Fieldproxy MCP for ChatGPT
Connect ChatGPT to your field service software
Ask ChatGPT to change your field service software, not just read it: new fields, screens and automations, built in a sandbox and live when you approve. Plus 145 tools for dispatch, customers, invoices and messages.
Ask anything
Illustration of Fieldproxy's MCP tools inside ChatGPT. Names and data are examples.
Short answer
How do I connect ChatGPT to my field service software?
Create a Fieldproxy app in ChatGPT's developer mode, sign in with your Fieldproxy workspace, and ChatGPT can answer from your live jobs, customers and invoices and do the work: plan routes, draft follow-ups, issue proposals. ChatGPT asks before each write, and Fieldproxy's own confirm step applies on top.
The part nobody else does
Your AI can change the software, not just the data
Every other field service connector we checked reads records and, at most, edits a few of them. Fieldproxy's MCP server can also build: new fields, screens, tables, automations and agents, made in a sandbox copy of your workspace and live only when you approve.
Add a field
Build a screen
Wire an automation
Create an agent
Add your own tools
Roll back anything
Read more: customise field service software with AI · build it with Claude or ChatGPT instead?
Setup
Connect ChatGPT in three steps
Turn on developer mode
In ChatGPT on the web (Plus, Pro, Business, Enterprise or Edu), open Settings, Security and login, and turn on Developer mode.Create the app
Create a developer-mode app named Fieldproxy with the Fieldproxy address from Settings, Connect an AI assistant and OAuth with dynamic client registration.Sign in
Sign in on the Fieldproxy page with workspace ID, email and password, and tick the permissions to allow.
- OpenAI Codex uses the same address: codex mcp add fieldproxy --url <address> --bearer-token-env-var FIELDPROXY_MCP_KEY, or codex mcp login fieldproxy to sign in instead.
In ChatGPT
What people ask it
"Give me today's operations briefing: late jobs, unassigned jobs and anything overdue."
"Which customers have invoices more than 60 days overdue? Draft a polite reminder for each."
"Plan tomorrow's routes for the North region and show me the drive time saved."
"Show me the before and after photos on job 4471."
See it on a workspace built on your own data.
Get DemoWhat your AI can do
145 tools, across the whole field operation
Not just lookups. The assistant can plan the schedule, send the follow-up and build the screen, and every change waits for a person to confirm.
Read
Customers, jobs, technicians, invoices, quotes and contracts. Accounts-receivable ageing, technician performance, funnels, and read-only SQL against your own schema.
Dispatch and maps
Spread jobs across technicians, plan daily routes and apply the schedule. Drive times, the nearest technician, live locations and job sign-off links.
Comms
Search the inbox and read threads. Draft emails and SMS from templates, then send after you confirm. Comments, notifications, email follow-ups and phone calls.
Sales documents
Preview and issue proposals, generate estimates and PDF documents.
Files and photos
Read a record's PDFs, Word, Excel and email files, and look at the photos themselves, for example to compare a job's before and after.
Monitoring
Dashboards, alerts, scheduled reports, the activity feed and usage.
Automation operations
Run a rule, retry or dismiss a failed run, and decide approvals (named approvers only).
Write, with undo
Every change is proposed, confirmed by a person, and can be undone. CSV import and export, record history and restoring deleted records.
Build
Apps, tables, automations, voice and text agents, and document templates, built as drafts in a sandbox copy of your workspace and promoted to live when you approve.
Guardrails
Your AI proposes. A person confirms.
Built for operations with branches, teams and an IT review, not a single login.
Nothing changes on one call
Every write, send or action is a proposal the user confirms within 30 minutes. Writes can be undone. More than 5 rows needs bulk permission, and more than 1,000 is refused.
Keys scoped like API keys
Limit a key to certain tables, certain rows ("only the North region's jobs"), hidden columns, insert/update/delete, and an expiry from 1 hour to 90 days. Postgres enforces the row limits.
