The Pulse Recruit MCP server

Who the agency already knows for a role: ranked people from the ATS and from the recruiters' LinkedIn networks, each with their reasons and which pool they came from. The one answer nobody else can give. Name a job_id, or a posting_id for a role a company is advertising. The consent filter applies. Costs at most the title-widening call; a repeat on the same title is free.

Since 2026-09.

Colleagues already connected to people at a company, with the relationship and who holds it. The consent filter applies: a person whose consent has lapsed is not returned. Give a company_id, an outbound_company_id, or a company_name.

Since 2026-09.

People NOT in the book, from the sourcing vendor, read and scored against the role, staged for a recruiter to accept in Pulse. Nobody becomes a person on the book until a recruiter accepts them. Always over the confirmation threshold: say the cost and get a yes first. Run search_the_book before this.

Since 2026-09.

One person: who they are now, their history on the agency's roles and their interview scorecards, their consent basis, what has happened with them recently, their upcoming interviews, and any records that may be the same person (merging happens in the product). Name exactly one of person_id (the ATS) or outbound_person_id (the recruiter's network). Ids come from search, search_the_book or who_do_we_know.

Since 2026-09.

One company: sector, size, whether it is tracked, its open roles with the agency, and who the agency already knows there. Name exactly one of company_id or outbound_company_id.

Since 2026-09.

One role: the client, how long it has been open, its stages with how many people are in each, the titles the book is searched under for it, and its upcoming interviews (booking one happens in the product).

Since 2026-09.

Find people, companies and roles by name, for when you have words rather than an id. People are also found by what their CV says, each such match with the line that matched (cv_line). People come from both pools with the consent filter applied, so a person whose consent has lapsed is not returned. Use the ids in get_person, get_company and get_job.

Since 2026-09.

The actions due now across the agency's campaigns, oldest first, grouped by where the work happens (LinkedIn, email, phone). Each is a person, the campaign, and the step. Also the recruiter's own tasks due by the end of today, as a separate list. Nothing here sends: the recruiter does each one in Pulse, and Pulse records it.

Since 2026-09.

People who replied and have not been answered, newest first: the thread, who they are, and the last thing they said. The most valuable list in the product.

Since 2026-09.

Advertised roles at tracked companies that match what the agency places: the function, the level and the place the desk has set. Newest first. Refuses until the desk's niche is set in Settings, because without it every posting would match.

Since 2026-09.

What a tracked company has done lately: funding, hires, departures, postings, news. Refuses for a company that is not tracked, because nothing is polled for one.

Since 2026-09.

How the campaigns are doing over a window, counted in people: contacted (any send in the window), accepted (a connection request accepted), replied (a reply, or a reply label a recruiter set), interested, meetings booked, roles opened from replies and placements on them, each later stage inside the one before. An accepted invitation is not a reply. Per live campaign, with its people by status now, and the workspace's total.

Since 2026-09.

A drafted message for a named person, in the org's style, from the facts Pulse holds and nothing invented. A draft, never a send: it lands where the recruiter reads it before anything goes. Give an enrolment_id for a person in a campaign, or a person id with a job_id for anybody else.

Since 2026-09.

Add people to a campaign. Every exclusion rule, the do-not-contact list and the off-limits register apply unchanged, and the answer says who was added, who was already in, and who was skipped and why. Nothing is sent: the first action appears in whats_due_today for a person to do. Refused for a draft or archived campaign.

Since 2026-09.

Record what a reply meant, as the recruiter's reading of it. Interested, not now (with a date), a referral, not interested, an objection, or an auto reply. It changes what happens next in the campaign, so quote what they said in why.

Since 2026-09.

Start marketing a candidate, free: the net starts from the candidate (companies like their employers when asked), and once judging starts the run adds the companies with jobs open now for their titles, the Dream 100 companies and market builds that hire for them, and a hiring search; minus every company the candidate worked for, applied to or is in process with, recruitment agencies, the do not contact list, and companies whose every open role is outside their area. Returns the counts, what was left out and why, and the projected cost of judging and drafting in credits. Nothing is judged and nothing is spent; market_candidate does that with this run_id and projected_credits.

Since 2026-09.

