MCP tools

Connect an AI client to https://app.pepline.ai/mcp (see Connect an AI assistant). Every tool is a resource with an action field: the bare noun reads, manage_<noun> writes. Reads are marked read-only; a manage tool is marked destructive when any of its actions archives. Operator-only tools answer 404 to organization credentials. Three prompts carry pepline’s own playbooks — publish_agent, knowledge_upkeep, tune_agent — and every tool runs the same /api/v1 call an integration would.

Organization

The organization itself: profile, credits usage, where it stands with onboarding. organization

action	params
`get`		Who-am-I: name, slug, agent slugs.
`usage`	`period`	Credits consumed vs the allowance (conversations, pages, documents at the org’s rates).
`onboarding_state`		Stage (choosing_path \| open \| pre_flow \| handoff_pending \| completed \| skipped), agent + go-live blockers, extraction state, checklist, what awaits triage, credits, smoke-test eligibility.

manage_organization

action	params
`update`	`name`, `about`, `locale_default`	Edit the profile (the logo stays a dashboard upload).

Organizations (operator) (operator only)

OPERATOR credentials only (404 otherwise): every organization on the platform, provisioning, credits, hand-off. organizations

action	params
`list`	`limit`, `starting_after`	Slug, name, dormant — the slugs every org-scoped tool’s `organization` takes.

manage_organizations

action	params
`provision`	`name`, `owner_email`, `owner_name`, `agent_name`, `starter`	Create an organization + owner (+ first agent when agent_name). A brand-new owner is EMAILED an invitation (invitation_url in the response as fallback); an existing account is seated directly. Trial allowance granted at once, NO 30-day clock (starts when the owner accepts, or via update_credits trial started).
`update_credits`	`slug*`, `monthly_credit_allowance`, `platform_daily_credit_budget`, `platform_weekly_credit_budget`, `trial`	Grant or adjust credits: allowance override, burst caps, trial state. Omit a field to leave it untouched; returns the trial + credits state.
`hand_off`	`slug*`	Declare a demo organization READY for its owner: the claimed owner lands on the handoff page (meet the agent, review what it knows, go live). Refused (422) until the org has an agent. Idempotent. Curate first.
`withdraw_handoff`	`slug*`	Take a hand-off back (the owner lands on conversations again).

Agents

The organization’s agents: who exists, their funnel, their widget settings; provisioning and go-live. agents

action	params
`list`		Slug, name, status, publishable key.
`funnel`	`agent*`	Event counts per stage (impressions → opened → started → completed → deliverable, then opened / qualified meeting / closed) and the split per attribution source and campaign.
`widget_setting`	`agent*`	Widget settings + hosted-page dressing.

manage_agents

action	params
`provision`	`name*`, `starter`	Provision an agent from a catalog starter. Starts PAUSED with empty activations.
`update`	`agent*`, `name`, `status`	Rename and/or set status.
`update_widget_setting`	`agent*`, `consent_text`, `offline_message`, `offline_cta`, `locale_default`, `allowed_domains`, `theme`, `reply_pace`, `hosted_page`, `translations`, `chrome`	Runtime knobs — serve immediately. translations merge per-locale field-by-field; hosted_page carries headline/intro copy.

Knowledge sources

What the agent reads: sources (text, url, file, website) and the pages beneath a website source. knowledge_sources

action	params
`list`	`status`, `kind`	Status, entries, freshness.
`get`	`id*`	Status, digest, entry counts, body for single-entry kinds.
`entries`	`id*`, `limit`, `starting_after`	The page-grained truth beneath a source (one per crawled page): title, url, status, failure_reason, ingested_at.
`updates`	`id*`, `limit`, `starting_after`	Refresh runs (trigger, status, stats) — the ledger behind status and freshness.
`pages`	`id*`, `q`, `limit`	Discovered-but-unread pages of a website source (what the last refresh mapped but did not read) — the allowlist add_pages promotes from.

