Fieldproxy MCP server

The official Model Context Protocol server for Fieldproxy, AI-native field service management.

It lets the AI assistant your team already uses (Claude, ChatGPT, Microsoft 365 Copilot and Teams, GitHub Copilot, Cursor, OpenAI Codex, Claude Code, or any MCP client) read and change a Fieldproxy workspace, and build on it, with a person confirming every change.

The part no other field service MCP server does: your AI can change the software itself. Ask for a new field, screen, table, automation or agent, and it is built as a draft in a sandbox copy of your workspace and goes live only when an administrator approves, with rollback.

• Overview: https://www.fieldproxy.ai/mcp
• Which field service software has an MCP server (verified comparison): https://www.fieldproxy.ai/mcp#which-fsm-has-mcp

This repository holds documentation and client configuration. The server itself is hosted by Fieldproxy; there is nothing to install or run.

Addresses

Use the address for your workspace's region.

Transport: streamable HTTP. Auth: OAuth 2.0 with dynamic client registration (sign in with workspace ID, email and password), or a bearer key (fpmcp.…).

Before you connect

AI assistant access is off by default. Fieldproxy switches it on per workspace (Portal, Org features, AI assistant access, Enabled, and optionally Allow build). Ask your Fieldproxy contact, or book a demo.

Keys (fpmcp.…) are made in the Fieldproxy web app under Settings, AI assistant, New key. Each key is shown once. An administrator can set up a key for someone else; that person claims it once from their own screen, so the administrator never sees the key.

Connect a client

Claude (Desktop and claude.ai)

Add a custom connector with your region's address, then sign in with workspace ID, email and password and tick the scopes. The sign-in lasts 1 hour and renews itself for 30 days. Guide: https://www.fieldproxy.ai/mcp/claude

Claude Code

claude mcp add --transport http fieldproxy https://api-us-east-1.fieldproxy.ai/mcp \
  --header "Authorization: Bearer fpmcp.…"

ChatGPT (web: Plus, Pro, Business, Enterprise, Edu)

1. Settings, Security and login: turn on Developer mode.
2. Create a developer-mode app named Fieldproxy with your region's address, and choose OAuth with dynamic client registration.
3. Sign in on the Fieldproxy page and tick the permissions to allow.

ChatGPT asks before each write action by default; Fieldproxy's own confirm step still applies on top. Guide: https://www.fieldproxy.ai/mcp/chatgpt

OpenAI Codex (CLI, IDE extension, desktop)

export FIELDPROXY_MCP_KEY="fpmcp.…"
codex mcp add fieldproxy --url https://api-us-east-1.fieldproxy.ai/mcp --bearer-token-env-var FIELDPROXY_MCP_KEY

Or in ~/.codex/config.toml: see examples/codex-config.toml. To sign in instead of using a key, add the server without --bearer-token-env-var, then run codex mcp login fieldproxy.

GitHub Copilot in VS Code

Copy examples/vscode-mcp.json to .vscode/mcp.json. VS Code asks for the key once and hides it. Leave out headers to sign in through the Fieldproxy page instead.

Microsoft 365 Copilot and Teams (Copilot Studio)

1. In your Copilot Studio agent: Tools, Add a tool, New tool, Model Context Protocol.
2. Name it Fieldproxy, add a one-line description, and enter your region's address.
3. Authentication: OAuth 2.0, Dynamic discovery (each person signs in with their own Fieldproxy account), or API key, type Header, name Authorization, value Bearer fpmcp.….
4. Create the connection, Add to agent, then publish the agent to Teams and Microsoft 365 Copilot.

Copilot Studio needs its own Microsoft licence, and a Power Platform data policy can block the connection. Guide: https://www.fieldproxy.ai/mcp/copilot

Cursor and other JSON-config clients

See examples/cursor-mcp.json.

What it can do

145 built-in tools, plus up to 14 Command Center ready-made reads and any tools the workspace defines itself (saved SQL tools).

Seven ready-made workflows (prompts): daily_ops_briefing, plan_the_schedule, triage_my_inbox, why_did_this_automation_fail, customer_360, build_an_app, build_a_dashboard.

Scopes

The user's own Command Center access sets the ceiling. A key can be limited further: expiry (1 hour to 90 days), tables, rows (for example "only team A's jobs"), hidden columns, insert/update/delete, and switches for SQL, sending and build areas. Postgres enforces row limits, and hidden columns are removed from every result.

Guardrails

• Confirm before change. 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; more than 1,000 is refused.
• Builds save as drafts. Saving or publishing is blocked if the linter, a SQL check against the database, or a navigation check finds a new error. New automations and agents are saved switched off. Public apps are never published or 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.
• Stateless and audited. Every request is authorised from scratch, so a revoked key is refused on the next call. Every call is logged with the tool, key, person and outcome.

Example asks

• "What did we finish last week that nobody invoiced?"
• "Mike is out sick. Move his jobs to whoever is closest and show me the plan."
• "Which quotes over $5,000 have had no follow-up in two weeks? Draft the emails."
• "Add a warranty expiry date to every job, and alert the service manager 30 days before one runs out."
• "Build a subcontractor compliance screen that flags missing insurance."

Links

• Overview: https://www.fieldproxy.ai/mcp
• Claude: https://www.fieldproxy.ai/mcp/claude
• ChatGPT: https://www.fieldproxy.ai/mcp/chatgpt
• Microsoft 365 Copilot and Teams: https://www.fieldproxy.ai/mcp/copilot
• Book a demo: https://www.fieldproxy.ai/book-demo?source=github-mcp