Quickchat AI

Add an AI Agent alongside the ones the account already holds, optionally with initial settings. Not the tool for a first build: call list_scenarios first, and if it shows an Agent with `configured: false`, set that one up instead — creating another leaves it blank. Use this when the user wants an Agent in addition to the ones they have. Provisions a real AI Agent on the free tier (no charge) and returns its scenario_id, so you MUST pass confirm=true to avoid creating one by accident. Configure it with update_assistant_settings, add_knowledge_base_article, or onboard_assistant_from_url; knowledge-base changes are embedded automatically.

create_assistant

Save the configuration for an HTTP request AI Action: an API call the AI Agent can make during conversations (e.g. look up an order, create a CRM lead). This tool only stores the definition in Quickchat and never sends the request itself; test_http_request_action is the separate tool that executes a saved action. The target API is the one you own at the url you supply, and the method you supply is stored as configuration. Define parameters the AI fills in and reference them as {{parameter_name}} in the url, headers, query/body items or body_json. By default the AI chooses when to call it; execution_mode can instead run it automatically before the first reply or the first time its run_conditions match, in which case the AI fills nothing and every parameter needs a resolvable default_value. A valid GET (read-only) action activates automatically on creation; other methods stay inactive until set_ai_action_active. You can run test_http_request_action at any time, before or after activation, and deactivate an action that misbehaves. Subject to the plan's action limit.

create_http_request_action

Distribution of customer INTENT across conversations in the date range, into seven fixed buckets: purchase, support, inquiry, complaint, feedback, subscription, other. Use for 'what are people contacting us about'. These are intent categories, not free-text topics. For free-text topic labels use the `topics_by_day` field of get_analytics_overview. Defaults to the last 30 days; `meta` flags sampling on very large windows.

get_topics

Permanently delete an AI action and its configuration. This cannot be undone, so you MUST pass confirm=true; prefer set_ai_action_active with is_active=false to switch an action off temporarily.

delete_ai_action

Permanently deletes ONE knowledge-base article by its id, removing it from what the AI Agent knows. Irreversible — the delete proceeds only with confirm=true plus an expected_title and expected_added_at that exactly match the article's current values as returned by list_knowledge_base_articles, so a stale or hallucinated id cannot delete another article. The response echoes the deleted content and a restore_type usable to re-add it with add_knowledge_base_article, but only the first 20000 characters, so for a longer article (see content_chars in list_knowledge_base_articles) that response is not enough to restore it. Requires EDITOR-or-above role.

delete_knowledge_base_article

Export conversations to CSV/XLSX in a date range (max 31 days). Returns an EmbeddedResource pointing at a signed download URL that expires in 5 minutes. NOTE: this tool runs synchronously and can take 30+ seconds for high-volume chatbots — set the user's expectation accordingly before calling.

export_conversations

Conversations the AI flagged for human attention (is_flagged with a flag_reason: abuse, safety, compliance, legal, distress) and market insights (free-text business/product signals like competitor mentions or feature requests). insight_type selects flagged, market_insights, or all. Defaults to the most recent matches across ALL time — pass start_date/end_date to constrain to a window. Page with the opaque `next_cursor` (max 100 per call).

get_insights

Read an AI Agent's current persona/configuration and knowledge-base descriptions, plus `onboarding_from_url_completed` — whether an onboard_assistant_from_url run has finished (poll this after starting one), `onboarding_from_url_blocked` — present when that run refused the source, which is terminal, so stop polling and follow its `message`, and `onboarding_progress_tracked` — false when this Agent has no completion signal to poll at all, in which case read the settings once instead of polling. Requires EDITOR-or-above role on the scenario (matches the dashboard settings page). For embed snippets and live links use get_deployment_info.

get_assistant_settings

Recent call log for one AI action: per-call status, HTTP status code, duration, input parameters, output preview, error message and the conversation id it happened in. Only calls made during real conversations are logged; test_http_request_action runs do not appear here. Set only_issues=true to see just failed or rejected calls when debugging.

get_ai_action_calls

Read one AI action's full configuration. For HTTP request actions this includes method, url, headers, query/body items, parameters, response filters, save-to-memory captures and run conditions, with {{parameter_name}} placeholders shown by name. Works for all action types (remote MCP auth tokens are redacted).

