crm_metrics
ChatGPTCounts, totals, rates and breakdowns over leads, calls and activities, aggregated INSIDE the CRM. ALWAYS reach for this before search_leads when the question is a count, a total, a rate, a share, a ranking, a breakdown or a comparison -- "how many", "what percentage", "which source", "top 5", "summarise", "versus last month". One call comes back with the answer in a handful of rows. The same question answered by paging search_leads costs one call for every 200 records and returns records nobody reads; over a few thousand leads that is dozens of calls for a number this tool already computed. Use search_leads only when the user needs the individual records themselves. Break the answer down by one or two dimensions, filter it, compare it with the previous period, or ask for each row's share. ⚠ Three leads metrics are POINT-IN-TIME and ignore the date window entirely: count_not_contacted (never called at all), count_no_response (called, never connected, nothing tried for over two hours) and count_missed_unresolved (the customer called, it was missed, and nobody has connected since). They describe the leads as they stand right now, so from, to and all_time make no difference to them -- a lead untouched for a year is the most important row in a Not Contacted report, and reporting one of these as "for last week" is wrong even though the number comes back. Returns at most 200 rows. Call describe_schema first if you have not -- stage, source and disposition values are specific to this organisation.
describe_schema
ChatGPTCall this first. It returns the exact values this organisation uses for stage, source and disposition -- do not guess them. It returns this organisation's own lead fields as a table of label, key, type, whether the field is writable and whether its screen marks it required -- that key is what create_lead and update_lead take under fields, and a field this CRM has not made writable is refused by name rather than ignored. It also returns this organisation's lead custom fields by NAME, with each one's type, whether it is required and a dropdown's option labels, which is what create_lead and update_lead accept under custom_fields. It also returns every entity, metric, dimension and filter key the other tools accept, and this connection's own limits: rows per call, how far back it may read, which scopes are on, and -- when this connection may write -- how many records it may write per day and how many of those are left today. Read that budget BEFORE starting any job that creates or updates more than a handful of records: past it every further write is refused until the organisation's midnight.
fetch
ChatGPTRetrieves one CRM lead by the id search returned, as a document with its notes, activities and call summary. Same data and same limits as get_lead.
get_call
ChatGPTOne call in full, AND the whole AI assessment of it -- this is the only tool that returns either. The record: who called whom, the disposition, the duration, the outcome, the telecaller's own notes. The assessment: the six skill scores behind score_overall, the customer's sentiment and how it moved through the call, the questions each side asked, the root cause, what else the customer raised, the coaching notes, and every one of this organisation's own assessment rules that fired -- grouped into Team Execution, Business and Tracking, each with the line it fired on and why. Get the call_id from list_call_records, which lists calls by lead, telecaller, disposition or date. The transcript and the assessment are included only when this connection also has the call_analysis scope; otherwise they are named as omitted, which is not the same as the call having none. ⚠ A rule fires only if this organisation wrote it, so a call with no findings is NOT a call with no problems -- the reply says how many rules exist, and zero findings against zero rules means nothing was looked for. The call id is the conversation id -- the same record, one id. The reply says which lead the call belongs to, or that it is not linked to a lead. Calls made by the AI Phone Agent are not included in this tool. A transcript comes in parts of 20,000 characters: this returns part 1, and when transcript.next_offset is set, get_call_transcript returns the rest.
get_lead
ChatGPTOne lead in full: 34 fields including this organisation's custom fields and the lead's tags, its most recent notes and activities, and a call summary. Notes and activities are capped at 20 each and say so when there are more; list_notes, list_activities and list_call_records read the whole history of each, for this lead_id, when 20 is not enough. Get the id from search_leads. A lead outside this connection's scope reads exactly like one that does not exist.
kpi_tiles
ChatGPTA row of headline KPI tiles, each with its previous period, its change, its unit and WHICH DIRECTION IS GOOD -- so a fall in lost leads is not coloured as a problem. Every tile is one aggregate computed inside the CRM, so the whole row is ONE call. Reach for it before search_leads for any "how are we doing" question; never count records yourself to fill a tile. Call it with set: "morning_review" for the curated group (new leads, calls, connect rate, won, won value, lost, overdue tasks, never-contacted), or pass your own list under metrics. Passing both gives the set plus the extras. Defaults to the last 7 full days against the 7 before them; pass from = to = yesterday for a single day. A tile the CRM did not report comes back with value null and an unavailable sentence -- that is NOT a zero and must never be reported as one. Point-in-time measures (not_contacted, no_response, missed_unresolved, open_leads, untouched_leads) describe the leads as they stand now, ignore the window, and carry no change. For a measure that is not a tile, or for any breakdown, use crm_metrics.
