Reference file

Connector zime

connector-zime.md

Connector profile: Zime

The Zime MCP connector at https://mcp.zime.ai/mcp. Users sign in with their Zime work email; every tool returns only what that person can already see in the Zime dashboard.

Seven tools. Five are free lookups, two run Zime's agents and use credits.

ToolJobCredits
list_meetingsResolve one meeting or recorded call, or browse themno
get_transcriptThe full verbatim transcript of one callno
list_dealsResolve one deal, or browse dealsno
list_accountsResolve one CRM account, or browse accountsno
get_contactResolve one CRM contactno
ask_zimeZime's agent: any question over calls, deals, signals, strategyyes
prep_noteZime's prep-note agent for one upcoming scheduled callyes

Resolve tools

All five return the same envelope:

  • {"status": "resolved", "data": {...}}
  • {"status": "multiple_matches", "candidates": [...]} → show them with the

fields that tell them apart, ask, re-call pinned by id. Never guess.

  • {"status": "no_match", "candidates": [...]} → not proof it does not

exist; usually the window excludes it or it is named differently. Show near-misses, offer to widen. Never substitute a similar record.

  • {"error": "<CODE>"} → INTERNAL_ERROR retry once; UNAUTHORIZED /

FORBIDDEN re-authorize or lack of access; INVALID_ARGUMENT fix the arguments.

list_meetings

Covers recorded calls and calendar meetings, upcoming or unrecorded. A row carries meeting_id, call_id and has_transcript; either id may be null, and that is information: call_id: null is exactly why no transcript exists.

  • query — company, attendee or topic. **Time words go in the date

fields, not here.**

  • call_id / meeting_id — pin from a prior candidates response.
  • start_date / end_date — YYYY-MM-DD. Defaults: 90 days ago to 14 days

ahead, so the window looks both back and forward.

  • recorded — true recorded only, false upcoming or unrecorded only,

omit for both. Set true when the next step needs a transcript.

  • list + limit — browse recent or upcoming instead of resolving one.

get_transcript

Returns the entire transcript, verbatim. Large payload; a timeout is a real possibility, so if it fails twice say so rather than filling in from memory.

  • call_id from list_meetings, or query plus start_date / end_date

(defaults 90 days ago to today).

  • This is the only source of quotable text. Anything in quotation marks

anywhere must have come from here.

list_deals

  • query — deal or account name. deal_id to pin. account_name to narrow.
  • start_date / end_date — deals are not date-scoped by default; a deal

that closed two years ago is a legitimate subject. Only pass dates if the user bounded it; a stray range hides the deal entirely.

  • list + limit — browse recent deals.
  • Questions about many deals (filters, totals, "what's stuck") are not for

this tool. Send them to ask_zime.

list_accounts

  • query — company name. account_id to pin. domain narrows better than a

name. list + limit to browse.

  • A company not in CRM is not a lookup; that is web research.

get_contact

  • query — name or email, e.g. "Jane at Acme". contact_id to pin.

account_name to narrow.

Agent tools

ask_zime

One argument, question. The agent has no memory of the chat and no separate filter or date parameters, so everything belongs in the question text: the record by name, the window, the segment, the rep, the lens, the output shape. Resolve every pronoun.

Answers descriptive questions (what happened, what was said, who attended, what are my action items) and prescriptive ones (how do I win this, why am I losing to X, what should I do next, what is working in won deals). Sees analysed call signals, CRM state and the playbook together.

Not for schema, SQL or how-the-data-is-stored questions. When a request matches a fixed-purpose tool exactly, use that tool instead.

Ask for the evidence explicitly. A risk call without a reason is not reviewable; the agent supplies reasoning when the question asks for it.

Skill rule: writing the recap, analysis or review yourself from raw records while ask_zime is available is a failure of the skill. The agent sees the extracted signals and CRM linkage a raw transcript does not carry.

prep_note

Only for preparing for one specific upcoming, scheduled call. Both conditions required. Anything about past calls, general deal strategy, or "what meetings do I have" goes to ask_zime or list_meetings.

  • query — the user's request minimally rewritten, keeping the company or

person and any time hint. Unlike every other Zime tool, time words belong here. Never invent a company or time that was not said.

  • calendar_event_id — only on a follow-up call after the tool returned a

candidate list and the user picked one.

Statuses beyond the note itself: a candidate list (show it, pick again), "no upcoming external meetings", "prep notes aren't configured for your workspace yet" (ask the Zime admin), "already being generated" (wait a few seconds, call once more).

Which tool for which job

The user wantsResolve withAnswer with
When is / who was on one meetinglist_meetingsstop, return the record
What happened on one calllist_meetings (recorded)ask_zime, naming the call and date
Exact words from one calllist_meetingsget_transcript
One deal's fieldslist_dealsstop, return the record
One deal's risk, strategy, stakeholderslist_dealsask_zime, naming the deal
One account's fieldslist_accountsstop, return the record
Many deals, pipeline, "what's at risk"noneask_zime with window and lens in the question
Themes across many callsnoneask_zime
Prep for an upcoming calllist_meetings, forward windowprep_note with the meeting_id
A person's role, email, accountget_contactstop, return the record
A company not yet in CRMnoneweb research, not Zime

House rules every Zime skill states

  • Answering from memory or chat history while the tool is available is a

failure of the skill. A remembered stage, amount or quote can be stale or wrong.

  • Relay agent output as-is. Do not re-rank, merge, condense, round or add. The

ordering is part of the judgment; "$47,300" stays "$47,300".

  • Keep evidence and dates the agent attached. "Raised on the Jun 18 call" is

what makes a risk actionable.

  • Unknown is a finding. "No objections surfaced" often means discovery was

thin, not that the deal is clean.

  • Quotes only from get_transcript.
  • Never merge two calls' transcripts into one answer.
  • The wrong date guess causes a miss, not a wrong answer. That is deliberate:

widen the window rather than trust a guess.

  • On multiple_matches, never pick for the user. "The latest" means the

newest.

Local mode conventions

When the connector is not connected, Zime skills work from files the user provides and say so in the first line:

  • Transcripts: .txt, .vtt, .json, .md. The answer covers only that

file, with none of the call's extracted signals or CRM linkage.

  • CRM or calendar exports: .csv. Look the row up; do not infer engagement

from a file with no activity columns.

  • No file and no connection → say so rather than reconstructing.

Output conventions

  • Header block naming the record and its hard facts (title, date, attendees,

linked deal; or stage, amount, close, owner).

  • Footer naming the source: _Live CRM record via Zime._,

_Live meeting record via Zime._, _Grounded in this deal's calls and CRM record via Zime._, _Live CRM and call data via Zime._.

  • Analyses render as a visual artifact: card sections, badges for short

labelled values, colour only on status fields (red at-risk, amber needs-attention, green on-track) and always with the word as well as the colour. Prose stays in the chat. Missing values stay visible as "not stated".

  • A transcript body stays plain text. No re-wrapping, no styling, no

truncation.

Sibling skills in the catalogue

account-research, actions-commitments, call-prep, competitive-intelligence, create-sales-asset, create-sales-to-cs-handover, daily-briefing, follow-up, review-account, review-call. Route to these by slug; do not invent others.