SeekOut MCP: What Chat-Window Sourcing Actually Replaces

In May 2026, SeekOut shipped a Model Context Protocol server that lets a recruiter run a billion-profile search from inside Claude. Greenhouse followed with 36 ATS tools exposed to Claude, ChatGPT, and Google Antigravity. Manatal gated its own MCP behind the Enterprise Plus tier and claimed first-to-market. The sourcing UI you spent three years learning is being quietly demoted to a viewer pane.

What SeekOut MCP actually ships

SeekOut MCP is a hosted Model Context Protocol endpoint that exposes 14 guided recruiting workflows and 6 data sources, backed by SeekOut's 1B+ profile index, to any MCP-compatible AI assistant. MCP, if you have not met it yet, is Anthropic's open protocol for letting a chat model call tools on a remote server with the user's credentials.

The specifics worth memorizing before you demo it to your team:

• Clients supported on day one: Claude, ChatGPT, Gemini, Microsoft 365 Copilot, and Llama.
• Also distributed as: a SeekOut-published connector for Copilot Studio, Power Automate, Power Apps, and Logic Apps.
• Hard rate limit: 1,000 requests per user per day by default.
• Setup gate: on Claude for teams, a Claude Owner or Primary Owner has to open Organization settings and add the connector. There is no self-serve toggle for an individual sourcer.

The 14 workflows are the interesting constraint. SeekOut did not expose its raw search API to Claude; it exposed a curated set of prompts. You can ask Claude for a Rust engineer in Berlin and it will route the question to one of those 14 routines, not to a free-form query against the index. That is a very different product than "SeekOut, but typed."

The quiet replacement no vendor is naming

The chat window is replacing the query builder, not the results view. Every MCP recruiting server I have looked at this quarter still returns a link back to the vendor's own UI the moment the result set grows past a handful of names.

SeekOut MCP's documented behavior is explicit: it returns a clickable link to open the results in the SeekOut web app for any multi-candidate review. Greenhouse MCP responses behave the same way for pipeline work: Claude will summarize and filter, then hand you back to Greenhouse for the actual stage move on 40 candidates. The "chat replaces the UI" narrative recruiters are hearing at conferences is half true. What chat replaces is the Boolean builder, the facet sidebar, and the saved-search dropdown. What survives is the grid.

That matters for how you budget time. A sourcer who used to spend most of a search in the query builder now spends almost none of it there, and the entire session in the grid. The skill shifts from "can you write a 14-clause Boolean" to "can you read a Claude-generated shortlist and know which 8 of the 50 are real."

14

Greenhouse MCP flips the governance problem

Greenhouse MCP inherits the signed-in user's existing ATS permissions, which solves the obvious data-leak question and creates a new one. The connection is OAuth 2.0 with PKCE and Dynamic Client Registration against the user's Greenhouse account, so a sourcer in Claude can only see what they would see in the Greenhouse UI. Admins set a ceiling; user permissions apply underneath. That is the two-check model.

The non-obvious failure mode: Claude is now a new attack surface for a system the sourcer already had read access to. A prompt-injected candidate note, a poisoned resume PDF, or a hostile email body sitting in a Greenhouse candidate record is one "summarize this candidate" call away from executing instructions against the signed-in user's token. Expect the first public incident in 2027. The vendor will not be at fault; the model will be.

One concrete limitation worth putting in your rollout doc before anyone complains:

What no scope covers is what applicants said during screening, because Greenhouse never stores it.

If your sourcing team treats recruiter screens as the primary source of truth about a candidate's motivation, the MCP server will feel strangely blind. It can read notes. It cannot read the call that generated them.

Native 36 vs third-party 175

Vendor-neutral MCP endpoints are already beating vendor-native ones on tool coverage. The CData Python MCP Server for Greenhouse exposes 175 tools against the Greenhouse API: pipeline, candidate, analytics, and bulk HR operations. Greenhouse's own native server covers 36 in open beta. That is 4.8x the surface area, with none of the polish and none of the governance guarantees.

Big agencies will run both and route by task: native for anything a junior recruiter touches, third-party CData for a staff sourcer running overnight automations. Boutique shops will pick one and live with the gap.

The actual numbers, side by side

Here is the comparable data in one place:

The 5,220 US sourcers are the exact wrong buyer persona for pure chat sourcing. Reviewers already call SeekOut's web product a specialized instrument best suited to skilled sourcers, with inconsistent contact data and email-only outreach in the self-serve Recruit product. Those sourcers' value is Boolean craft and spreadsheet discipline. Chat abstracts both away and hands the sourcing primitive to the hiring manager, who is the real buyer MCP is built for. If you lead a sourcing team, that is the strategic shift to brief your director on this quarter, not next.

Rate limits are the new seat limits

The economic unit of sourcing is shifting from seats to tool calls, and nobody in procurement has caught up. SeekOut's 1,000 requests per user per day sounds generous until you watch an agent loop. One autonomous "find similar to this engineer, then rank by intent, then enrich contact details" chain can burn 50+ calls per candidate. A 20-candidate overnight run blows through the daily ceiling before breakfast.

Three practical consequences:

1. Agent design matters more than prompt design. A chatty agent will hit the ceiling on day one. A disciplined agent that batches queries and caches intermediate results will not.
2. Seat math stops working. The old question was "how many Recruiter seats do we need?" The new one is "how many tool calls per req, times how many reqs per quarter, times how many sourcers?" Vendors will price around it within 18 months.
3. Overage pricing is coming. None of the three vendors has published it yet. All three will.