get_ai_action

Read the step-by-step playbook for a Quickchat task before doing it. Each one names the right tools in the right order and the traps to avoid, so call this first whenever the user asks for something bigger than a single tool call — building or launching an Agent, improving or auditing one, changing what it knows or how it looks, testing it, triaging conversations, or reviewing performance. The `task` enum lists what is available. The result carries `required_arguments` for every tool the playbook names, read from the live schemas — trust that over the prose.

get_playbook

Read this AI Agent's plan, billing interval and AI credit position: plan allowance used and remaining for the current billing period, plus any purchased top-up credits, which are a separate pool spent after the plan allowance runs out. `replies_these_credits_can_fund` is the only field that says whether the agent can still reply: each reply costs `credit_cost_per_message` credits and is charged to one pool in full, so a balance below that rate funds nothing even though it is above 0. Use this instead of guessing or asking the user to read their billing screen. When `billed_externally` is true the account pays under a contract, the dashboard's Billing tab is disabled for it, and there is no `manage_url`. `subscription` carries live status and any already scheduled plan change: a plan card is greyed out in the dashboard only when it is the current plan or a scheduled change, so a greyed-out lower plan means that downgrade is already scheduled (read `subscription.next_tier`); downgrades are never blocked by usage. Use this for any plan, usage, credits, renewal, downgrade or cancellation question instead of guessing billing policy. Results are cached for 60 seconds. Read-only. Requires ADMIN-or-above role, matching who sees billing in the dashboard.

get_billing_info

Returns ONE knowledge-base article by its id, with its stored body up to 100000 characters. It is the only tool that returns more than a 500-character preview, so it is where the whole body comes from for the full-content replacement update_knowledge_base_article performs, and for a copy of an article kept before an irreversible delete. `content_truncated` marks articles longer than the cap: those cannot be rewritten through MCP at all, and no tool here returns a complete copy of one. Also returns content_chars, added_at and whether the article still auto-refreshes from its source url. Requires EDITOR-or-above role.

get_knowledge_base_article

Read Testing page (Simulation Testing) results. Poll this after run_simulation until status is completed, failed or cancelled. With run_id: the run's status, progress and per-message results — the AI's actual reply, the 1-5 AI score with the judge's justification, and a stable error code for failed messages. With dataset_id: its test runs newest first with status, progress (completed/failed of total), average AI score and any manual thumb-feedback score, paged by offset/limit, without loading test messages. Add view='dataset' to read only the dataset's messages, conversation history, evaluation rubric and conversation metadata. With no ids: the AI Agent's test datasets — named sets of visitor messages replayed against the AI and scored by an LLM judge — each with message_count, run_count and the latest run's AI score.

get_simulation_results

How fast HUMAN agents send their first reply after a conversation is handed off to them, in wall-clock SECONDS (not minutes, not business-hours-adjusted, so overnight gaps inflate it). Returns average and median seconds, a daily series, and a per-agent breakdown. Measures human responders, NOT the AI's speed, and only covers handed-off conversations; prefer the median. Defaults to the last 30 days.

get_ttfr

Everything needed to put an AI Agent in front of real users: the copy-paste website widget embed snippet, the hosted public chat page link (and whether it is enabled), the Agent's own MCP server endpoint for ChatGPT/Claude/Cursor (and whether it is active), the Discord servers it is live in, and the verified state of the other channels (Slack, WhatsApp, etc.) in `channels`. Treat not_configured there as a definite no; a channel in not_checked, or absent from all four lists, was not verified, so say you could not check it rather than guessing. Call this as the FINAL step after building or configuring an assistant and hand the snippet and links to the user. An assistant that is never deployed has no users. Requires EDITOR-or-above role on the scenario.

get_deployment_info

In-chat visitor feedback collected in the widget: thumbs/emoji answers to 'Was the AI helpful?' and 'How was your experience?', plus optional reasons. Returns raw counts per question and response, with no average score. Distinct from get_csat (external 1-5 CSAT). Defaults to the last 30 days.

