Every capability is a command your agent can call. Postings collect applications, assessments score themselves, automations move the routine motion. Humans make the decisions. Software does the rest.
claude mcp add --transport http hiring-orchestrator https://www.hiringorchestrator.com/api/mcpAuto-on
Legacy ATSs are databases that wait for clicks. Orchestrator closes the loop: a published posting feeds stage one directly, the native assessment sends and scores itself, and automations you set in one sentence handle the routine motion. What reaches you is the part that deserves a human.
Application received on the hosted board
j.lindqvist → Applied
Assessment invitation sent automatically
30-min multimodal · HiringTest
Scored. Composite 86, scorecard attached
fraud-checked · domains on record
Automation: score 80+ → advance to Interview, notify the team
undo handle t_4c1 · audit-logged
You: “prep me for the 2pm with Maya”
briefing: background, domain gaps, prior feedback
When you want eyes on it
Drag-and-drop board, candidate profiles, scorecards, settings. Every card shows who moved it: a person, an agent, or an automation. The UI calls the same 62 commands your agent does, which is the point. Headless means the interface is optional, not absent.
How it works
Add the connector in Claude or ChatGPT, or run one command in Claude Code. OAuth handles the rest. No keys to wrangle, no seats to provision.
Walk in with your whole pipeline. Promote your Greenhouse, Ashby, or Gem state into Orchestrator: jobs, candidates, stages, and history.
Postings collect applications. Assessments send and score themselves. Automations move the routine motion and tell you what needs a human. You decide; it executes.
Agent-native, not chatbot-bolted
“Headless” here is 62 typed commands over the Model Context Protocol, with a Claude Skill that teaches the playbooks: triage, weekly review, req-to-offer, migration day one. Reads are instant. Writes are graded: routine moves carry an undo handle, and anything destructive asks first.
Natively connected
An assessment stage that runs a scored, fraud-checked, 30-minute multimodal assessment in-process. The composite and scorecard land back on the application automatically. Graded auto-advance moves strong candidates forward, and never rejects anyone for you.
When a candidate has a published case, it surfaces on their ATS profile: verified endorsements and demonstrated work, candidate-controlled. The candidate-side network the incumbents can't offer.
Questions
The product is a capability layer, not an app. One capability layer powers the MCP server your agent calls, the REST API, the in-product copilot, and the UI. The interface is whichever one you're already in. Most ATSs added an API to a UI; this is built the other way around.
The routine motion runs without you touching it. A published posting collects applications straight into stage one. An attached assessment sends itself and scores flow back automatically. Automations advance, tag, note, and notify on triggers you set. The only thing that never automates is the decision: nothing here rejects a candidate on its own.
It's the inverse. The ATS is a set of typed capabilities, and the agent is the front door. Every action a recruiter takes in a legacy ATS is a tool an agent can call, scoped to your company, audit-logged, and reversible.
No. There's a pipeline board, job and candidate views, and settings for the moments you want hands on the records. But the day-to-day runs from the agent you already live in.
Only upward, and only past thresholds you set. Stage moves carry an undo handle, destructive actions require confirmation, offer approvals are bound to the named approver's own account, and it never auto-rejects. We suggest, and you decide.