Remoet is now listed in Claude's connector directory. You can find it at claude.ai/directory/remoet, connect it from there, and start asking Claude about tech jobs without setting anything up.
This was our second attempt. The first submission went in back in May and was never listed. No rejection, no email, nothing. (In fairness to Anthropic, by August a good part of what we had written in that form was no longer true, because we had made Remoet free and deleted the paid tier it described.)
So on 2 October we did it again, properly this time, and I wrote down every step. This post is that log: what the portal asks, what it caught, what we had to fix in our own server before we were willing to hit submit, and how long it all took.
What you can do with it
The short version, for anyone who just wants to try it:
Search jobs with no account. Ask for roles by technology, title or company ("find me remote Rust jobs") and Claude reads the public board directly. Every role links to the employer's own posting.
Find companies by the stack they hire for. Also no account. The stack comes from the roles a company has open right now, not from tags it picked for itself.
Star companies, read your feed, keep a profile. These need a free Remoet account, and Claude asks you to sign in the first time you use one. Paste a CV and Claude fills in your profile.
What it does not do: apply for you. Applications to roles on public careers pages happen on the employer's own site, and Claude hands you the link. (That one is on purpose.)
If you would rather wire it into another agent, the same search works there too.
The portal is not the form we remember
In May you needed a Team or Enterprise organisation to submit at all. Now any paid Claude plan can, at claude.ai/directory/manage. It is ten steps (connection, tools, listing, use cases, company, authentication, data handling, test and launch, compliance, review) and the draft saves in your browser tab as you go.
The part I liked: you paste your server URL and the portal connects to the live server and pulls your tools and prompts itself. No copying tool lists into a text box. Which also means it tells you, immediately, what it does not like.
Every tool was "missing a title" (none were)
Every one of our tools had a title. The portal flagged all of them anyway: "Missing title annotation. Add annotations.title".
The MCP spec moved the tool title from annotations.title to a top-level title field, and the portal reads the old location. We now set both (one line in the helper every tool goes through), re-synced, and the warnings disappeared.
There was one more suggestion, which we left alone: "Reference the target API docs in the description" on update_profile. As we read Anthropic's review criteria, that rule is aimed at custom query tools, the kind that take freeform endpoints or request bodies. update_profile is a plain typed tool, so we judged it a false positive and moved on. (The listing was approved without it, so I think we read that one right?)
The rule we were quietly breaking
Anthropic's review criteria say a tool description should describe what the tool does, not tell Claude how to behave. Descriptions that push Claude to call tools the user did not ask for are exactly what they warn against.
Ours were full of it. "IMPORTANT: Always call this first before making any changes." "Typical flow: read get_profile for the user's stack and seniority, search..., then star the genuine fits." "TIP: Suggest the user adds their link tree URL to their CV." It felt helpful when we wrote it: we were coaching the agent. Read with fresh eyes, it is exactly the pattern a prompt-injection scanner should be suspicious of.
We rewrote 14 descriptions so they state facts (what the tool returns, where an id comes from, which tool reports your star budget), then added a test that fails if a description starts scripting other tool calls again.
Some guidance stayed. "Confirm with the user before deleting" and "only star companies whose stack overlaps yours" are about the tool's own action, so we kept them. The compliance step asks you to confirm "no instructions about model behavior", which is broader than the criteria, so we said so in the notes field rather than tick the box and hope.
If you maintain an MCP server, this is the one to check before you submit. It is very easy to write and very hard to notice in your own code.
Making the connector match the website
The prerequisites field ("what does a user need before connecting?") turned up an inconsistency I did not like. On remoet.dev you can browse companies by tech stack without an account. Through the connector, the same company search asked you to sign in.
We could have explained that gap in the form. We fixed it instead, the same day. Company search now works without an account through the connector too, same as job search.
(Then came the boring but necessary part: updating every place that described what works without an account. The docs, the agent skill, the website, llms.txt, the server's own instructions. Five repositories for one sentence.)
The portal guessed our auth wrong
The portal probes your server without credentials. Ours answered, because search is public, so it detected "None: anyone can use this server without signing in".
That is true for search and wrong for everything else. Our account tools answer with a 401 and an OAuth challenge, and the right option was "Required when the server asks": Claude connects without credentials and prompts you to sign in when a tool needs it. Left on "None", I suspect starring would have been broken for everyone arriving from the directory. Worth checking if your server mixes public and private tools like ours does.
Things that broke before we got this far
None of these were the portal's fault. They were ours, and the submission is what surfaced them:
A security bug in our OAuth redirect check. We fixed and deployed it before submitting, since every sign-in from the directory goes through that flow.
Our reviewer login had quietly stopped working when we moved hosting earlier this year. A config value did not make the move, and a redirect between our two domains turned a POST into a GET. Nothing told us. We only found out because we needed a reviewer login link.
We tested every tool as the reviewer account before saying we had. They all worked. (Then we cleaned up after ourselves: every test write was undone and the temporary API keys are deleted.)
Use cases: write what people actually type
The form asks for up to three use cases with example prompts. Our first drafts were wrong in an instructive way: "Find senior roles that use Rust, Go or TypeScript." Who types that? Nobody. A real person asks about their own stack, in plain words.
What we submitted:
Search open tech jobs: "Find me remote Rust jobs."
Build a profile from a CV: "Here's my CV, set up my Remoet profile."
Find and star companies by stack: "Which companies are hiring Go engineers? Star the ones worth watching."
We also dropped "check your starred companies for new roles" as a showcase. It is the habit we want people to build, but for a brand new user it returns an empty feed, which is a bad first impression for a reviewer too.
How long it took
We started on 2 October and submitted late that afternoon, every fix above included.
Then about four and a half days of waiting. The approval email arrived late on 6 October (UTC), and publishing the listing was one more step on our side.
What came back from review
Nothing, apart from the approval. No questions, no change requests. The email said the listing was "approved as a community connector by an automated review".
Which means the reviewer login we fixed so carefully was probably never used. (I am not mad about it. It was broken, and now it is not.)
Community is the label an automated approval gives you. The listing says plainly that Anthropic has not verified it, and that you should only connect developers you trust. Verified is a separate step, and Anthropic explains the difference here. Fair enough, honestly: an automated pass is not a person reading your OAuth flow, so holding that bar is on us.
Was it worth it?
Ask me in a month. Back in July we decided that directory listings are credibility, not distribution: they make you findable by name, and anything beyond that is earned by people actually using the thing. I still think that is right, and I am not expecting a traffic spike.
What I can already say is that the submission paid for itself in what it made us fix. A security bug, a broken reviewer login, tool descriptions that bossed the agent around, and a connector that was stricter than our own website. We found all four because we sat down to fill in a form.
If the listing brings in nobody, I will say so in a follow-up.
If you want to try it, Remoet is in the directory, and searching needs no account. And if you have been through this review yourself and it went differently, I would genuinely like to hear about it: [email protected].
