The Greenhouse MCP server is official, it reads and writes, and as of October 2026 it's still in open beta. Ashby launched its own in late June and got listed in Claude's connector directory. If you're a CTO who hires through either ATS, this changes what your assistant can do on a Tuesday morning: pull a pipeline, summarize interview feedback, move a candidate to the next stage, all from a chat window.
I've spent the last few months on the other side of this, building an MCP server for a staffing company, so I read these launches with one question: what's safe to hand to a model, and what isn't?
Key Takeaways
- Greenhouse announced its MCP on May 7, 2026, with rollout from June. The help center says it is "not limited to read-only access."
- Ashby launched its MCP server on June 29, 2026 in open beta on all plans, and it's listed in Claude's directory.
- Both are beta. Plan for tool names and scopes to change.
- Read access is low risk and high value. Write access needs scoping, an owner, and someone checking what the assistant actually did.
What the Greenhouse MCP server does
Per Greenhouse's help page, it supports read and write actions, connects to clients including Claude, ChatGPT, Copilot Studio, Glean and Amazon Q, and uses OAuth with a site admin enabling scopes in the Dev Center. It's available on the Core, Plus and Pro plans; legacy tiers have to ask their account team about access. Greenhouse frames the launch as "a governed way" to connect AI tools, and the governance is real: an admin decides which scopes exist before any user connects.
It isn't in Claude's connector directory as of this writing. You add it as a custom connector. That's a small thing, but it means your team needs the server URL and admin approval, not a one-click install.
What the Ashby MCP server does
Ashby's server reads candidates, applications, jobs, interviews, offers and feedback, and it can take a few actions: add notes, move stages, consider a candidate for a job. Its directory listing also says it can send emails, which is the first permission I'd lock down. The directory listing shows tools such as `get_candidate`, `get_job_pipeline`, `get_submitted_feedback`, `change_application_stage` and `create_candidate`, at `mcp.ashbyhq.com/mcp/v1`, and marks it "Anthropic verified." Auth is per user, so the assistant sees what that user can see.
| Greenhouse MCP | Ashby MCP | |
|---|---|---|
| Status (Oct 2026) | Open beta | Open beta |
| Announced | May 7, 2026 | June 29, 2026 |
| Reads | Yes | Candidates, jobs, interviews, offers, feedback |
| Writes | Yes, per enabled scopes | Notes, stage changes, create candidate, emails |
| In Claude's directory | No, custom connector | Yes |
| Plans | Core, Plus, Pro | All plans |
| Who controls access | Site admin enables scopes | User-level OAuth |
What about Lever and Workday? I couldn't find an official Lever MCP server, only third-party ones. Workday made its MCP "Agent-Ready Tools" generally available on October 13, 2026, but the release doesn't say recruiting data is included, so I wouldn't assume it.
What I'd let an assistant do, and what I wouldn't
Yes, from day one: summarize feedback across interviewers, list candidates stuck in a stage for more than five days, draft the debrief agenda, compare a scorecard to the job's requirements. All reads. All things a hiring manager does badly at 6 p.m.
Yes, with a person watching: add a note, move a candidate from "application review" to "phone screen." Cheap to undo.
No: rejecting candidates, creating offers, or anything that sends a message to a candidate. A wrong stage move is a click to fix. A rejection email sent to your best applicant is not. If you use AI in decisions, also read our AI hiring compliance checklist; some rules attach to automated screening.
There's a quieter risk too. Candidate data is personal data. Once it flows through an assistant, your vendor's retention terms apply to it. Check those before a recruiter pastes 40 resumes into a chat.
A Concrete Version
A 50-person B2B company runs Ashby and has four engineering reqs open. The head of engineering connects Claude with the Ashby connector on Monday.
- 9:00. "Which candidates have been in technical interview for more than a week, and what did interviewers say?" The assistant calls the pipeline and feedback tools. Answer in about 30 seconds: 6 candidates, 2 with split feedback.
- 9:10. "Add a note to both split ones: schedule a calibration call with both interviewers." Two write calls.
- 9:15. Two reqs have no viable candidates after three weeks. The assistant drafts a brief for a staffing vendor. The head of engineering sends it.
Saved: maybe 45 minutes of clicking. Not saved: the judgment on the split candidates, which still takes a calibration call. If the vendor route makes sense, you'd see blind profiles within 72 hours and meet the engineers after the agreement is signed. Our post on MCP vs API for AI agents covers how a vendor's side of this connection works.
The Honest Counterpoint
Beta software in your hiring system of record is a real risk. Tool names change, scopes change, and a model that confidently moves the wrong candidate creates a mess your recruiter finds a week later. If your team is three people and you hire twice a year, the ATS UI is fine and the connector is a toy.
And none of this improves candidate quality. An assistant reading a pipeline full of inflated resumes produces a fast, well-formatted summary of inflated resumes. The vetting still has to happen somewhere. AI candidate screening covers what automated tools catch and miss.
Frequently Asked Questions
Does Greenhouse have an official MCP server?
Yes. Greenhouse announced it on May 7, 2026, and it's in open beta. It supports read and write actions, with scopes enabled by a site admin, and connects to Claude, ChatGPT and other MCP clients.
Is the Greenhouse MCP server in Claude's connector directory?
Not as of October 2026. You add it as a custom connector. Ashby's MCP server is listed in the directory.
Can an AI assistant reject candidates through the Greenhouse MCP server?
Depending on the scopes your admin enables, write actions are possible. I'd keep rejections, offers and candidate emails with a person. Use the assistant to read and prepare, not to decide.
The Bottom Line
The Greenhouse MCP server and Ashby's connector make your ATS readable by an assistant, which is useful today for summaries and pipeline hygiene. Turn on reads, add writes slowly, keep decisions human. And when the pipeline itself is the problem, request a vetted shortlist or see available engineers.
Roberto Espinoza is CEO of Ruzora, which helps US startups hire pre-vetted senior LATAM engineers, with a vetted shortlist in 72 hours. See available engineers.
