If your recruiting team runs on Greenhouse and you have spent the last six months trying to glue an AI agent to it, this post is for you. The agent itself is the easy part. The hard part is everything around it: pulling the right candidates without burning through your Harvest API rate limit, writing scorecards back into the right interview kit, advancing applications through stages that your team renames every quarter. It is a lot of plumbing for what should be a five-minute job.

Ribbon MCP is the shortcut. It is a connector that gives any AI agent live, structured access to your Greenhouse data, and lets the agent write results back without anyone hand-rolling a Harvest client. Plug it in once and Claude, ChatGPT, Cursor, or whatever agent you are running can read your pipeline, score candidates, and update applications inside Greenhouse without leaving the conversation.

What MCP actually is

MCP stands for Model Context Protocol. It is an open standard, introduced by Anthropic in late 2024, that lets AI models talk to outside systems through one consistent interface, the way USB lets any laptop talk to any printer. The agent does not need to know that your ATS is Greenhouse, that authentication uses a Harvest API key, or that the candidate stage you call "Recruiter Screen" is a job stage attached to a particular job. It asks the MCP server in plain language and gets back structured data.

For Greenhouse specifically, that abstraction earns its keep fast. Greenhouse has one of the cleaner ATS APIs out there, but "clean" still means you are juggling Candidates, Applications, Jobs, Job Posts, Job Stages, Scorecards, Scheduled Interviews, and Custom Fields, and each one has its own pagination, filtering, and update semantics. MCP collapses that surface into a small set of tools the agent can reason about.

What gets connected

There are two halves to a real ATS integration: what the AI can see, and what it can change. Ribbon's Greenhouse connector covers both, scoped to whatever your Greenhouse permissions allow.

On the read side, the agent gets live access to:

• Jobs and job posts: open requisitions in your Greenhouse instance, filterable by status, name, or department. The agent can see what is being hired for right now, not what was true at the time of the last sync.
• Candidates and applications: the full pipeline, with each application carrying its current stage, the linked candidate, source, custom fields, and any scheduled interviews. Greenhouse models a candidate as a person and an application as that person against a specific job, so the agent can pull a candidate's full history across roles in one shot.
• Scheduled interviews and interview kits: upcoming interview slots, who the interviewer is, which interview kit applies, and which scorecard attributes the rubric expects.
• Custom fields: whatever your Greenhouse admin has defined on candidates, jobs, or applications shows up unchanged. If your team tracks "Visa Status" or "Compensation Floor" as custom fields, the agent can read them.

On the write side, the agent can:

• Create new applications against an existing job, with candidate name, email, phone, and resume attachments.
• Move an application between job stages. Stage matching uses the actual stage names on that specific job, so the connection survives the routine renaming that every TA team does mid-quarter.
• Attach a structured assessment result to an application. Ribbon writes the interview score, summary, and a link to the full transcript and recording, plus a breakdown across attributes like motivation, communication, role-specific skills, language metrics, background fit, logistical fit, tools and skills coverage, and the candidate's own questions for the team.
• Trigger downstream Greenhouse automation. Because the move and the result attachment happen inside Greenhouse, any existing automation rule keyed off a stage change (notifications, scheduling triggers, scorecard reminders) fires the way it always does.

Authentication is a single Harvest API key, scoped to whatever permissions you grant. There is no proxy, no shared inbox, no separate user account that ends up looking like a person on your team page.

Five workflows that actually pay for themselves

The point of all this is not to demo MCP. It is to delete recruiting work that should never have been manual. A few of the workflows that show up most often once a team plugs Ribbon into Greenhouse:

1. End of week pipeline review. The agent reads every application that moved into "Recruiter Screen" in the last seven days, summarizes the cohort by source and seniority, and flags candidates that have been sitting in the same stage for over ten days. The output is a Slack message your head of TA actually reads, not a dashboard nobody opens.

2. Auto-attach interview results. When a Ribbon AI screen finishes, the agent locates the right candidate and application in Greenhouse, attaches the score and transcript, and moves the application to the next stage if the score clears whatever bar the team set. The recruiter sees the result inside Greenhouse, on the same screen they already use, with no copy paste.

3. Sourcing on real-time data. The agent looks at currently open requisitions in Greenhouse, then drafts personalized outreach for candidates already in your database who match. Because the read is live, it never pitches a candidate for a job that closed yesterday.

4. Stuck pipeline reports. Every Monday the agent pulls applications older than fourteen days in stages like "Take-Home" or "Onsite," cross references them with custom fields like priority or hiring manager, and emails the right person a per-job nudge. Greenhouse can do some of this through stored reports, but stored reports do not write follow-up emails.

5. Recruiter copilot. A recruiter pings the agent, "what is our pipeline for the Senior Backend role and which candidates are slowest to hear back," and gets a one-paragraph answer with names, stages, days idle, and links straight back into Greenhouse. The recruiter never opens a separate tool.

What setup actually looks like

The setup overhead is small. A Greenhouse admin generates a Harvest API key with the relevant Candidates, Applications, Jobs, and Scorecards permissions. They paste it into Ribbon's integrations page. From there the agent inherits whatever scope that key has, no more and no less. Ribbon does not store candidate data outside what is needed to maintain the connection, and the connection is read mostly, with writes bounded to the operations listed above.

If your team uses Ribbon for AI screens already, this is a one-time additive change. If you are starting fresh, you can connect Greenhouse first and bring agents online afterward.

FAQ

Does Ribbon need a separate Greenhouse seat? No. Ribbon authenticates with a Harvest API key, not a user license. The actions the agent takes show up in Greenhouse as API actions, attributable to the integration.

What stops the agent from moving the wrong application to the wrong stage? Two things. First, the stage transition is name-matched against the actual job stages on that specific job, so an agent cannot move an application to a stage that does not exist on its job. Second, write actions are scoped behind explicit tools, so an agent reading data cannot accidentally mutate it.

How does this interact with Greenhouse's own automation? Cleanly. When Ribbon advances an application, Greenhouse treats the change like any other stage transition. Any rule, notification, or scheduled email tied to that stage fires normally.

Can we restrict what the agent can write? Yes. The Harvest key controls the ceiling. If you grant read-only Harvest permissions, every write tool fails closed and the agent can still answer pipeline questions.

Does this cover Greenhouse Onboarding or only Recruiting? Today the connector focuses on the Recruiting product, which is where AI agents save recruiters the most hours. Onboarding objects are on the list but not the priority yet.

What happens to scorecards and interview kits? Ribbon writes a structured assessment record back to the application, clearly identified as the Ribbon AI result and linked to the full transcript and recording. It does not overwrite human scorecards; the assessment lives alongside them.

What is next

The Greenhouse connector is one of the most-used surfaces Ribbon supports, and the next round of work is on the write side: richer scorecard mapping so AI assessments can populate scorecard attributes one-to-one when teams want them to, plus optional auto-rejection flows for candidates that fall below a threshold the recruiter sets. If you have a workflow you want the agent to handle and it is not covered here, that is the kind of thing that goes into the next release.

If you want to try this on your Greenhouse instance, head to ribbon.ai and connect from the integrations page. The setup is genuinely a couple of minutes, and the first useful workflow tends to show up the same day.