get_ratings

List the AI Agents (also called assistants or chatbots; internally scenarios) this connection can access. Returns scenario_id, name, and the user's role per scenario, so an AI Agent named by the user can be resolved to its scenario_id. Pass name_contains to filter a long list instead of scanning it (the response's `total` is the unfiltered account count; an empty `scenarios` with name_contains set means no name matched, not an empty account). Each entry carries `configured`: false means nothing has been set up in that Agent yet — no name, persona, guidelines or knowledge — and its `name` is only its id standing in. An account whose Agents are all unconfigured is a new user who has not started: offer to set that existing Agent up now — pass its scenario_id to onboard_assistant_from_url if the user gives a website, or to update_assistant_settings if they describe the business in a few sentences. Do not provision another Agent; that leaves this one blank.

list_scenarios

List an AI Agent's AI Actions (the external operations it can perform mid-conversation: HTTP request actions, remote MCP servers, Shopify MCP, knowledge-base search). Returns each action's id, type, name, active/valid state and 7-day usage stats (call counts, success rate, average duration; all zeros until the action has been used in real conversations). Use get_ai_action for an action's full configuration.

list_ai_actions

Lists the knowledge base articles an AI Agent answers from, most-recently-updated first — an audit of what the Agent knows. Returns per article: id (the handle for update/delete_knowledge_base_article), title, type, source url (empty for manually added content), added_at (the immutable value delete_knowledge_base_article's expected_added_at takes), content_chars, whether the article auto-refreshes from its url, and a 500-character content preview — `content` here is a preview, never the stored body, which get_knowledge_base_article returns. Cursor-paginated via `next_cursor`, max 100 per call. Filter with `type` (e.g. 'URL' for website sources added via add_website_source). Content management happens through add_knowledge_base_article, update_knowledge_base_article and delete_knowledge_base_article. Requires EDITOR-or-above role.

list_knowledge_base_articles

List metadata keys available for AI Action placeholders and run conditions. Returns keys observed in this Agent's 20 most recent conversations plus keys its HTTP or Remote MCP actions are configured to save to memory. occurrence_count shows recent frequency, example_value is a bounded scalar example, has_scalar_value says whether the key is suitable for scalar comparisons, and is_memory_key identifies action-produced values. Use keys as {{metadata_<key>}} placeholders; pass the bare key to run_conditions.

list_observed_metadata_keys

Probe a remote MCP server URL and list the tools it exposes, without creating or changing anything. Connects to the server the way the AI Agent would (including the headers you supply), verifies it speaks MCP, and returns the discovered tool names — or requires_oauth=true when the server needs an OAuth sign-in, which can only be completed in the Quickchat dashboard. Use this before create_remote_mcp_action to preview a server's tools and decide which to enable.

list_remote_mcp_server_tools

Create or edit a test dataset on the Testing page to simulate multi-turn customer chats against the Agent, QA it before deployment, or run a pre-launch red-team suite. action='create' (name; optional description, evaluation_criteria, conversation_metadata, messages) makes the dataset, optionally with its messages. action='update' (dataset_id) changes only the fields you pass; existing messages are untouched. action='add_messages' (dataset_id, messages) appends visitor messages (at most 100 per dataset). action='delete' (dataset_id, confirm=true after the user confirmed) irreversibly removes the dataset, its messages and run history. evaluation_criteria sets how the LLM judge scores replies (1-5); conversation_metadata can e.g. satisfy AI Action run conditions. Creating a dataset runs nothing and spends no credits — follow with run_simulation; read datasets back with get_simulation_results.

manage_simulation_dataset

Refresh, auto-refresh or delete website pages already in the assistant's knowledge base. action=refresh: re-fetch the given `urls` from the live site in the background (omit `urls` to refresh every URL article) and overwrite the stored snapshots; retraining runs automatically afterwards. action=set_auto_refresh: pass `urls` and `enabled` to enrol (true, paid plan) or unenrol (false) pages in daily auto-refresh, which re-scrapes them when their content changes. action=delete: permanently delete the given `urls` from the knowledge base — all listed URLs must exist or nothing is deleted, and you MUST pass confirm=true; the response says whether the removal applies automatically or needs a retrain_assistant call. URLs must exactly match `url` values returned by list_knowledge_base_articles. Poll get_website_source_status for progress. Requires EDITOR-or-above role.

