Most tools staple a chatbot onto the sidebar. FairHire ships an 200-tool MCP catalog so your agent operates the real hiring pipeline — the same permission-checked, audited surface a recruiter uses. It sources, screens, scores trials, and reads the numbers. The decisions that matter stay human.
# discover what this key may do
› fairhire.whoami()
{ org, role: "agent", permissions: […] }
# reads are broad
› fairhire.workflow.needs_attention()
› fairhire.trial.score({ trialId, score: 9.2 })
→ composite: 9.2 × .50 = 4.60
# the decision is a human gate
› fairhire.stage.advance({ to: "offer" })
✋ human_gate — post a note, @mention a human
› fairhire.notes.add({ mentions: ["@lena"] })Every tool is a thin, typed wrapper over the exact service function the human UI calls. Permission checks, org-scoping, and actor-tagging live in the service — so an agent can never reach a code path a recruiter couldn’t.
Composite scores, funnel metrics, and bottlenecks come from the one shared scoring the dashboard uses. The number your agent reads is the number your committee sees. No drift, by construction.
Agent writes are tagged as agent actions and land in the same activity feed, stage history, and immutable hash-chained audit log as human ones. The fairness trail doesn’t care who acted — only what happened.
The fairhire.* catalog covers every stage of hiring, from
sourcing a passive candidate to assembling a fairness page. Each tool is
input-validated and returns structured results your agent can act on.
Open roles as the agent’s unit of work — with live health, pipeline config, and hiring teams.
Browse, search, create, dedupe, and merge. Resume parsing is local and $0 — candidate data never leaves the stack.
Source passive candidates, tag, pool, filter, dedupe, and re-engage — email-deduped, DNC honored.
Move candidates through exit gates. A blocked gate returns the missing requirement — the agent never force-moves.
Weighted criteria roll into one composite via the shared scoring — the same number the UI shows.
Scope, track deliverables, score the real work, and record payment. The heaviest signal in the composite.
Draft rubric-mapped kits, propose slots, schedule, attach transcripts, and score panels.
Notes, threaded comments, and blind-by-default reviews. Agent actions are badged in the feed.
Templates, messages, and multi-step sequences — with per-candidate contact preferences respected.
Draft, read, and template offers. There is deliberately no tool to approve or send one — that stays human.
Assemble and preview a fairness snapshot freely; publishing and external grants are human-gated.
Whoami, dashboard stats, the same funnel/bottleneck/conversion numbers as the UI, and the immutable audit chain. Fairness stats are aggregate-only.
The agent can do almost everything a recruiter can — but the
consequential decisions are hard-gated in code. A default operator key
hitting one of these gets a human_gate error and is told
to post a note and @mention a human, not to retry. There is no flag
that lets an operator agent override them.
A key is minted as fairhire_sk_… and shown a single
time. Only its SHA-256 hash is stored — the plaintext is never
recoverable, even by us.
Every MCP call re-resolves the key into a fresh, org-scoped actor context. Revoke or expire a key and the very next call fails — no session to wait out.
A key is bound to one organization and one role. It never sees another workspace’s reqs, candidates, or tags — org-scoping is enforced in the service on every read and write.
Key issuance and role management are admin-only tooling — never exposed as agent tools. Elevated permissions (publishing, DEI reads, audit reads, gate overrides) are opt-in, not the default.
Mint a scoped key, point your MCP client at FairHire, and let it run the funnel while the decisions stay yours.