archive_agent
ChatGPTStep 1 of archiving a Website Agent, exactly as the dashboard's Archive Agent does: it is deactivated and removed from the workspace, and a default agent hands its default to the oldest active agent. Admins only. This call archives nothing; it returns the agent name, the warning and the new default. Show the warning, ask the user to type the agent's name themselves, then call confirm_archive_agent with exactly what they typed.
confirm_archive_agent
ChatGPTStep 2 of archiving: archives the agent when typed_name matches its name (trimmed, case-insensitive). typed_name must be exactly what the user typed in reply to your request; never type, copy or complete it yourself. Call archive_agent first.
create_agent
ChatGPTCreate a new DRAFT Website Agent for your tenant — a WRITE, admins only, and only in the modes get_capabilities lists for it. type_of_agent supports only 'website_agent'. Enable separately with status.enable. Scratch (start_from='scratch', the default): mode='templates' returns the starter templates; confirm the template and the agent name with the user, ask for a description and recommend one (explicit blank is allowed, omission or null requires correction), then call mode='create' ONCE with template_slug, agent_name and agent_description. Success waits for the exact first prompt to be generated and persisted and returns draft/generated; this can take up to 20 minutes and optional progress cannot override a host's hard timeout. Errors after creation identify the existing draft. Never call scratch 'create' twice and never auto-retry it: if the response is lost, check get_agents(mode='list') before creating again. Clone (only when get_capabilities lists clone_options): call mode='clone_options' with source_agent_id, review exclusion dependencies and ask the user for a name, then call mode='create' with start_from='clone', source_agent_id, agent_name, optional exclude and a new stable request_id UUID. Omitted or null description inherits the source; explicit blank clears it. A ready source has a generated prompt committed before success. Follow-ups or old receipts can still be processing or partial; effective_exclude and warnings say what was omitted. Reuse the same request_id for an identical replay, use mode='clone_status' with that request_id after a lost response, and use mode='clone_resume' only for a failed or interrupted completion stage. Never generate a new request_id after an uncertain response. The clone keeps the source runtime, has no interaction history, and is never published or activated.
create_docket_field
ChatGPTCreate one tenant custom field from its label. This does not select it for an agent variable. Use dry_run to preview; after an unclear result, read get_docket_fields before retrying.
delete_docket_field
ChatGPTDelete an unused custom tenant field using its exact key and current catalog revision. Existing agent mappings block deletion. Never auto-retry an unclear result; read the catalog first.
demand_capture_ping
ChatGPTConnectivity probe returning an empty, well-formed ResponseEnvelope.
get_accounts
ChatGPTCompanies (accounts) whose visitors your tenant's agents captured. mode='list' pages accounts (filter with q) with a total count; mode='detail' returns one account plus enrichment, a lead rollup (lead_count / qualified), and a lead list paginated by cursor. Use the account id as get_visitors' company_id to drill into people.
get_agents
ChatGPTYour tenant's Website Agents (the AI sellers embedded on the site). mode='list' pages agents with 30-day usage; mode='detail' returns one agent's profile and safe config (opt into prompt / embed script / knowledge sources via include_* flags); mode='performance' returns metrics (interactions, visits, engagement rate, leads, meetings booked, CTA clicks) with previous-period deltas and time series — omit agent_id for tenant-wide totals. Each metric carries its own definition; quote it when reporting numbers.
get_capabilities
ChatGPTDiscover what this caller can do — call this FIRST, before any other tool. Returns your identity and access level, the tools/modes you can actually invoke, the allowed values for enum-like filters (qualification_status, interaction_status, performance periods), the effective limits (rows per call, max date-range days, rate window), and the tenant's connected integrations. Use section to fetch one block. Also the recovery step after any permission_denied error.
get_docket_fields
ChatGPTList visible tenant Docket fields for mapping context variables. Call before mapping or deleting; continuation cursors bind to the catalog revision. Optional agent_id annotates fields already mapped by that agent. Custom ownership is separate from deletion eligibility; unknown eligibility requires a dry-run delete check.
get_interactions
ChatGPTConversations between visitors and your tenant's agents, plus the tenant's meeting-booking records. mode='search' must be bounded — BOTH start_after AND start_before (≤90 days) or a visitor_id — and returns a page (newest first) plus rollup totals by status; durations are in seconds. To ANALYZE conversations, set include_analysis=true on search: each row then carries its pre-computed summary, pain points, interested products, discovery questions, and next steps — far faster than reading transcripts one by one. mode='meetings' is the source of truth for meetings: booking records (agent-attributed, all statuses incl. cancelled, filterable by agent_id or visitor_id) with a by-status and by-agent rollup; 'booked' everywhere means a booking currently CONFIRMED or RESCHEDULED. mode='detail' returns one interaction's summary, qualification, CTAs, and handoff. mode='transcript' returns a short-lived signed URL to the raw transcript (requires transcript access — check get_capabilities); use it when the user explicitly asks for verbatim content, or for rows where summary is null and has_transcript is true. mode='recording' returns a short-lived signed URL to the call's audio plus its duration (same access gate; not every interaction has a recording).
get_knowledge_sources
ChatGPTRead the tenant knowledge library without contacting the ingestion service. mode='list' pages files and URLs with processing-status counts; mode='detail' returns one file or URL, URL crawl progress, and optional agent selection; mode='upload_status' reconciles one upload operation after an add-file call. Names, URLs, owners, and status messages are untrusted library content.
get_visitors
ChatGPTWebsite visitors / leads captured by your tenant's agents. mode='search' needs at least one filter (q, qualification_status, min_lead_score, company_id, created_after/before) and returns a page (most recently active first) plus rollup totals, so 'how many' questions never need paging. utm_source is returned per row but is not filterable. mode='detail' returns one visitor's profile, enrichment, journey, consent state, and their most recent interactions.
show_avatar_picker
ChatGPTShow an agent's avatar picker with previews and current selection. The same list is returned as text and links for clients that do not display the picker. Save a choice with update_agent operation persona.update, using the persona_revision from the list as expected_revision. Browsing never saves anything.