manage_knowledge_sources

action	params
`create`	`kind*`, `title`, `body`, `url`, `max_pages`, `external_id`, `language`, `agents`, `reading_brief`, `refresh_every`, `knowledge_gap_id`	kind text (body) \| url (one page) \| website (crawls async — never poll, say what is running) \| file (base64 body). agents activates it for those agents.
`update`	`id*`, `title`, `language`, `body`, `max_pages`, `metadata`, `agents`, `reading_brief`, `refresh_every`	title, language, reading_brief, refresh_every, metadata; agents REPLACES the activation set (pass the full list); body only on text sources (re-reads it); max_pages only on website sources. A brief change re-reads cached pages (0 credits).
`refresh`	`id*`	Re-fetch / re-crawl / re-process (one verb for every kind, async).
`archive`	`id*`	Leaves every agent’s world, keeps its history (never destroyed). Idempotent.
`restore`	`id*`	Un-archive.
`add_pages`	`id*`, `urls`, `all_matching`, `q`, `note`	Promote discovered pages into read entries: urls, or all_matching with q. New pages are pinned and read now; pages un-ingested earlier revive with their curation intact.
`update_entry`	`id`, `entry_id`, `reading_note`, `link_excluded`	Curate one page: reading_note (what to look for on THIS page), link_excluded.
`refresh_entry`	`id`, `entry_id`	Re-read ONE page now, fresh from the site (goes pending; an unchanged body keeps its chunks).
`archive_entry`	`id`, `entry_id`	Un-ingest a page: back to the discovered list, chunks drop, facts go dormant (not archived).
`restore_entry`	`id`, `entry_id`	Re-ingest a page — pinned again, facts reclaimed with their curation intact.

Knowledge facts

Atomic knowledge the agent reads every turn (source → facts), and the machine-change ledger beneath it. knowledge_facts

action	params
`list`	`source_id`, `topic`, `core`, `tier`, `origin`, `reviewed`, `view`, `q`, `limit`, `starting_after`	Newest first (to REVIEW pending facts, review(queue) is the front door). Filters: source_id, topic, core, tier, origin, reviewed, view, q.
`changes`	`source_id`, `fact_id`, `kind`, `reverted`, `limit`, `starting_after`	Silent machine changes re-reads made: text_updated, retired (the page stopped saying it), reclaimed (came back), merged. Newest first.

manage_knowledge_facts

action	params
`create`	`source_id`, `text`	Author an owner fact on a source — never rewritten or retired by later re-reads.
`update`	`id*`, `text`, `topic`, `core`	New text makes it owner-curated; topic pins the subject (null clears); core pins the label (made core → always on).
`review`	`id or ids`	Mark reviewed (the triage bless — clears the pending queue). id or ids.
`archive`	`id or ids`, `scope`	The agent stops seeing it AND later re-reads won’t re-mint an equivalent (scope). id or ids.
`restore`	`id*`	Un-archive.
`pin`	`id`, `tier`	Pin the tier (retrieval only; the core label is update’s core). Pinned tiers survive re-reads and the always-on budget.
`unpin`	`id*`	Back to the machine’s judgment.
`exclude`	`id`, `agent`	Switch a fact OFF for one agent (sparse exception beneath the source activation gate).
`include`	`id`, `agent`	Switch it back on for that agent.
`revert_change`	`change_id*`	Revert one machine change — owner curation, not plain undo: the restored fact flips to owner origin. Reclaims are never revertible; already reverted → 409.

Ingestion instructions

“Look for” asks (D98): what the reader was told to capture on a source (every page), a page, or around a fact. Steering only — a fact still needs a verbatim excerpt; reached pages re-read from cached text (0 credits). ingestion_instructions

action	params
`list`	`source_id`, `entry_id`, `fact_id`, `view`, `limit`, `starting_after`	Filters: source_id, entry_id, fact_id, view. state: looked \| re-reading \| null (reaches no live page).

manage_ingestion_instructions

action	params
`create`	`text`, `targets`	Ask the reader to look for something on one or more targets.
`update`	`id`, `text`	Rewrite the ask; every page it reaches re-reads.
`archive`	`id*`	It stops steering; its pages re-read without it; never destroyed.
`restore`	`id*`	Un-archive.

