MimikFlow

The user's LinkedIn prospecting KPIs and funnel for a period: connection invites sent/accepted, first messages (cold DMs) sent, prospect replies, conversion goals reached (goal_reached = a won lead, usually a booked meeting), with per-step drop rates. Use it for recaps ('how is my prospecting going?') and to diagnose where the funnel leaks. A frequent leak: invites accepted but few replies, which usually means the first message's approach (formality, style, hook) does not fit the audience -- read real threads and the campaign profile, then tune the cold_dm_* settings. window is '7d', '30d' or 'lifetime'. Report the numbers to the user in plain words in their language (invitations envoyées / acceptées, premiers messages, réponses, objectifs atteints), never as raw field names.

Read one campaign's WhatsApp settings: the wording of the first message, the reminders and the AI that answers on this channel, its sending days/hours, plus the plan of messages around a booked meeting (see update_meeting_moments). What the campaign SAYS -- the offer, the target, the goal, the voice, the proof, the media library -- is shared with LinkedIn and is read with get_campaign_profile: only the channel's own wording lives here. An empty value means MimikFlow's default applies, which is a working configuration, never a gap. Name each setting to the user by what it does in their language, never by its key.

Raw material to coach the user on their sales conversations: full threads of recently WON leads (conversion goal reached) and LOST leads (dropped after the prospect had replied). YOU do the analysis: recurring objections, drop-off points, what worked in wins. Then translate EVERY finding into a concrete settings change (update_targeting for who gets contacted; update_campaign_profile for my_objections, the custom instructions, style or proof; update_pipeline_settings for the first-message tone when the opener does not fit the profiles contacted) and offer to apply it. Frame findings as settings to improve, never as defects of the product. outcome is 'won', 'lost' or 'both'.

The user's connected LinkedIn accounts. Most users have ONE; an AGENCE user runs several (e.g. one per client) from this single connector. Returns each account's id, name (label), LinkedIn plan, status (active/disconnected), how many campaigns run on it, and its lead-dedup pool (accounts sharing a non-empty pool never prospect the same people). Use this to know which accounts exist, then list_campaigns to see the campaigns of each, and target any campaign by its campaign_id -- one connector pilots every account.

The user's prospecting campaigns. A campaign bundles one ideal customer profile (ICP) + one offer + one tone + its own leads and conversations; most users have one, agencies run several in parallel. Returns id, name, status (active/archived), lead count, and the LinkedIn account it runs on (linkedin_account_id + account_name -- an Agence user targets a client by picking that client's campaign). Optional `account_id` filters to one account's campaigns. When a tool gets no campaign_id it uses the primary (oldest active) one.

Conversations MimikFlow's AI deliberately handed over to the human (status talk_to_human): the prospect asked something out of the AI's scope, requested a human, or a custom stop condition matched. The machine will NOT answer these; they wait for the user. Read each thread with get_conversation, then help the user craft an answer and send it via reply_to_lead. Covers every campaign by default (pass campaign_id to narrow to one client). total / has_more / next_offset describe escalation_cards; leads_waiting_total is the real number of prospects waiting for the user, of which the 30 most recent are listed.

List the user's prospects (leads found and worked by their automated LinkedIn prospecting). Filterable by status (comma-separated; 'connected' expands to every status where the LinkedIn relationship exists), tag name, or free-text search on name/position. Status meanings: new = found not yet contacted, connection_request = invite sent, connection_accepted = invite accepted, first_contact = first DM sent, discussion = the AI is talking with them, follow_up = being reminded, snoozed = re-engage at a future date, goal_reached = WON, talk_to_human = needs the user, dropped = abandoned. These codes are internal: report a status to the user in plain words in their language ('invitation envoyée', 'en discussion', 'objectif atteint'), never as the raw code. Paginated: `total` is the real number of matching prospects and `has_more` says whether you only saw a slice -- never present one page as the whole base, page with `next_offset` or narrow the filters.

LinkedIn conversations in the user's inbox that MimikFlow (the machine automating their LinkedIn prospecting 24/7) is NOT tracking: threads the user started or handled by hand, outside the automated pipeline. Scans the most recent chats and returns the untracked ones with the contact's name, headline and a chat_id. Use it when the user wants to see their manual conversations or hand one over to the machine; then call adopt_conversation with the chosen chat_id. Read-only: nothing is saved or contacted.

The campaign's media library: the videos, documents, images and links MimikFlow can attach to a message. Each one carries a title, a description (what it is and when it is relevant) and, when the user set one, an explicit rule saying when the AI may send it. Use it whenever a conversation would land better with something to SHOW -- a demo video before a call, a case study after one, the document a prospect asked for -- and to get the `media_id` that attaches a file to a message around a booked meeting (update_meeting_moments). Most users upload one file and never wire it anywhere, so an unused media is worth mentioning once when their conversations show it would help. Files are uploaded in the app; this tool only reads what exists.