Run the market for a candidate inside a budget (the projected credits, reserved once): every company is judged against the candidate's characteristics from its own postings and signals, the proposal (fits, maybes with a live role, close matches) is ticked on the shortlist, up to two people a shortlisted company are found by the workspace's seniority ladder (those with a verified address and a clear greeting approved by themselves), and an email, a LinkedIn message and a connection note are drafted for each approved person. It stops at review: no message is ever sent until a person approves it on the run page. Always asks first with the projected cost.

Since 2026-09.

Save a run as a campaign in one transaction: a list of the approved people at the shortlisted companies and the campaign marketing the candidate with the Market a candidate sequence. Saving again adds anybody approved since. The campaign opens in draft, and each message sends only once a person has approved it. Always asks first.

Since 2026-09.

Turn an interested reply on a campaign that markets a candidate into a role: a job at the replier's company, created as a prospect when the company is not on the book, with the marketed candidate on its first stage and the reply linked to it. Needs the candidate's OK to be named to that company, which is recorded. Admins only, as every role is. Always asks first. Nothing is sent.

Since 2026-09.

Add a note about a person or a role. It appears in the activity stream marked as written by an external assistant, beside the notes people wrote. Name at least one of person_id, outbound_person_id or job_id.

Since 2026-09.

Create a task for the recruiter: a reminder or to-do of their own. It lands in their task list marked as created by Claude, private to them unless it is about a person or company. Nothing is assigned to anybody else and nothing is sent.

Since 2026-09.

The agency's candidate tags, such as Python, Hot or Met at event: each tag's name, colour and how many people carry it, most used first.

Since 2026-10.

Add or remove tags on one person on the ATS, by name. A new name becomes a tag, as it does in the product. Up to 10 names a call. Tagging many people at once stays in the product.

Since 2026-10.

Refresh what Pulse holds about a person from the vendor path the sourcing panel uses. For a person in the network that is their LinkedIn profile, read once a day at most, through the agency's own connected account. Refused when their consent has lapsed.

Since 2026-09.

Read a prospect's LinkedIn company page and keep a dated observation of its headcount beside the record, leaving the record itself as it was. One observation per company per day.

Since 2026-09.

Fill empty fields on a person Pulse already holds, from evidence you can cite. A field that already has a value is left alone whoever wrote it, a fact with no evidence is skipped, and an email or a phone is never accepted here. The answer says what was filled and what was skipped and why.

Since 2026-09.

Start tracking a company: its signals are polled weekly from now on and appear in company_signals. Refused by the database past the plan's tracking allowance. Use set_tracking_tier for daily.

Since 2026-09.

Move a tracked company between weekly and daily polling. Daily is a small allowance, refused by the database when it is full. A company not yet tracked is tracked at the tier you name.

Since 2026-09.

What is tracked, at which tier, what fired in the last week, and how much of the tracking allowance is used.

Since 2026-09.

Store a market definition once the person has confirmed it: the ICP paragraph with its exclusions, a headcount band, places, the seed accounts the client likes, and eight to twelve slices with real industry picklist values and thematic keywords. The first call refuses with the whole definition in the confirmation; the retry with the id and identical content writes. Give id to update, from_id to start from an existing market. Refuses while the desk niche is unset.

Since 2026-09.

Every market definition the workspace has saved: name, paragraph, exclusions, seeds, slices, who confirmed it and when, and the builds made from it. Pick one by name for the next build rather than redrafting.

Since 2026-09.

Retire a saved market definition. Its builds and the companies on the book are untouched. Refused while a build on it is in progress.

Since 2026-09.

Build a Dream 100 from a saved definition: every slice pulled from every configured source, deduplicated by domain, relationship accounts set aside, each company judged FIT, MAYBE or NOT against the ICP and the desk, and the best hundred kept. Always asks first with the projected cost, which is a ceiling. Runs unattended for minutes to an hour; ask tam_status for progress. The full market behind the hundred, usually five to twenty five thousand companies, is a service the RecruiterGTM team builds by hand and is asked for in the product, not here.

Since 2026-09.

Where a build is: the stage, rows in and out, spend so far against the projection, which sources it used and which were missing, what it was built from, and how many companies wait on a person's verdict.

Since 2026-09.

