to the MCP endpoint. The agent should:

1. Send a POST request to `/mcp`
2. Include the `Authorization: Bearer <token>` header
3. Use the JSON-RPC 2.0 format for requests
4. Handle the protocol methods: `initialize`, `tools/list`, `tools/call`

## Troubleshooting

### "unauthorized" error

**Symptom:** Requests return HTTP 401 with "unauthorized"

**Fix:** Verify one of these is true:

- The OAuth flow completed successfully and the client is sending the issued bearer token
- The static token matches `MCP_API_TOKEN`
- If you are using the static-token flow, `MCP_USER_EMAIL` matches an existing Sure user

### Static token works, but the user still gets rejected

**Symptom:** Requests return HTTP 401 even though the bearer token matches `MCP_API_TOKEN`

**Fix:** The `MCP_USER_EMAIL` probably does not match an existing user. Check that:
- The email is correct
- The user exists in the database
- There are no typos or extra spaces

### Pipelock connection refused

**Symptom:** AI assistant cannot connect to Pipelock's MCP proxy (port 8889)

**Fix:**
1. Verify Pipelock is running: `docker compose ps pipelock`
2. Check Pipelock health: `docker compose exec pipelock /pipelock healthcheck --addr 127.0.0.1:8888`
3. Verify the port is exposed in your `compose.yml`

## See Also

- [External AI Assistant Configuration](ai.md#external-ai-assistant) - Configure Sure's chat to use an external agent
- [Pipelock Security Proxy](pipelock.md) - Set up security scanning for MCP traffic
- [Model Context Protocol Specification](https://modelcontextprotocol.io/) - Official MCP documentation