Meetings the AI setter booked with prospects in the user's real calendar (the usual conversion goal). Returns prospect, datetime and status. upcoming_only=false includes past meetings. CAVEAT: only meetings stamped through a CONNECTED calendar appear here; a meeting agreed in conversation or self-booked via a link often does not. goal_reached in get_pipeline_stats is the source of truth for won leads -- never conclude 'no meetings booked' from this list alone.

The user's approval inbox: messages MimikFlow's AI drafted for prospects but is holding until the user validates them (validation mode, or a coherence check flagged the draft). A processing item is visible but must not be acted on until MimikFlow finishes or asks for a calendar check. Workflow: read each lead's thread with get_conversation, then approve_pending_message (send as is), edit_and_approve_pending_message (improve then send), or reject_pending_message (never send). Covers every campaign by default (pass campaign_id to narrow to one client), oldest first, and paginated: `total` is the real backlog, `has_more` says a slice is missing -- page with `next_offset` before telling the user their inbox is empty or counting it.

List the user's lead tags (free labels used to segment prospects, e.g. by cohort, event or offer) with the number of leads carrying each. Tags are usable as a filter in list_leads. Complete list, never truncated.

The WhatsApp numbers connected to the user's MimikFlow, and what each one is allowed to do right now. WhatsApp is a second channel on the SAME prospects as LinkedIn: it answers people who write first (a wa.me link, an ad, a QR code), carries the messages around a booked call, fills and converts a live, re-activates an imported base, and can take over when LinkedIn goes silent. Read `quiet_reason` before telling the user anything is broken: a freshly connected number stays SILENT for 24 hours and then writes only a few messages a day for its first two weeks, on purpose. That is WhatsApp's own rule against numbers that behave like machines, not a malfunction and not something to work around -- explain it and say how long is left. `paused` means the number was put on hold automatically because too few of the conversations it opened got an answer; the fix is better targeting and a better first message, never more volume. Cold-outreach and reply ceilings, credits left and whether unknown senders are accepted are reported as they are: none of them is writable from here.

Reject a pending draft (it will never be sent). Setter drafts move the lead to talk_to_human, cold DM drafts drop the lead, follow-up drafts restore the previous status. Post-booking cards preserve the lead's current status. Also resolves escalation cards.

Send one message to a prospect AS the user, in their existing conversation. Meant for conversations the AI handed over (talk_to_human) after reading the full thread with get_conversation, and for a quick courtesy message on a WON prospect right after a meeting ('merci pour ce call, je t'envoie un mail') -- a won prospect keeps its won status. Write in the conversation's language and the user's voice, and get the user's OK on the wording before sending: this message goes out immediately. SAFETY: the live LinkedIn thread is refreshed first, then the tool REFUSES to write when we already spoke last and the prospect has not answered since -- that is how a prospect the user answered by hand ends up reminded twice. The refusal returns the real last messages: read them, tell the user what is actually in the thread, and only pass allow_consecutive=true when they confirm they do want a second message in a row. channel: 'linkedin' (default) or 'whatsapp' when the prospect's number is connected and the conversation lives there. WhatsApp goes out under the number's own protections (daily ceiling, spacing, credits, opt-out): a refusal there is a safety rule doing its job, never a bug to work around.

After checking the real calendar for a draft marked calendar_outcome_unknown, record whether the meeting exists. This never sends a message. Use 'booked' only when the event is visible, otherwise use 'not_booked'. The draft then returns to normal review.

Give a conversation back to the AI. When MimikFlow (the machine automating the user's LinkedIn prospecting) hands a prospect over to the human -- the user answered by hand on LinkedIn, the AI escalated, or a reply landed while the AI was off -- that conversation stays manual forever: the AI stops answering and the reminders stop. This is the return path. It first syncs the real LinkedIn thread (so the messages the user typed by hand are recorded and stop being read as a new manual takeover), then puts the prospect back under the AI: it answers on its next run when the prospect spoke last, otherwise the reminder sequence resumes. The prospect's history and the notes the user wrote are kept, and the 'needs your attention' cards for that prospect are cleared. Only works on a prospect currently handled by the human. If the user answers by hand again afterwards, the conversation goes back to them, which is intended. Confirm with the user before calling: the AI will write to this prospect again.

Trigger an immediate one-off run of a MimikFlow pipeline instead of waiting for its schedule (e.g. lead_scraper to go find prospects right now, connection_sender to send today's invites). All daily LinkedIn safety limits still apply, so 'now' never means 'more'. Not available for setter (it runs whenever a prospect replies) and tracking. To the user, describe the action in product words in their language ('je relance la recherche de prospects'), never with the internal pipeline id.

Full-text search across every prospect conversation (e.g. 'pricing', 'not interested', a competitor name). Returns matching leads with the matched snippets. Scans the 200 most recent matching messages only: `truncated` true means older matches exist beyond what is shown, so present the results as examples, never as an exhaustive count.