How big a market is and how it segments by slice, sector, size and place; the overlap between sources; twenty FIT companies sampled at random with their reasons, for the SOP's spot check; the companies waiting for a person's verdict, the first 300 by name with the total; and the relationship accounts set aside. Counted over every company in the build, however large.

Since 2026-09.

Record a person's decision on a company the model could not decide. The verdict becomes fit or not, stamped as the recruiter's, and no later model verdict overwrites it.

Since 2026-09.

Put companies from a finished build onto the Dream 100, the set Pulse watches for signals, and take others off in the same press. The list holds a hundred: the database refuses the whole call, writing nothing, when what is coming in does not fit what is going out, and says how many more to release. A company already watched is counted, not re-added. A company with no domain cannot be watched at all and is reported. Only a qualified company joins: one still for review joins once settled as a fit, and anything else is reported as not qualified.

Since 2026-09.

Every lead list in the workspace: its source (hand-picked, people who match, buyers at companies), the rule in prose, how many people are on it, when it was refreshed, and which campaigns point at it. Also what a new rule may use: the finished market builds, the signal kinds and job functions that exist, and the approved buyer titles.

Since 2026-09.

How many people a rule would put on a list, with a sample of forty and the skips, without saving anything. Run this, show the person the count and a few names, then save_lead_list once they agree. Thirty a minute.

Since 2026-09.

Create a lead list from a rule, or change the rule of one, and refresh it. Pulse keeps it current: every campaign pointed at it follows on the fifteen-minute sweep. The first call refuses with the rule in prose; the retry with the confirmation id and identical arguments writes. Free.

Since 2026-09.

The recruiter's open tasks, their own reminders and to-dos, soonest first: title, due time, priority, repeat, list, and the person or company it is about. Not campaign actions; those are whats_due_today.

Since 2026-09.

Mark one of the recruiter's tasks done, as them. A repeating task makes its next occurrence. Nothing is sent and nothing else changes.

Since 2026-09.

Every candidate pool in the workspace: the agency's own people kept for future roles, built from stacked filters (CV words, stage reached, client, silver medallists, consent and so on) with people filed from roles or added by hand. Each with its filters in prose, how many people are in it and when it was refreshed. get_person names a person's pools.

Since 2026-09.

Research briefs, candidate profiles and submission packs, newest first, each at its latest version: the kind, title, version, whether it is frozen (shared or sent, so it no longer changes), and for a profile whether a newer CV is on file than the one it was built from (stale). get_artifact reads one.

Since 2026-10.

One version of a research brief, candidate profile or submission pack: its body as Pulse keeps it and the sources it rests on, each with its type (cv_file, profile_field, answer or search), label, the line it quotes and any link.

Since 2026-10.

How the client answered on each candidate in a pack: per recipient the pack was sent to, interested or not for us, any note they left, when, and whether somebody has acted on it in Pulse. Candidates nobody has answered on are listed as unanswered.

Since 2026-10.

Keep research you did as a research brief in the agency's library, version 1, frozen, with the sources it rests on. Refused without at least one source. It can then be shared from Pulse like any brief.

Since 2026-10.

Ask Pulse to build, or rebuild, a candidate's profile from their CV in the agency's house format. Pulse's own builder reads the CV with the person's name and contact details removed, and keeps only lines it can quote from the CV; the current employer is described, not named. About 2 credits for a text CV, more for a scan. With no CV it is built from the record and costs nothing. Returns the profile's id for get_artifact.

Since 2026-10.

Start a submission pack for a role from candidates on it, in the order given. Returns the pack's id, who still needs a profile (build_pack_profiles) or a pitch (draft_pack_pitches), and anything that would stop the pack going out, per candidate (off limits, no lawful basis to share). Nothing is frozen or sent: a person finishes and sends it in Pulse.

Since 2026-10.

Build a profile for each named candidate on a pack who has none yet, through Pulse's own builder, one at a time. About 2 credits each for a text CV. Reports honestly: what was built, what it cost, and who failed and why.

Since 2026-10.

Draft the pitch, why the agency puts them forward, for each named candidate on a pack. Pulse drafts it from the candidate's profile and the role's brief, and keeps it only if every line it quotes is in the profile; it is saved on the pack for a person to edit. About 1 credit each. Replaces a pitch already there.

Since 2026-10.