Builds save as drafts
Apps and automations are checked against the database before they save, new automations and agents are saved switched off, and public apps are never changed.
Sends are limited
Email goes only from the user's own mailbox, at most 20 sends an hour per key, inside the workspace's quiet hours. A sandbox never sends.
Everything is logged
Every call is written to an audit log with the tool, key, person and outcome. Administrators see every key and connected app, and revoking one applies on the next request.
Access levels
Three levels, and your own role is the ceiling
Each connection gets one or more scopes. A person can never grant an AI more than they can do themselves.
| Scope | Requires | Unlocks |
|---|---|---|
| Read (data:read) | AI assistant access on for the workspace, and the user has Command Center access | Your schema, ready-made reports, and read-only SQL on your own data |
| Write (data:write) | Tables the workspace lets the AI write, and no approval gate on them | Changes to records, confirmed by a person and undoable; email and SMS drafts, sent on confirm |
| Build (build) | An administrator, plus Allow build switched on for the workspace | New fields, tables, screens, automations, voice and text agents, and sandboxes |
A key can be narrowed further, like an API key: an expiry from 1 hour to 90 days, only certain tables, only certain rows (for example one region's jobs, the key owner's team, or a service zone), hidden columns, and any mix of insert, update and delete. Limits can be described in plain words, and Fieldproxy shows how many rows the key would see before it is created.
See it built around how your team already works.
Get DemoDeliberately not possible
What the AI can never do, however it is asked
- Change anything on a single call: every write, send or action is a proposal a person confirms within 30 minutes
- Change more than 1,000 records in one go (more than 5 needs bulk permission)
- Drop a table or a column: builds only ever add
- Publish or change a public app
- Read a hidden column, however the query asks for it
- See data the signed-in user cannot see: their own access is the ceiling
- Send email from anyone's mailbox but the user's own, or more than 20 sends an hour per key, or outside quiet hours
- Send anything at all from a sandbox
- Run JavaScript tools or tools tied to an app screen (those stay inside Fieldproxy)
- Call a tool that is not in its list, even by name
Security
How the connection actually works
Every request starts from scratch
Nothing is kept between calls, so a changed setting or a revoked key applies on the very next request.
Sign-in on every call
Each request carries a key (fpmcp.…) or a sign-in token that lasts 1 hour and renews itself for 30 days. The key is looked up every time, so a revoked or expired one is refused at once.
A tool list built per person
The tools the AI sees are assembled for each request from the key's scopes, the person's role, the workspace settings and the key's own limits. A tool missing from the list is refused even when called by name.
Reads run locked down
Reads run as a read-only database role, one statement at a time. Postgres applies the key's row limits and hidden columns are removed from every result.
Changes become proposals
Writes, sends, runs and builds are shown to you first and happen only after you confirm. Builds are saved as drafts, and new automations and agents are saved switched off.
Everything is on the record
Every call is logged with the tool, key, person and outcome. Administrators can read the log, see every key and connected app, and revoke any of them.
FAQ
Questions
- Can ChatGPT connect to field service software?
- Yes. ChatGPT supports MCP apps in developer mode on Plus, Pro, Business, Enterprise and Edu plans. Fieldproxy's MCP server connects with OAuth, so each person signs in with their own Fieldproxy account and sees only what their role allows.
- Will ChatGPT send messages to my customers without asking?
- No. Emails and SMS are drafted first, and sent only after you confirm. They go from your own mailbox, at most 20 an hour per key, inside your quiet hours.
- Is my data used to train models?
- Fieldproxy passes only the results of the tools you call, and access is limited by scopes and row-level limits. How ChatGPT treats conversation data is set by your OpenAI plan and settings.
Explore Fieldproxy MCP
Start here
Use cases
More use cases
See ChatGPT run your operation
We connect it to a Fieldproxy workspace built on your data and run your real questions.