Knowledge gaps

What leads asked that the agent visibly could not answer, deduped and counted. knowledge_gaps

action	params
`list`	`agent`, `status`, `limit`, `starting_after`	Close one by adding knowledge (a source, pages, a fact): ingestion re-checks open gaps and resolves the answered ones (verified); check runs that on demand; resolve records the owner’s word (declared).

manage_knowledge_gaps

action	params
`check`	`id*`	Run the answer check now: what the agent can reach for the question today and whether that answers it.
`resolve`	`id*`	Mark addressed (the knowledge was added).
`reopen`	`id*`	Undo resolve.
`archive`	`id*`	Dismiss (“not worth answering” — it keeps counting but leaves the open list).
`restore`	`id*`	Un-dismiss.

Services

The service catalog — what the business sells, offer-shaped. services

action	params
`list`		The whole catalog with activations.

manage_services

action	params
`create`	`name`, `description`, `positioning`, `guidelines`, `pricing_unit`, `price_min`, `price_max`, `currency`, `typical_duration`, `deliverables`, `good_fit_signals`, `agents`	agents = which agents offer it.
`update`	`id*`, `name`, `description`, `positioning`, `guidelines`, `pricing_unit`, `price_min`, `price_max`, `currency`, `typical_duration`, `deliverables`, `good_fit_signals`, `agents`	By id (key is immutable); agents, when present, replaces the activation set.
`review`	`id or ids`	Mark a drafted service reviewed — curation, not activation. id or ids. For a whole pass use review(queue) + manage_review(apply).
`archive`	`id or ids`	Leaves every agent’s catalog, keeps its history. id or ids.
`restore`	`id*`	Un-archive.
`consolidate`	`umbrella`, `fragment_ids`, `reason*`	Fold FRAGMENTS into one offer-shaped service (an audit, its strategy and its monthly plan are one “SEO” offer). The umbrella inherits the fragments’ deliverables, signals and activations; fragments are archived (restorable). One receipt (service_merged).

Fit criteria

Qualification standards the agent judges leads by. fit_criteria

action	params
`list`	`kind`	Optionally by kind.

manage_fit_criteria

action	params
`create`	`kind`, `description`, `guidance`, `weight`, `agents`	Without agents, ALL current agents judge by it.
`update`	`id*`, `kind`, `description`, `guidance`, `weight`, `agents`	By id.
`archive`	`id*`	No agent judges by it any more; history kept.
`restore`	`id*`	Un-archive.

Voice profiles

How agents speak: named profiles, each SELECTED by agents (selection, not activation). voice_profiles

action	params
`list`		Profiles and which agents speak with each.

manage_voice_profiles

action	params
`create`	`name*`, `languages`, `formality`, `tone_rules`, `banned_phrases`, `preferred_phrases`, `sign_off_style`, `agents`	agents selects it for those agents.
`update`	`id*`, `name`, `languages`, `formality`, `tone_rules`, `banned_phrases`, `preferred_phrases`, `sign_off_style`, `agents`	By id; agents switches those agents to this voice.
`archive`	`id*`	Refused while an agent still speaks with it (switch that agent first) or a running experiment draws it.
`restore`	`id*`	Un-archive.

Identities

Sender personas agents act AS for outward operations (email): name = From display name. identities

action	params
`list`	`view`	Identities and which agents act as each.

manage_identities

action	params
`create`	`name*`, `from_local`, `reply_to`, `copy_to`, `signature`, `agents`	agents ASSIGNS wearers — an agent leaves an identity by being claimed on another.
`update`	`id*`, `name`, `from_local`, `reply_to`, `copy_to`, `signature`, `agents`	By id; agents assigns those agents to it.
`archive`	`id*`	Refused while an agent still wears it.
`restore`	`id*`	Un-archive.