MimikFlow's official help center content. Use it to answer ANY how-to or support question about MimikFlow (setup, features, plans, LinkedIn safety, billing, calendar, broadcasts, MCP...) with accurate product answers instead of guessing. Without arguments: the catalog of topics with their questions. With topic_id: that topic's full content. With query: every topic matching the text. Base your answer on this content and answer in the user's language.

Preview-search LinkedIn prospects matching a free-text query (e.g. 'expert comptable Bordeaux' or 'SaaS founder Paris hiring SDR'). Returns candidate profiles WITHOUT contacting or saving anything, marking those already in the campaign. Review with the user, then pass the chosen URLs to add_prospects_to_campaign.

Give the machine fresh news from the user (a new case study, a new result, a new offer detail). For the next 30 days, MimikFlow swaps ONE reminder per lead with a re-offer message built from this news, a strong way to revive silent prospects. Max 600 chars, strictly factual (the AI must never invent numbers). Empty text clears it.

Schedule a future re-engagement for a lead. Two uses: (1) the prospect asked to be re-contacted later ('come back next month') -- the lead is snoozed and MimikFlow wakes it at the date with one warm re-engagement message, then resumes its normal cadence; (2) after a meeting happened (lead already won / goal_reached) -- the lead KEEPS its won status and gets ONE follow-up at the date, with no automatic cadence after it. Pass the optional `note` to give the AI real context for that message (summary of the meeting or call, the offer made, the agreed next step): the follow-up will reference it naturally instead of shipping a generic re-engagement. callback_at is an ISO date clamped to 1-90 days from now; null cancels the schedule and clears the note.

Attach a tag (free label used to segment prospects) to a list of leads, creating the tag if it does not exist yet. Idempotent: already-tagged leads are unaffected. Useful to mark a cohort before filtering or reporting on it later.

Turn one of MimikFlow's automated pipelines on or off for a campaign. The prospecting chain: lead_scraper ('la recherche de prospects', finds new prospects that match the ICP), connection_sender ('les demandes de connexion', sends LinkedIn invites), cold_dm_sender ('le premier message', first personalized message), follow_up ('les relances', spaced reminders), setter ('l'IA qui répond aux prospects'). Extra lead sources: engagement_scraper ('les réactions à tes posts', people who like/comment the user's posts), profile_views ('les visiteurs de profil'), invite_manager ('les invitations reçues'), first_degree_scanner ('tes relations de 1er degré', existing connections). These ids are tool-call vocabulary only: to the user, always name the step by its product name in their language, never by its id. Changes apply within minutes; use run_pipeline_now for an immediate run. Warn the user before turning OFF part of the chain: leads then pile up unprocessed at that step.

Update the campaign's settings. Writable keys: who_am_i, my_product, my_objections (answers to common objections), sender_name, communication_style, style_examples, social_proof_stories, goal_type ('call' or 'link'), goal_link, goal_label, and the OPTIONAL behavior overrides custom_cold_dm_instructions / custom_follow_up_instructions / custom_setter_instructions plus setter_custom_stop_condition (extra hand-off situations). These overrides personalize MimikFlow's built-in behavior (which already varies reminders, respects a no, stops on silence); leaving them empty is a valid, normal state, so only propose filling one to address a SPECIFIC pattern seen in the user's own conversations. Read get_campaign_profile first, change only what the user validated, and APPEND to existing instructions rather than overwriting them unless the user asks. Present each change in plain words ('tes réponses aux objections'), never by raw key name.

Update the machine's fine-grained settings (same keys as get_pipeline_settings): reminder cadence and count, the selected offer-pitch follow-up, per-phase relance styles, first-message and setter tone, sending days/hours, sending modes, market/language. `target_regions` narrows prospecting to provinces / states / cities INSIDE the chosen countries (free text the user writes their own way, comma-separated, e.g. 'Quebec, Montreal' on a Canada campaign); empty means the whole countries. The first message's BASE STYLE is one choice made of two mutually exclusive keys: cold_dm_template picks a ready-made style, cold_dm_custom_template is the user's own example message the AI imitates. Setting either one automatically clears the other (they are reported in `cleared`), because an example message left behind silently overrides the chosen style. Daily and weekly VOLUME limits are readable but NOT writable here: how much a LinkedIn account may send is a human decision taken in the app, never reshaped from a chat. Each key is validated. CAUTION on modes: 'semi_manual' means NOTHING is sent until the user approves each draft in the inbox -- a forgotten semi_manual silently freezes the whole machine, so only set it on an explicit request and remind the user to process the inbox. Confirm every change with the user before saving, describing each setting by what it does in plain words, never by its raw key.

Rewrite the campaign's ideal customer profile: my_target (who the machine should find and contact) and/or my_anti_target (who it must avoid). Write in the user's language, concrete and specific (roles, sectors, company size, geography, pains). MimikFlow's prospect finder and evaluators apply the new text on their very next run, so confirm the wording with the user before saving.