How to run an AI sourcing agent for engineering roles in Claude, ChatGPT or Cursor
An AI sourcing agent is a chat assistant like Claude, ChatGPT or Cursor that you connect to a people-search tool over MCP (Model Context Protocol). You give it a role brief, and then it searches, narrows the results and tells you why each person fits. For engineering roles, the agent is only as good as the search behind it, so connect one that ranks engineers by the work they've actually done.
We start with a brief for a Chicago role
Our brief was for a senior Python engineer in Chicago, or someone willing to work Chicago hours, who builds developer tooling. Experience with MCP servers or agent frameworks was a big plus, and they had to have shipped open source that other people use.
If you paste that into a chat model with no search connected, the model still answers. You get ten plausible names with plausible titles, and some of those people may not exist. We asked for engineers, and we got back text that looks like engineers. That's the first way sourcing agents fail, so we keep one rule: if the agent can't point to a profile, a repository or a role for a person, we treat that person as made up.
The second failure is harder to notice. When you connect the agent to a search that only covers profiles, the names become real, but the reasons don't get any better. You get whoever wrote the right words on their profile the last time they updated it, and most engineers rarely update theirs. The third failure comes from the second. The agent tells you someone has a "strong Python background", and you can't check that. If it tells you someone is the second most active contributor to fastmcp, you can check that in a minute.
We fix all three the same way, so we connect the agent to a search that returns real people along with the work that made each person match.
We connect the agent to a real search
We build Vamo, and we rank engineers by what they've built on GitHub, so we used Vamo for this run. Once Vamo was connected, we pasted the brief as written and added one line about evidence: "Use Vamo to find engineers for this brief. For each person, name the repository or role that made them match. Show me ten."
The agent splits that brief into two parts. The facts become filters: Chicago, not a bot, and not an organization account. The work becomes a short ranking phrase, "MCP servers and Python agent tooling". We keep places and job titles out of that phrase, so people get scored on their work and not on the words in their profile.
Then we narrowed the results the way we would with a sourcer sitting next to us. First we asked it to "drop anyone who has not pushed code in the last six months" and to keep people who contribute to projects they don't own. Then we asked, "Of these, who has worked on fastmcp or another MCP framework?" and had it explain each fit in two sentences.
You can check the first person in one click
What we wanted back was a person with a reason attached. One engineer who fits this brief is nate nowack, a senior software engineer at Prefect in Chicago. He's the second most active contributor to fastmcp, the Python framework for building MCP servers and clients, and only its creator has contributed more. He also contributes to Prefect itself.
On nate's card at the top of this post, you can open fastmcp and Prefect from his repositories, and his garden shows how often he commits in public.
You can check every claim in that paragraph with one click, because his name is right there in fastmcp's contributors. The ten names from the run without a search would have taken us an afternoon to disprove.
Then we asked the follow-up a recruiter would ask: "Find five more people like him who are in the Midwest."
The agent answers that with the similar tool, and that's usually how we turn one good person into a short list. A lot of the best people we find sit two hops out from a strong person or team.
Where you should take over from the agent
After a good run, you'll want to hand the agent the rest of the job. Ashby and Greenhouse both ship MCP servers now, so Claude or ChatGPT can read your pipeline and move candidates between stages (Greenhouse MCP, Ashby MCP server). Those servers only return people who already applied. A sourcing agent works on everyone else, and we stop it earlier than you might think.
We let the agent find people and explain why they fit, and then a person on the team decides what happens next. In practice we do four things:
- We review before anything goes out. If the agent drafts anything, a person reads it and decides whether to send it.
- We check the evidence ourselves. We open the repository and read one pull request, and that takes about a minute.
- We leave hiring decisions to the team. We use the ranking to decide who to talk to first, and nothing more.
- We check the brief for bias, and the results too. If you write "top school" in the brief, the agent will filter on it. The Chicago brief worked because we wrote it around the work.
In Ashby, for example, each user can require approval before an agent takes any action in the ATS (Ashby announcement). We'd turn on settings like that wherever you find them.
For some jobs we wouldn't use the agent at all. If you're hiring at high volume for non-technical roles, you need job boards and fast screening. Some excellent engineers have never pushed public code, and you find them through referrals, their employer history and the people they work with. Our hidden engineers post covers some of that.
Run the same moves on your own brief
Take the role you have open and write the brief around the work, the way we wrote the Chicago brief. Then run the same moves. Paste the brief with the evidence line, narrow the results, ask about the one project that defines the role, and then ask for more people like the best person. If the agent can't name a repository for every person in its first answer, you're back in the run without a search, and you know what to fix.
We use these prompts for most briefs, with the role details swapped in.
The role brief. "Use Vamo to find [level] engineers in [place] who have built [the work, in one phrase]. Must have: [two or three facts]. Nice to have: [one or two]. For each person, name the repository or role that made them match."