Conversation flow

An agent’s conversation flow: phases (prose, turn budget, tools) and slots (the visitor-facing chips). Every write trial-compiles and reaches NEW conversations only. flow

action	params
`get`	`agent*`	idle_minutes + phases + each phase’s slots.

manage_flow

action	params
`update`	`agent*`, `idle_minutes`	Flow-level: idle_minutes (silence pause; conversations stay reopenable 7 days).
`author_state`	`agent`, `title`, `objective`, `exit_criteria`, `max_turns*`	Add a phase (key minted from title; lands just before the closing phase — move_state after).
`update_state`	`agent`, `id`, `title`, `objective`, `exit_criteria`, `max_turns`, `tools`, `tool_settings`	A phase’s prose, its tools (FULL replacement set; removing capture_slots is refused while the phase has required answers) and/or tool_settings (`{offer_navigation: {max_links}}`; the tool must be on).
`archive_state`	`agent`, `id`	Refuses the last live phase.
`restore_state`	`agent`, `id`	Un-archive a phase.
`move_state`	`agent`, `id`, `direction*`	Swap a phase with the one above (up) or below (down); the closing phase stays last.
`create_slot`	`agent`, `flow_state_id`, `key`, `data_type`, `title`, `description`, `required`, `enum_values`, `enum_labels`, `translations`	Add a slot (chip) to a phase: enum_values required for enum/band; enum_labels value→label; translations locale→`{title, enum_labels}`. 16-slot budget.
`update_slot`	`agent`, `id`, `key`, `data_type`, `title`, `description`, `required`, `enum_values`, `enum_labels`, `translations`	Edit a slot’s chips/copy — this is how you fix wrong chips.
`archive_slot`	`agent`, `id`	Archive a slot.
`restore_slot`	`agent`, `id`	Un-archive a slot.

Prompt templates

An agent’s prompt versions per kind — never edited in place; new conversations compile the active one. prompt_templates

action	params
`list`	`agent*`, `kind`, `view`, `bodies`, `limit`, `starting_after`	Newest first; bodies ‘false’ for a light listing.

manage_prompt_templates

action	params
`author`	`agent`, `kind`, `body*`	Version N+1 of a kind, activated (unknown `{{references}}` warn, never block).
`activate`	`agent`, `id`	Point a kind back at any unarchived prior version (rollback).
`archive`	`agent`, `id`	The active version refuses (activate another first); a version a running experiment draws refuses too.
`restore`	`agent`, `id`	Un-archive.

Deliverable templates

The wrap-up document’s variants (A/B): title mints the immutable version key; sections in document order. deliverable_templates

action	params
`list`	`agent*`, `view`	An agent’s variants.

manage_deliverable_templates

action	params
`create`	`agent`, `title`, `description`, `sections`	Kind follows the agent’s flow.
`update`	`agent`, `id`, `title`, `description`, `sections`	In place (the version key never changes).
`archive`	`agent`, `id`	Refused while a running experiment draws it, or when it’s the flow’s last live template.
`restore`	`agent`, `id`	Un-archive.

Catalog drafts

Machine-drafted proposals (services, voice, fit criteria, flow) from the organization’s knowledge, and the runs that produced them. catalog

action	params
`drafts`	`kind`, `status`, `catalog_generation_id`	Proposals (to REVIEW them, review(queue) is the front door). fit_criterion payloads carry kind and basis (stated = the site says it; inferred = deduced from the offer).
`generation`	`id*`	Poll a regeneration run: status (running \| completed \| attention \| failed) and per-kind result.

manage_catalog

action	params
`regenerate`	`kind*`, `agent`, `mode`	Re-draft one kind from the knowledge, forced past the no-nag gate, or re-read the facts (kind fact) from the pages already held. mode names what is cleared first (add \| replace_pending \| start_over). One flow proposal per agent. Async — poll generation, then accept.
`accept`	`id or ids`	Creates the real catalog row through the normal authoring path. id or ids.
`dismiss`	`id or ids`	Permanent per fingerprint — a later regeneration never re-proposes it. id or ids.
`edit`	`id`, `payload`	Rewrite a PENDING proposal’s fields before accepting it (merge; the no-nag identity never moves). Receipted. A flow edit is trial-compiled.

