Introducing the WorkOfficially MCP Server: Your AI, Working Inside Your Account

Every freelancer we talk to now has an AI assistant open next to the marketplace: pasting job posts into it, asking it to tighten a proposal, copying the answer back. Every client does the same with briefs and candidate lists. The assistant is doing real work, but it is doing it blind, one paste at a time, with no idea what is actually in your account.

On 6 July 2025 we closed that gap. The WorkOfficially MCP server lets your assistant work inside your account: it can see your jobs, proposals, contracts, hours, earnings and inbox, and act on them, as you, within limits you set. It runs at mcp.workofficially.com and works with Claude, ChatGPT, Claude Code, Cursor and any other client that speaks the Model Context Protocol.

What is MCP, in one paragraph

MCP is the open standard AI assistants use to reach outside tools. A service publishes a set of named tools with clear inputs; the assistant decides which to call, calls them over HTTPS, and reads the answer. Think of it as a USB port for AI: one plug, every device. Instead of scraping pages or guessing at your data, the assistant asks WorkOfficially directly, and WorkOfficially answers with exactly what you are allowed to see.

What your assistant can now do

If you freelance

• Find the right jobs, ranked for you. search_jobs returns open jobs with a fit score against the skills on your profile, so "show me the five jobs I am most likely to win this week" is a single question.
• Draft proposals that sound like you. draft_proposal reads the job and your profile and writes a first draft; you edit, then submit_proposal sends it, spending Credits exactly as the website would.
• Run your week. my_contracts, time_diary, my_earnings, my_invoices, my_reputation and wrk_wallet answer "how did this week go?" without opening a single page.
• Stay on top of the inbox. inbox_threads, read_conversation and send_message, with the same contact-sharing rules and blocks the site applies.

If you hire

Hiring is a bigger story than a list; the client guide walks through it end to end.

• Post a job from a sentence. expand_job_brief turns "I need a Shopify store migrated to Hydrogen" into a complete brief with a budget and skills; create_job saves it as a draft and publish_job puts it live when you say so.
• Shortlist in minutes. job_proposals lists every proposal with the freelancer's job success score and history; rank_proposals orders them against your priorities; shortlist_proposal marks the ones you like and tells the freelancer.
• Find talent directly. search_talent, get_freelancer, save_freelancer and invite_freelancer bring the whole hiring loop into the conversation.

Thirty-seven tools in all, plus resources your assistant can read on its own (your profile, contracts and inbox, the platform guide, live platform status) and ready-made prompts for the common jobs: write a proposal, post a job, weekly review, shortlist.

What it can never do

We drew one line before writing a single tool: nothing that moves money or changes who you are. There is no permission, and no tool, for funding or releasing escrow, signing or ending a contract, accepting an offer, identity verification, WRK withdrawals, refunds or disputes. When a conversation reaches one of those steps, the assistant hands you the link and you do it on the website, the same way you do today. The admin side of WorkOfficially is not reachable at all: staff accounts cannot hold an MCP credential, and there is no admin tool to call.

Permissions you actually control

Reading is always included. Every kind of write is its own scope, and you tick them one by one:

Irreversible actions ask twice: closing a job, which declines every open proposal, needs an explicit confirmation from you before the tool will run.

Two ways to connect

Claude and ChatGPT: paste the URL

Add https://mcp.workofficially.com/mcp as a custom connector. The app sends you to WorkOfficially to sign in, shows you exactly which scopes it is asking for, and you approve or deny. That is the OAuth flow; there is no key to copy and you can disconnect the app at any time from Settings, Connected apps.

Claude Code, Cursor and scripts: an API key

Create a key at Settings, API keys, name it after the tool that will hold it, tick the scopes and pick an expiry. It is shown once. In Claude Code it is one line:

claude mcp add --transport http workofficially https://mcp.workofficially.com/mcp \
  --header "Authorization: Bearer wo_key_..."

The full walkthrough for every client is at workofficially.com/developers/mcp.

Built to be trusted with your account

• Keys and tokens are never stored in the clear. Only a one-way hash is kept; the key itself is shown once and never again.
• You hear about every connection. An email goes out each time a key is created or an app is connected, so a key you did not make never goes unnoticed. Creating a key asks for your password again.
• One reset disconnects everything. A password reset, "sign out everywhere" or a support sign-out revokes every key and every connected app at once.
• The assistant sees data, not orders. Job posts, proposals and messages are written by other people. Every result that carries their text is marked for the assistant as data to read, never instructions to follow.
• Nothing you say to the assistant is stored by us. We keep a call log (which tool, when, how long) so you can see what an app did; we never store the arguments or the answers.

Why we built it this way

The easy version of this feature is a read-only feed. The tempting version is an assistant that can do anything you can. We built the one in between on purpose: everything that helps you win and manage work, nothing that can cost you money without your hand on it. Your assistant should make you faster, not make you nervous.

Connect yours today, and tell us what it should learn to do next.