fieldwork-mcp
# Fieldwork MCP
**Ask your pest control business questions in plain English.**
Works with Claude, Cursor, Codex, and other AI apps that support [MCP](https://modelcontextprotocol.io/).
> “How many customers do I have?”
> “Who owes me money?”
> “How much Alpine did we use this week?”
> “Who did the most jobs?”
Your AI asks Fieldwork. You get a clear answer.
**Read-only** — it can’t change jobs, charge cards, or edit customers.
Site: [`landing/index.html`](landing/index.html) (Relay home) · [`landing/fieldwork.html`](landing/fieldwork.html) (Ask Fieldwork) · Roadmap: [`ROADMAP.md`](ROADMAP.md)
---
## Who this is for
Pest and lawn-care owners who run **[Fieldwork](https://fieldworkhq.com/)** and want answers without digging through reports.
---
## Why can’t they just ask Claude or Cursor to set it up?
**Often they can** — if they’re already in Cursor or Claude Code with this project open, say:
> Set up Fieldwork MCP for me. I’ll paste my API key into `.env`.
The agent can create the venv, write MCP config, and wire the launcher.
You still need to **get the API key yourself** (it’s your Fieldwork password-equivalent).
Don’t paste the key into a random chat if you can drop it straight into `.env`.
**When a copy-paste snippet still helps**
- You’re on the **website** and don’t have an agent in the folder yet
- You’re on **Claude Desktop**, which doesn’t edit your disk the same way
- Someone else is setting up a machine without opening this repo
The endgame (see Roadmap) is a **one-line remote URL + login**, like LandingFolio — no Terminal, no JSON.
---
## What you can ask
### Customers
- How many customers do I have?
- Do I have a customer named ___?
### Money
- Who owes me money?
- How much did we make this month?
- How is business doing?
### Product
- How much product was used?
- What chemicals do we carry?
### Schedule, routes & technicians
- What’s on the schedule this week?
- Who are my technicians?
- How are my routes doing?
- Who did the most jobs?
**Bonus prompt:** `monday_morning_briefing` — one bundled ops snapshot.
---
## Quick start
### 1. Get your Fieldwork API key
Easiest path (in the Fieldwork app):
1. Go to **[Settings → Users & Routes](https://app.fieldworkhq.com/settings/users)**
2. Open your user (or an API Integration user)
3. Open the **Integration** tab and copy the **API Key**
Help: [Fieldwork APIs](https://intercom.help/fieldwork/en/articles/2406552-fieldwork-apis)
### 2. Install
```bash
git clone https://github.com/daisydaines/fieldwork-mcp.git
cd fieldwork-mcp
python3 -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -e .
cp .env.example .env
# put FIELDWORK_API_KEY=... in .env
```
### 3. Connect your AI app
**Cursor** — `~/.cursor/mcp.json`:
```json
{
"mcpServers": {
"fieldwork": {
"command": "/ABS/PATH/TO/fieldwork-mcp/scripts/run_mcp.sh"
}
}
}
```
**Claude Desktop** — same JSON in
`~/Library/Application Support/Claude/claude_desktop_config.json`
**Claude Code:**
```bash
claude mcp add fieldwork -- /ABS/PATH/TO/fieldwork-mcp/scripts/run_mcp.sh
```
**Codex** — add a stdio MCP server pointing at `scripts/run_mcp.sh` (same idea as Cursor).
Then restart the app and ask: *How many customers do I have?*
### Hosted connect (no Dropbox-style OAuth)
Fieldwork does not offer a third-party OAuth login popup. Closest path:
```bash
./scripts/run_mcp_http.sh
# open http://127.0.0.1:8000/connect
```
Paste your API key once. You get a bearer token and a one-line snippet for Claude / Cursor / Codex.
Also serves the Relay landing at `/`.
| URL | What |
| --- | --- |
| `http://127.0.0.1:8000/` | Landing |
| `http://127.0.0.1:8000/connect` | Paste API key → connect token |
| `http://127.0.0.1:8000/mcp` | Streamable HTTP MCP (`Authorization: Bearer <token>`) |
For a shared deploy, set `FIELDWORK_VAULT_SECRET` (required) and `FIELDWORK_PUBLIC_BASE_URL`
(e.g. `https://mcp.example.com`). See `Dockerfile`. Keep `FIELDWORK_MCP_ALLOW_ENV_FALLBACK=0`.
---
## Safety
| Can do | Cannot do |
| --- | --- |
| Read customers, jobs, invoices, usage | Create or edit customers |
| Summarize routes & techs | Cancel or reschedule jobs |
| Report product used on jobs | Take payments or change prices |
Tools are annotated `readOnlyHint`. The HTTP client **refuses non-GET** requests.
---
## For builders
```bash
pip install -e ".[dev]"
pytest
ruff check src tests
./scripts/run_mcp.sh # stdio
./scripts/run_mcp_http.sh # streamable-http
```
Auth to Fieldwork: `api-key` header (default).
Unofficial. Not affiliated with Fieldwork / Anstar Products.
## License
MIT — see [LICENSE](LICENSE).
TDQS
Scored across 28 tools
Many tools appear to serve the same purpose, such as what_products_do_we_carry and list_materials, how_much_money and get_financial_summary, and whats_on_the_schedule and list_calendar_work_orders. The natural-language question tools overlap heavily with the structured API tools, creating ambiguity about which to use. An agent would struggle to pick the right tool consistently.
The tool names mix full-sentence questions (how_many_customers, who_owes_me_money) with standard verb_noun patterns (get_invoice, list_materials, search_customers). There is no consistent convention; even similar actions use different verbs such as find_customer vs search_customers. This inconsistency makes it harder to predict tool names.
With 28 tools, the server exceeds the 25-tool threshold for 'too many' and includes many near-duplicate question-answering and structured-data tools. The count feels inflated by redundancy rather than justified by distinct functionality.
The server provides broad read-only coverage across customers, work orders, invoices, materials, routes, technicians, and summaries. It lacks any create/update/delete operations, which may be acceptable for an analytics assistant, but there are minor gaps such as no all-invoices list endpoint and no single-technician fetch.