Upwork MCP in Claude Code: What 88 Calls Taught Me
The day after an Upwork rejection I rebuilt my profile through Claude Code and Upwork's MCP. 35 calls, zero errors, and two entries that said "saved" and never appeared. What 88 calls taught me about its tools, its previews and the permission rules I'd set.

Table of Contents9 sections
Real tests on Claude and SEO tools, with the numbers and what failed. About once a week.
Almost there. Check your inbox and click the link to confirm. Not there in a minute? Look in spam.
Something went wrong on my side. Try again in a minute.
Too many tries. Give it a minute.
That email doesn't look right. Check it for a typo.
Unsubscribe anytime. Privacy

On 24 October, the day after an Upwork client turned me down, I rebuilt my Upwork profile from Claude Code. 35 calls to Upwork's MCP in about 75 minutes, and not one error came back.
Two of the entries it confirmed never showed up on the profile.
The short version: it reads jobs, client hire rates and rival profiles well, and most profile and proposal changes come back as a preview you confirm.
But a confirmed write can still fail without a word. And one tool, Send Message, skips the preview, so that's where your own Claude Code rule matters.
I've made about 90 calls in six sessions, the last nine today. The counts below come from those session logs.
Only here for the setup? Go straight to the permission rules I'd set.
Key Takeaways
- Upwork's MCP launched on 10 August 2026. It's free, hosted by Upwork, and Claude Code adds it with one command.
- Every profile and proposal change comes back as a preview. Nothing happens until a second call confirms it, and the preview expires after about 15 minutes.
- Send Message has no preview.
- "Confirmed" didn't always mean done: two experience entries came back with new ids and never appeared, and four skill names landed as other skills.
- The
/mcplabels don't set your prompts. Your rules do.
What the Upwork MCP is
Upwork announced its MCP server on 10 August 2026, free for every client and freelancer. It's a hosted server, so there's nothing to install.
In Claude Code it's one line, then /mcp to log in to Upwork in the browser:
claude mcp add --transport http upwork https://mcp.upwork.com/mcp
Upwork's setup page has the same steps for ChatGPT, Cursor and Codex. Sign-in is OAuth 2.1, and you can revoke it any time from your Upwork account settings.
Money never moves through it. Upwork says payments and offer acceptance finish on upwork.com.
The 33 tools, sorted by what they can do
On 24 October, /mcp listed 34 tools. The screenshot is in my vibe coding post, where I used them to read 306 jobs.
Today there are 33. Two "Legacy" draft tools are gone, and a Help Center search is new.
The labels say read-only, destructive or open-world. I find it more useful to sort them by what stands between Claude and your account:
| What happens on a call | Tools | Count |
|---|---|---|
| Reads only | Find Jobs, Find Freelancers, Get Profile, Get Messages, List Contracts, the dashboard, Help Center search and the rest | 18 |
| Prepares a preview, changes nothing yet | Update Profile, Manage Proposals, Respond to Offer, Submit Milestones, Manage Meetings, Boosts | 6 |
| Carries out a preview | Submit Preview (confirm_preview) | 1 |
| Acts at once | Send Message, Save Job, Set Tool Mode, Set Tool Permission, the three upload tools, Submit Feedback | 8 |
That last row is the one to watch. Save Job and the upload tools are harmless. Send Message goes straight to a client, and its help text only tells Claude to ask you first.
What 88 calls looked like
Here's every session I've run it in, counted from the session logs.
| Date | The job | Upwork MCP calls | Errors |
|---|---|---|---|
| 24 Oct | Rebuild my profile, then read top earners for the portfolio | 35 | 0 |
| 31 Oct | Search jobs matched to the new profile | 13 | 0 |
| 1 Nov | Find "finish my AI-built app" jobs | 7 | 0 |
| 5 Nov | Read four competitors, rewrite title, overview and skills | 21 | 0 |
| 12 Nov | Check invitations and proposals | 3 | 0 |
| 13 Nov | Re-check the tools for this post | 9 | 0 |
On 24 October only 7 of the 34 tools did all the work: Update Profile 11 times, Find Freelancers 10, Submit Preview 7.
The same session made almost 300 Claude in Chrome calls for the rest. Its transcript is gone now, so these figures come from the run log I kept that night.
Visibility (it was Private), my rate, the portfolio and the "Open for work" badge all needed the browser.
Every write is a preview first
On 5 November I changed my title. Update Profile didn't change it. It sent back this, with a preview id:

Nothing on Upwork moves until Claude calls Submit Preview with that id, and the preview expired about 15 minutes after the call.
Two more things I learned the hard way:
- On 24 October I made three skill previews in a row, and only the last could be confirmed. The limit is one pending preview per type.
- On 5 November title, overview and skills were pending together, and all three went through. Different types don't block each other.
The overview preview carries the old text and the new one, so it doubles as a diff.
Where it fails without telling you
Zero errors doesn't mean zero failures. I hit two kinds.
The first was Add Other Experience.
Claude confirmed it twice and got a fresh record id back each time. Neither entry ever showed up on the profile, even after two Get Profile checks, so I added both by hand.
The second was skill names. Upwork matched what I typed to the nearest skill it knows, and didn't say so:
| I typed | Upwork saved |
|---|---|
| GEO | Google Earth |
| Generative Engine Optimization | Generative AI |
| Generative Engine Optimization (GEO) | AI Overview Optimization |
| Answer Engine Optimization | Search Engine Indexing Optimization |
Upwork has no GEO skill. The AEO tags that work are "aeo" and "AEO Strategy", the ones job posts use.
By 5 November exact names went in cleanly: Claude Code, Claude Skills, SEO Strategy, Website Migration. Read the skill list in the preview before you confirm it, then check with Get Profile.
Today's help text says a skill name that doesn't resolve now shows as rejected in the preview. I haven't tested that yet.
Reading is where it earns its keep
On 31 October the feed matched to my profile held 74 jobs. Most were GEO, AEO or AI visibility, the role I read 36 remote postings for, so the rewrite had steered Upwork's recommender.
Opening a single job shows what the search list hides: the client's real hire rate. Here are the four I opened that day:
| Job, as posted | Client hire rate | Connects to apply | Could I apply? |
|---|---|---|---|
| eCommerce Figma Designer For Landing Pages | 100% | 22 | No, balance 0 |
| Home page + Nav bar fixes | 41% | 20 | No |
| AI Architect Consultant | 27% | 22 | No |
| Looking for expert UI/UX Designer | 25% | 15 | No |
A 25% client hires once for every four jobs they post. Would you spend 15 Connects on that?
Today the balance history explained something else. The "Open for work" badge took 1 Connect every 12 hours, 14 a week, until the free ones ran out:

The MCP can't switch that badge on, by design. Its Boost tool only reads the badge and turns it off, because turning it on starts a weekly charge.
Two details in the raw output surprised me.
Anything another person wrote arrives inside untrusted_participant_content tags, so Claude treats a client's post as data, not as orders.
And every job link carries utm_source=claude. Upwork can count what came through Claude.
Competitor research works well too: Find Freelancers takes filters like earnings over $100,000 and job success over 90%, and Get Profile reads each result in full.
I read four top SEO profiles on 5 November, then reordered my portfolio to open on +562K organic visits.
The permission rules I'd set
Three gates sit between Claude and your Upwork account: Claude Code's rules, Upwork's preview, and the steps that only finish on upwork.com.
| Gate | Who runs it | What it stops |
|---|---|---|
| Allow, ask and deny rules | Claude Code, from your settings | Calls that match an ask or deny rule |
| Preview, then Submit Preview | Upwork's server | Profile, proposal, offer and milestone changes |
| Finish on upwork.com | You, in the browser | Accepting offers, paying, turning on the badge |
The labels in /mcp aren't a gate. Claude Code's MCP docs say a read-only hint lets it repeat a call and run it beside other reads.
They don't say the label skips your prompt. Your permission rules do that, checked deny first, then ask, then allow.
This is what I'd put in .claude/settings.json:
{
"permissions": {
"allow": [
"mcp__upwork__upwork__list_*",
"mcp__upwork__upwork__get_*",
"mcp__upwork__upwork__find_*",
"mcp__upwork__upwork__search_help_center"
],
"ask": [
"mcp__upwork__upwork__confirm_preview",
"mcp__upwork__upwork__send_message",
"mcp__upwork__upwork__boost_profile",
"mcp__upwork__upwork__set_tool_mode",
"mcp__upwork__upwork__set_tool_permission"
]
}
}
The server is named upwork and its tools start with upwork__ too, hence the double name. So the reads run without a prompt. The previews run too, because a preview changes nothing.
You're asked at the two points where something real happens: Submit Preview, and Send Message, which has no preview at all.
Upwork also has its own switch, Set Tool Permission. Write tools default to "always ask", and "always allow" never skips the preview. I'd leave it alone.
I run Claude Code in auto mode (my settings post covers the modes). Its docs say an explicit ask rule still forces a prompt there, so these two still stop and wait for me.
Common Questions
How do I connect the Upwork MCP server to Claude Code?
Run claude mcp add --transport http upwork https://mcp.upwork.com/mcp in your terminal, then type /mcp in a Claude Code session to log in to Upwork. It's free, hosted by Upwork, and signs in with OAuth 2.1.
Can the Upwork MCP submit proposals and spend my Connects?
Yes, but in two steps. Manage Proposals first returns a preview showing the Connects cost and your balance. Connects are only spent when Claude calls Submit Preview with that preview's id, so a Claude Code ask rule on that tool keeps you in control.
Does the Upwork MCP pay freelancers or accept offers for me?
No. Upwork says funding escrow, paying a milestone and accepting an offer all finish on upwork.com. The MCP can prepare an offer or decline one, but the binding steps happen in your browser.
Why did my Upwork skills change to the wrong tags through the MCP?
Upwork matches each typed skill to the nearest one in its own list. In October 2026 "GEO" became Google Earth and "Answer Engine Optimization" became Search Engine Indexing Optimization. Read the skill list in the preview before confirming it.
I run real tests on Claude and SEO tools, like 10 voice models on one Mac or 25 audits of a site with 20 planted problems. You get the numbers, including what failed. One email per post, about once a week.
Almost there. Check your inbox and click the link to confirm. Not there in a minute? Look in spam.
Something went wrong on my side. Try again in a minute.
Too many tries. Give it a minute.
That email doesn't look right. Check it for a typo.
Unsubscribe anytime. Privacy