list_activities
ChatGPTLists tasks and reminders: what is overdue, what falls in a due-date range, what is open, whose it is. This is the ONLY tool that lists them -- crm_metrics counts them without naming any. Filter by overdue, status label, type label, priority label, a due-date range, one lead, or the owner -- by user id, exact name or email. For one lead it returns the whole task history, not the 20 get_lead embeds. Every row says whether it belongs to a lead or is standalone, so that is never something to infer. Rows are bounded by what this connection can see, ordered by due date (no due date last). Returns at most 200 rows and 50 by default; when next_cursor comes back, send it as cursor for the next page. Call describe_schema first if you have not -- stage, source and disposition values are specific to this organisation.
list_call_records
ChatGPTThe individual calls, one row each -- and the ONLY tool that returns a call id, which is what get_call needs for a transcript or an AI analysis. Use it for one lead's complete call history, for a telecaller's calls, for calls with a given disposition or over a date range, and for the calls that reached no lead at all. Every row says whether it is linked to a lead or unlinked, so that is never something to infer. Use summarize_calls or crm_metrics instead when the question is a count, a rate or a breakdown -- they answer it in SQL rather than returning the rows to count. With the call_analysis scope it also finds calls BY WHAT HAPPENED ON THEM: the customer's sentiment, an AI score range, why the call ended as it did, the customer's interest, and which of this organisation's own assessment rules fired. That is how "find the calls where X happened" is asked -- then pass the id to get_call for the transcript and the full assessment. ⚠ Those five narrow to calls the analyzer has ALREADY processed, which is a subset: a call too short to analyse carries no assessment, and a rule this organisation never wrote can never fire. An empty answer from them is not evidence the thing never happened, so check the unfiltered count before reporting a zero. Returns at most 200 rows and 50 by default; when next_cursor comes back, send it as cursor for the next page. Call describe_schema first if you have not -- stage, source and disposition values are specific to this organisation.
list_notes
ChatGPTEvery note on one lead, oldest history included, with who wrote each one and when. get_lead returns only the most recent 20 and says so when there are more -- this is how the rest are read. A lead with no notes and a lead this connection cannot see both come back empty, so an empty answer is not evidence the lead is missing. Returns at most 200 rows and 50 by default; when next_cursor comes back, send it as cursor for the next page.
list_users
ChatGPTThe CRM users this connection can see, with their ids for user_ids and assigned_to. An admin connection sees the whole organisation; any other connection sees its own team only. Deactivated users are returned marked, so do not assign new work to anyone whose is_active is false.
search
ChatGPTSearches this CRM's leads and returns id, title and url for each match. Pass an id to fetch for the full record. Same data and same limits as search_leads. When next_cursor comes back, send it as cursor for the next page.
search_leads
ChatGPTFinds leads by name, company, email, phone or lead number, and filters them by stage, source, owner, city, tag, creation date, expected closing date, next follow-up date or last activity date (or no activity at all). Returns 12 fields per lead including the id you pass to get_lead; send fields to add up to 25 more (close dates and values, remarks, tags, custom fields, score, territory, last call) -- then the default page is 50. Last activity moves when a call on the lead is logged, edited or deleted, a note is added, a task is created on the lead or marked complete, or one of 21 core lead fields (stage, owner, source, contact details, deal and close fields, follow-up, custom fields) is edited. It does NOT move on lead creation (a new lead has none), tags, territory, score, imports, WhatsApp or email messages, AI Phone Agent calls, or linking an existing task to the lead later. Use crm_metrics instead when the question is a count, a total, a rate or a breakdown, and always past about 200 matching leads: it computes that in SQL and returns the answer, not the rows to count. Returns at most 200 rows and 50 by default; when next_cursor comes back, send it as cursor for the next page. A total shown as "1000+" is a floor, not a count: narrow the filters for an exact figure. Call describe_schema first if you have not -- stage, source and disposition values are specific to this organisation.
summarize_calls
ChatGPTCall volume and outcomes, broken down by disposition, status, sub status, call type, telecaller or day. Filter by disposition, call type, status, sub status, telecaller and date. It answers HOW MANY and names no individual call: for the calls themselves -- one lead's history, a telecaller's day, or the id that get_call needs -- use list_call_records. Returns at most 200 rows and 50 by default. Call describe_schema first if you have not -- stage, source and disposition values are specific to this organisation.
weekly_report
ChatGPTThe whole weekly review in ONE call: headline numbers with their week-on-week change, sources, team, pipeline and overdue follow-ups. ALWAYS prefer this to assembling the same report from a dozen crm_metrics calls, and NEVER build it by paging search_leads -- every section is aggregated inside the CRM, so this is one call where the other routes are dozens, and the comparison is composed server-side, so it cannot come back empty. Any request for a review, a summary, a digest or "how did we do" over a week belongs here. Defaults to the last 7 full days against the 7 before them.