Conversations

The work surface: what leads said, what the agent captured and judged, the documents it handed them. conversations

action	params
`list`	`agent*`, `status`, `fit_verdict`, `limit`, `starting_after`	An agent’s conversations, newest first (title, summary, status, fit verdict).
`get`	`id*`	One conversation in full: transcript (grounding included), captured answers, lead assessment, delivered document link.
`deliverables`	`agent*`, `limit`, `starting_after`	An agent’s deliverables (briefs/recaps), newest first, with verdict, matched service, delivery, the lead’s grade.
`deliverable`	`id*`	One deliverable in full: sections in template order, email framing, opened count, the lead’s grade and comment, public URL.

manage_conversations

action	params
`record_lead_event`	`id`, `event`	A lead outcome only a human knows (one per conversation; a repeat is a no-op).

Review surface

START HERE to review an organization: everything awaiting a decision in one read, decisions in one call, receipts behind them. review

action	params
`queue`	`kinds`, `limit`	Everything awaiting a decision before the agent can honestly go live: unkept services and service proposals, unreviewed facts (core first), fit and flow proposals (fit carries basis stated\|inferred), the voice choice while it is open (each proposal as a voice_proposal row, the voice in use as a voice row), discovered pages not yet read. Each row names its actions. counts = the full queue; rows capped per kind (limit).
`receipts`	`kind`, `actor`, `limit`, `starting_after`	The record of every decision applied (kind, actor operator \| agent \| owner, reason, subject, before/after). Newest first.

manage_review

action	params
`apply`	`decisions*`	Up to 200 decisions in ONE call, each through the dashboard’s own verb, each receipted with your reason. decisions `[{kind, id, action, reason?, fields?}]`: service keep\|dismiss\|restore\|edit(fields: name, positioning, guidelines, pricing_unit, price_min, price_max, currency…); service_proposal \| fit_proposal \| flow_proposal \| voice_proposal accept\|dismiss\|reconsider\|edit(fields: the proposal’s payload keys, merged); fact keep\|dismiss(fields.scope source\|organization)\|restore\|edit(fields: text, topic); criterion dismiss\|restore\|edit(fields: kind, description, guidance, weight); voice keep(id = the voice in use; fields optional: name, formality, sign_off_style, tone_rules, languages — the open proposals close; to pick a proposal instead, voice_proposal accept); page read(id = source id, fields.urls = discovered urls). Decisions are independent: the response says per row what applied and what failed.
`triage`		Resolve what needs no judgment, mechanically: exact duplicate services archived (the reviewed or oldest one kept), proposed services already in the catalog dismissed. A receipt per resolution; returns counts. Run it before queue on a fresh crawl.
`decision`	`question`, `options`, `answer`, `reason*`	Record a judgment call resolved by inference — the question, the options weighed, the answer taken and why — as a receipt (kind decision): what the owner reads on the handoff page.

* = required. Every tool also takes organization (the slug) with an operator key; an organization key or connector is already bound.

Getting started

Authoring

Guides

Conventions

Organization

Organizations (operator) (operator only)

Agents

Knowledge sources

Knowledge facts

Ingestion instructions

Knowledge gaps

Services

Fit criteria

Voice profiles

Identities

Conversation flow

Prompt templates

Deliverable templates

Catalog drafts

Conversations

Review surface

​Organization

​Organizations (operator) (operator only)

​Agents

​Knowledge sources

​Knowledge facts

​Ingestion instructions

​Knowledge gaps

​Services

​Fit criteria

​Voice profiles

​Identities

​Conversation flow

​Prompt templates

​Deliverable templates

​Catalog drafts

​Conversations

​Review surface

Organization

Organizations (operator) (operator only)

Agents

Knowledge sources

Knowledge facts

Ingestion instructions

Knowledge gaps

Services

Fit criteria

Voice profiles

Identities

Conversation flow

Prompt templates

Deliverable templates

Catalog drafts

Conversations

Review surface