What chat sourcing is actually good at

Chat sourcing wins at ambiguous, cross-source questions where a Boolean would take 40 minutes to write and a shortlist of 12 is more useful than a list of 400. The strength is not scale; it is translation.

Three questions a chat window handles well that SeekOut's dedicated UI handles poorly:

• "Who are the Rust maintainers in Berlin who also worked on cryptography projects at a FAANG before 2022?" (Multi-source join across GitHub, LinkedIn, and employment history.)
• "Among the 300 candidates in this Greenhouse pipeline, who has shipped a side project in the last 90 days?" (Cross-reference ATS data with the open web.)
• "Find me the five people most likely to leave Stripe in the next six months based on their public signals." (Soft inference the facet UI cannot encode.)

Those are the exact shape of questions Refolk was built for. The underlying thesis is the same as MCP's: natural language is a better query interface than a form when the question involves more than three dimensions. The difference is index. The MCP vendors expose their own data through Claude; Refolk reads across GitHub, LinkedIn, and the open web, which is where most senior ICs actually live.

Where it falls down, and what to keep in the dedicated UI

Chat sourcing is bad at anything requiring a list, a column, or a stage change at scale. Keep the following work in the vendor's native UI:

• Reviewing more than ~15 candidates at once. The chat transcript is a terrible grid.
• Boolean refinement with named facets. "Add 'kubernetes' to the required skills, remove companies under 50 employees" is three clicks in SeekOut, three paragraphs in Claude.
• Stage moves on more than a handful of candidates. Greenhouse's bulk-action UI is faster than any chat prompt will ever be.
• Any workflow a compliance team has to audit. Chat transcripts are not an audit log, no matter what the vendor tells you.
• Reading recruiter screen content inside Greenhouse MCP. It is not there to read.

The honest division of labor for the next 18 months: chat for discovery and reasoning, dedicated UI for review and action. The vendors shipping MCP know this. They are not trying to replace their own grids. They are trying to replace the five minutes you spend assembling a Boolean so hiring managers can skip straight to the output.

Chat replaces the query builder, not the results grid. Budget your sourcers' time accordingly.

Who else is in the race

Four named vendors have shipped recruiting MCP servers you can actually connect to today:

1. SeekOut MCP (seekout.com/solutions/mcp) - 14 workflows, 1B+ profiles, Claude / ChatGPT / Gemini / Copilot / Llama.
2. Greenhouse MCP (mcp.greenhouse.io/mcp) - 36 native tools, OAuth 2.0 with PKCE, Core/Plus/Pro, Claude / Claude Code / ChatGPT / Google Antigravity, plus Amazon Q, Copilot Studio, Glean, and Grok AI through the same OAuth path.
3. Manatal MCP Server - Enterprise Plus tier only, claimed first-to-market in recruiting, Claude and ChatGPT.
4. Recruit41 (Bengaluru) - shipped April 2026, framed by CTO Sripathi Krishnan as "headless intelligence," targeting Claude, Microsoft Copilot, and ChatGPT.

Notice who is missing. LinkedIn Recruiter has not shipped an MCP server. Neither has Workday, iCIMS, Lever, or Gem. The holdouts are mostly the incumbents with the most to lose from a hiring manager who can source without a Recruiter seat.

What to do this month

Three concrete moves if you lead a sourcing function:

1. Stand up Greenhouse MCP in a test workspace. It is included on Core/Plus/Pro. The marginal cost is zero and the learning curve is a week. Connect it to one Claude account, not fifty.
2. Instrument tool calls, not sessions. If you cannot see how many MCP calls your team made yesterday, you cannot see the 1,000-request ceiling until you hit it.
3. Decide where SeekOut's 14 workflows leave gaps for you. The MCP servers are tied to each vendor's own index, so a general-purpose people search that reads the open web stays in the stack for anything the 14 do not cover.

The chat window is not replacing your sourcing stack. It is replacing the five most tedious minutes of every search in it. The teams that notice the distinction first will spend 2026 shipping more roles with the same headcount.

FAQ

What is an MCP server, in one sentence?

Model Context Protocol is Anthropic's open standard for letting an AI assistant call tools on a remote server using the signed-in user's credentials, and an MCP server is any endpoint that implements the protocol so Claude, ChatGPT, Gemini, or Copilot can read and write to it on the user's behalf.

Is Greenhouse MCP an extra line item?

No. Greenhouse MCP is in open beta and included at no extra cost on Core, Plus, and Pro plans. The native server covers 36 ATS tools, with the connection inheriting the signed-in user's existing Greenhouse permissions through OAuth 2.0 with PKCE. The thing that will cost you is tool-call volume if you build autonomous agents on top.

Can I just use the CData third-party Greenhouse MCP instead?

You can, and you get roughly 4.8x the tool coverage (175 tools vs the 36 in Greenhouse's native server) by doing so. The tradeoff is governance and support. Greenhouse's native server has permission inheritance baked in and is backed by the vendor. The third-party build gives you raw API breadth but puts the responsibility for scoping, auditing, and incident response entirely on you.

Will MCP servers replace tools like SeekOut and LinkedIn Recruiter?

Not in 2026, and probably not in 2027. MCP replaces the query-building step of sourcing, which was five to ten minutes of a search. The results view, the stage-move UI, the audit log, and the bulk operations all still live in the dedicated app. The vendors shipping MCP know this; they are using chat to lower the barrier to entry for hiring managers, not to kill their own grids.