More like this person. "Find ten more engineers like [GitHub login], based in [place]. Same kind of work, not the same employer."
Contributors to a repository near a place. "Who has contributed to [owner/repo] and lives in [city]? Rank them by how much they have shipped there." We use this one a lot for infrastructure roles, because if your team depends on a project, the people who maintain it already know your problems.
Rank people you already have. "Here are twenty GitHub logins from our referral sheet. Rank them against this brief and tell me which three to call first and why." The agent uses deep_research for this, and we often get more out of it than out of finding new people.
Worked alongside a team. "Find engineers who have worked on the same repositories as the fastmcp maintainers, in Chicago."
If you want a vendor comparison of recruiting agents instead, see top AI recruiter agents and our AI recruiting agent reviews.
Appendix: run the search yourself
Add the connector. The Vamo server address is https://mcp.vamotalent.ai/mcp/search. The MCP page has the steps for each client.
- Claude: Settings, then Connectors, then Add custom connector. Name it Vamo and paste the address. The Claude guide walks through it.
- Claude Code: run
claude mcp add --transport http vamo https://mcp.vamotalent.ai/mcp/search. - ChatGPT: Settings, then Apps and Connectors. Turn on Developer mode under Advanced, choose Create and paste the address (ChatGPT guide).
- Cursor: add the server to
~/.cursor/mcp.jsonin your home directory, then open Cursor Settings, then MCP, and click Connect (Cursor guide).
Sign in. Vamo uses OAuth for sign in. Your AI tool opens a Vamo page and you approve it, so you never paste a key into a chat. If your tool can't sign in that way, it can send an API key as a bearer token instead.
The four tools. Your assistant picks the right one for each ask:
- search finds engineers by place, employer, school, title, languages and the repositories they built or contributed to, ranked by the work you describe.
- similar finds more engineers like one person you already like, anywhere or in one place.
- enrich looks up people you already know, by GitHub login or LinkedIn, many at once.
- deep_research writes a longer report on one developer or one repository, or ranks people you already have against your brief.
The request we sent in the Chicago run. If you'd rather call the API directly, send this:
Ask Claude
API request
curl -X POST https://api.vamotalent.ai/v2/developers/search
-H "Authorization: Bearer vamo_sk_..."
-H "Content-Type: application/json"
-d '{
"rank_by": ["v", "ANN", ["Embed", "MCP servers and Python agent tooling"]],
"filters": ["And", [
["location_ids", "ContainsAny", { "place": ["Chicago"] }],
["is_bot", "Eq", false],
["is_org_account", "Eq", false]
]],
"rerank": "clef"
}'You'll find every filter, sort and field you can send in the search API reference.
Put facts like place, employer and languages in filters. Describe the work in the ranking phrase, and keep that phrase short, with no places or company names in it. Each person comes back with their devRank and the repository or role that made them match. You can browse public engineer profiles at /engineers.
Best for: Finding software engineers by what they've built on GitHub, joined with their LinkedIn career and location.
Not for: Non-technical roles such as sales, operations or finance, high-volume hourly hiring, and engineers whose work all sits in private repositories, because we only read public work.
Price: Usage-based plans that we set up with you on a call. We don't publish a list price or offer a self-serve free trial.
FAQ
How do I source engineers from Claude Code or Cursor? Connect the Vamo MCP server, because the same server address works in any MCP client. In Claude Code, you add it with one command. In Cursor, you add it to ~/.cursor/mcp.json and click Connect. Then paste the brief and name Vamo in your ask so that the agent uses it.
What recruiting platform lets me set up a sourcing agent? Any AI tool that supports MCP, such as Claude, ChatGPT or Cursor, becomes a sourcing agent once you connect it to a search that returns real people. We connect Vamo, which ranks engineers by what they've built on GitHub and names the repository or role that made each person match.
What does it cost? You get the MCP server and the API on the same Vamo plan. Plans are usage-based, and we set them up with you on a call. We don't publish a list price or offer a self-serve free trial.
Where does the data come from? We read public engineering work on GitHub and join it to each engineer's work history, so you can filter by place, employer and school. We show our own signals, rankings and summaries of that work.
What is the difference between a sourcing agent and other recruiting agents? Sourcing agents find people who are not in your pipeline yet. Screening agents sort the applications that already came in. Scheduling agents book interviews, and outreach agents write and send messages. Your ATS vendor is moving into screening and scheduling.
Are AI recruiting agents effective for niche roles? For niche technical roles they often do better than for general ones, because the work narrows the field fast. We ask for contributors to the two or three projects that define the niche and start there.
Can an AI agent automate outbound recruiting? The agent can find people, explain why they fit and draft the outreach, but we keep a person reviewing each message before it goes out. A person on the team decides what happens next.
If you have a role open now, connect Vamo and paste the brief, or book a demo and we will run it with you. For a side by side of the MCP servers you could connect, read recruiting MCP servers compared.