manage_website_sources

Per-message forensics for AI replies in one conversation: the tools the AI actually executed (arguments, status, HTTP code, duration, error) plus the Why AI Said That analysis (summary, grounding sources with influence scores, improvement hint) when it exists. analysis_status=not_generated means the background analysis has not run for that message; this tool never triggers generation. By default it covers the ten most recent AI messages that ran tools, plus the latest AI reply; pass message_ords with ord values from get_conversation_detail to inspect specific replies. Messages are always returned oldest first, so a set reads in conversation order. Use it to answer why the AI said or did something, or whether an action actually fired.

get_message_diagnostics

Start or cancel a test run — simulate customer conversations against the Agent to QA it before launch, red-team it, or check that a change helped: try the Agent and see its real replies without leaving this chat. action='start' (dataset_id; optional label and config_overrides; confirm=true after the user confirmed) sends every dataset message to the REAL AI Agent in an isolated test conversation; an LLM judge scores each reply 1-5 against the dataset's rubric. Spends one AI credit per message (2 or 4 on advanced models) and the AI may fire its active AI Actions for real. One run at a time per dataset. config_overrides A/B-tests settings changes without saving them. action='cancel' (run_id) stops an in-progress run: answered messages keep their results and their credits stay spent; pending messages are not sent. Poll get_simulation_results with run_id for progress and per-message scores.

run_simulation

Browse or find individual conversations in the Inbox, newest first, optionally filtered by channel, status, assignee, or a full-text query. Cursor-paginated via `next_cursor` (do not increment a numeric offset), max 100 per page. This is for locating or reading specific conversations, NOT for counts or statistics — for totals, volume, resolution rate, or any aggregate over a period use get_analytics_overview instead. To read the transcripts of many listed conversations, pass their ids to get_conversation_detail in batches of up to 20; to pull a whole date range at once use export_conversations. Cost guidance for high-volume chatbots: keep `query` to one or two distinctive words. Natural-language search broadens with each extra word. Dates keep returned conversations relevant but do not bound the full-text index scan. For a full-text search without dates, matching messages default to the last 30 days. After a timeout or `meta.fts_subquery_capped: true`, retry with fewer and rarer words, or omit `query` and use the other filters.

list_conversations

Send a message to an AI Agent as if you were a visitor and return its reply, so you can try the Agent you just built without leaving this chat. Consecutive calls continue the same test conversation, so you can hold a multi-turn exchange; pass new_conversation=true to start fresh. The message reaches the REAL Agent: it spends one AI credit (2 or 4 on advanced models) and any active AI Action fires for real, so confirm with the user before sending. Test turns are recorded as preview conversations, like the dashboard's AI Preview, and the Agent's analytics count them the same as any other conversation — totals and the source breakdown alike. To score many messages against a rubric instead, use the simulation tools.

send_message_to_agent

Record feedback about this MCP for the Quickchat product team — this is the main way tool gaps get found and fixed, so err on the side of sending it. Call it whenever: the user asks for something these tools cannot do (missing tool, filter, or field); you had to send the user to the dashboard UI to finish a task; a tool result did not match its description or was missing data you expected; a call was slow or failed in a way that disrupted the workflow (include any error text and correlation ID in `context`); or the user voices frustration with the assistant-management workflow. Summarise the gap in `summary`, put what the user was trying to achieve in `context`, and set `client_name` to the app you are running in (e.g. 'Claude Code', 'ChatGPT', 'Cursor'). It is lightweight and sends only the text you write, so no user confirmation is needed — but submit at most once per distinct gap per session, and mention to the user that the feedback was sent.

submit_mcp_feedback

Sets an AI Agent's profile picture — its avatar, profile photo, display picture, logo or icon — the image shown in the widget header and launcher, on the public chat page, and on other channels that display the Agent's face. Use this whenever the user asks to change, upload or replace how the Agent looks; there is no avatar field on get_assistant_settings. Accepts exactly one of: image_url (a publicly reachable direct image link, downloaded server-side), image_base64 (a base64 data URI, e.g. an image the user shared in chat), or use_default=true, which restores the default avatar. Replacing an avatar permanently deletes the previous one. Requires EDITOR-or-above role.

