What it actually is
The Upwork MCP Server is built on the Model Context Protocol, an open standard for connecting AI agents to outside services, and it's hosted directly by Upwork at mcp.upwork.com. It works inside Claude (Web, Desktop, and Code), ChatGPT, Cursor, Codex, and any other MCP-compatible agent that supports OAuth. There's no local install, you add the server URL to your agent, log in with your Upwork account through a browser popup, and your credentials refresh automatically from there.
What a freelancer can actually do through it
Once connected, the agent can act on your account across most of what you'd normally do on the website:
- Pull a live summary of invitations, offers, messages, matched jobs, and your Connects balance in one call.
- Search jobs by skill, category, budget, or type, and pull matches based on your profile.
- Draft a proposal grounded in your actual profile and past work against a specific job's requirements, then review it before it goes anywhere.
- Respond to invitations and offers, accept, decline, or ask for changes.
- Track active contracts, submit milestone work, and check logged time.
- Check earnings and update your availability, employment history, languages, education, and portfolio.
- Read and send messages, and manage your availability badge or profile boost settings.
Nothing binding happens without you
This is the detail that matters most. Every write action, submitting a proposal, sending a message, updating your profile, is drafted first and requires your confirmation before it goes through. Connects are only spent when you confirm a proposal, exactly like bidding on the website. Binding actions like accepting an offer complete on upwork.com itself, not inside the AI conversation. Upwork's own FAQ is direct about this: an agent cannot hire, or in your case, act, on your behalf without you confirming, full stop, at least in the current release.
The security model
Access runs through OAuth 2.1, so there are no API keys for you to manage or leak. You can revoke access at any time from Account Settings, which immediately kills both the access and refresh tokens. Upwork states plainly that it never sees your chat logs, your conversation history stays inside your AI agent, not on Upwork's servers.
Why this matters more than it looks like
Unofficial "Upwork MCP" tools have existed on GitHub and elsewhere since well before this official launch, and some of them work in ways worth knowing about before you trust one with your account. A few rely on browser automation, driving a real Chrome session through the Chrome DevTools Protocol specifically to get around Upwork's bot detection. That's not a minor technical detail, it's the exact category of behavior our account bans guide flags as one of the top scraping triggers on the platform right now. An unofficial tool built to evade detection is a real risk to the account you're trying to protect. The official connector authenticates the normal way and doesn't need to hide what it's doing from Upwork, because it's Upwork's own infrastructure underneath.
Setting it up
Exact steps vary slightly by agent, Claude, ChatGPT, Cursor, and Codex each have their own connector or MCP settings panel, but the shared pattern is: add the server URL https://mcp.upwork.com/mcp, authorize through the browser popup that opens, and you're connected. Upwork's own setup page walks through each agent's specific steps if you want the exact clicks.
Where this fits with everything else changing right now
This connector is infrastructure, not a new place clients discover you. It sits underneath the visibility changes we've already covered, how clients now find and hire through ChatGPT and Claude is the discovery layer, this MCP server is what lets an agent actually act on an account once it's connected. It's also one of several AI-native pieces Upwork shipped in 2026 alone, alongside Uma's instant interview feature and the full rundown of what changed on Upwork this year. Investing in AI infrastructure this heavily isn't incidental either, Q2 2026 earnings show AI-related work is one of the few categories still growing while overall GSV contracted. If you're not sure your profile holds up once an AI agent is the one reading it on a client's behalf, run it through our audit tool.
your turn
Want this analysis run on your own profile?
UpBRO audits your Upwork profile section by section against the same knowledge base and returns concrete fixes in about a minute.