set_assistant_avatar

Execute a saved HTTP request AI Action once with the parameter values you supply and return the live response (status code, body after response filters, equivalent curl command). Read the verdict from `test_passed` and `outcome`, never from the presence of a response: a failed run still returns one. This sends a REAL request to the configured URL — for actions that create or modify data, use test values you can clean up. Run conditions are evaluated exactly as at runtime; pass metadata_values to satisfy them. Test runs are NOT recorded in the AI Agent's call log (get_ai_action_calls only shows calls from real conversations).

test_http_request_action

Update an AI Agent's persona and configuration: display name, personality (an enum id, see the field), profession, creativity_level, greeting(s), header(s), language, reply_length, and the knowledge-base one_word_description / short_description / ai_commands. Only the fields you pass change, and changes take effect immediately — no retrain needed. NOTE: applying settings syncs the Agent's servers synchronously and can take up to ~90 seconds — tell the user before calling, wait it out, and do not retry while a call is in flight. Requires EDITOR-or-above role on the scenario.

update_assistant_settings

Update the stored configuration of an HTTP request AI Action. This tool only edits the saved definition and never sends the request itself; test_http_request_action is the separate tool that executes a saved action. Only the fields you pass change; each list field (headers, query_items, body_items, parameters, response_filters, response_metadata_captures, run_conditions) REPLACES the existing list wholesale, so call get_ai_action first and re-send every item you want to keep. Changes to an active action are rejected if they would make it invalid.

update_http_request_action

Assign, resolve or reopen one Inbox conversation, or hand it back to the AI. Replaces assign_conversation and set_conversation_resolution in older playbooks; their assign_to_me=true is assign_to='me' here. assign_to: 'me' (the authenticated user), a teammate's email (must already have access to this chatbot), 'ai' or 'unassigned'. status: 'resolved' to close, 'open' to reopen. Only the current assignee can change resolution status, so pass assign_to='me' together with status='resolved' to take over and resolve in one call (the dashboard flow); no other assign_to value can be combined with status. Resolving ends the conversation and may trigger the visitor's final-rating flow, so ask the user and pass confirm=true; reopening and assigning need no confirmation, but ask the user before reassigning. The widget and dashboard Inbox update live, exactly as from the UI. Read the result back with get_conversation_detail. This changes one conversation — for counts of resolved conversations use get_analytics_overview.

update_conversation

Updates an existing knowledge-base article's title and/or content by its id (ids come from list_knowledge_base_articles or add_knowledge_base_article). The new content REPLACES the article's full content — there are no append semantics, so a partial change still means supplying the whole edited body, which get_knowledge_base_article returns; every other tool returns only a 500-character preview. Content equal to that preview is rejected rather than committed, as is any content for an article longer than 100000 characters, which no tool here returns in full — such an article is editable only in the dashboard, though its title can still be changed here. The response echoes `previous_title` and `previous_content`, which restore the previous title and body when `previous_content_truncated` is false (auto-refresh is not restored by them). The article is re-embedded automatically; an edit to a URL-sourced article also turns off its auto-refresh, so the edit is not overwritten by the next refresh. Requires EDITOR-or-above role.

update_knowledge_base_article

Check the progress of website scrapes for an assistant: active and failed scrapes with pages saved so far, plus the knowledge base retrain status. Pass `url` to check one source (reports completed/not_found once no scrape is active for it); omit it to list all activity. Use after add_website_source (with or without crawl=true) — poll every ~15 seconds until done.

get_website_source_status

Identify the dashboard user this MCP connection is authenticated as. Returns the user's email and the count of scenarios accessible. Use list_scenarios to enumerate the scenarios themselves. `configured_scenarios` counts the Agents that hold anything — a name, a persona, guidelines or knowledge; 0 means the account was created by the connector and nothing has been set up yet, which is the moment to offer to build that existing Agent from a website (onboard_assistant_from_url) or from a description of the business (update_assistant_settings), rather than to report the count or provision another Agent.

whoami