MCP Server

Kitbase speaks MCP (Model Context Protocol). Connect it to claude.ai, Claude Code, Cursor, or any MCP-compatible client and ask questions about your traffic, events, users, sessions, AI visibility, and bot traffic in natural language. Over an OAuth connection it also exposes management tools — add AI-visibility competitors and prompts, start scans and site audits, build workflows, manage webhooks and team members.

The server is built into the Kitbase API — there is nothing to install or run.

https://api.kitbase.dev/mcp

Self-hosted deployments serve the same endpoint at https://<your-host>/api/mcp.

No Dashboard Required

Once connected, your AI assistant can answer questions like "How many visitors are on the site right now?" or "Which domains does ChatGPT cite instead of us?" by querying Kitbase directly — and, on an OAuth connection, act on requests like "Add Vercel as a competitor we track in AI answers."

Authentication

The server accepts two kinds of credentials. Both are scoped to exactly one project — tools default to it, so no organization parameter is ever needed (project-scoped tools do accept an optional projectId to target another project in the same organization).

Method	Best for	How
OAuth	claude.ai / Claude Desktop connectors	Add the server URL as a custom connector; a browser window asks you to sign in and pick the org + project to grant access to
Private API key	Cursor, headless clients	Send Authorization: Bearer sk_kitbase_... (or X-API-Key) with each request

Read tools work with either credential. Management tools work only over OAuth: each one checks the signed-in user's live permission (the same permission the dashboard enforces, capped by what the connection was granted). A private API key carries no user to check, so key-authenticated calls to management tools are rejected.

WARNING

Private API keys (sk_kitbase_) are secret keys — not the public SDK token used in browser SDKs. Generate one in your Kitbase dashboard under Project Settings > API Keys.

Revoking a connection

Revoke an OAuth connection from the dashboard under Account Settings → MCP Connections. Revocation takes effect on the connection's next token refresh, and the connection cannot renew itself afterwards.

Refresh tokens rotate on every use and are single-use. If a rotated token is ever presented again — the signature of a copied credential — the whole connection is revoked immediately rather than renewed, so a leaked token buys at most one refresh before both parties are cut off. Nothing is required of you for that to happen.

Client Setup

claude.ai (web / desktop)

1. Go to Settings > Connectors > Add custom connector
2. Enter https://api.kitbase.dev/mcp
3. Complete the sign-in flow and choose the organization and project to connect

Claude Code

claude mcp add --transport http kitbase https://api.kitbase.dev/mcp \
  --header "Authorization: Bearer sk_kitbase_your_key_here"

Cursor

Add to your Cursor MCP settings (.cursor/mcp.json):

{
  "mcpServers": {
    "kitbase": {
      "url": "https://api.kitbase.dev/mcp",
      "headers": {
        "Authorization": "Bearer sk_kitbase_your_key_here"
      }
    }
  }
}

Other MCP Clients

The server speaks the standard MCP Streamable HTTP transport. Any client that supports remote MCP servers works — point it at the URL and supply either credential.

Discovery

Clients and agents that look for MCP servers automatically, rather than being handed a URL, find Kitbase through the standard discovery documents. None of them require authentication.

Document	URL	Media type
Server card — name, version, description, transport, how to authenticate (SEP-2127)	https://api.kitbase.dev/mcp/server-card	application/mcp-server-card+json
AI catalog — the domain-level index that points at the server card	https://kitbase.dev/.well-known/ai-catalog.json and https://api.kitbase.dev/.well-known/ai-catalog.json	application/ai-catalog+json
Protected-resource metadata — which authorization server protects /mcp (RFC 9728)	https://api.kitbase.dev/.well-known/oauth-protected-resource/mcp	application/json
Authorization-server metadata — authorize/token/revocation endpoints, PKCE, scopes_supported (RFC 8414)	https://api.kitbase.dev/.well-known/oauth-authorization-server	application/json

An unauthenticated request to /mcp is answered with 401 and a WWW-Authenticate: Bearer resource_metadata="…/.well-known/oauth-protected-resource/mcp" header, which is how an OAuth-aware client finds its way to the consent flow without any configuration. The server card carries the same facts in one place:

curl -s -H "Accept: application/mcp-server-card+json" https://api.kitbase.dev/mcp/server-card | jq '.remotes[0]'

The card deliberately does not list tools — they depend on what the connection was granted, so clients read them at runtime with tools/list. The REST side of the same platform is described by the OpenAPI spec, which also documents POST /mcp with its OAuth security scheme.

Available Tools

The server exposes 180 tools. Most are reads for querying analytics, reports and configuration; the rest are management tools that change configuration — AI-visibility setup and runs, site audits, webhooks, projects, team, and workflows. Analytics data itself (events, sessions, visitors) can never be modified or deleted through MCP.

The tables below cover the tools you are most likely to reach for. Your client's own tool list is the complete one — ask your assistant to list the Kitbase tools if you want to see everything it has.

Unless noted otherwise, every tool accepts the standard date parameters described in Date Filtering: preset, from and to.

Web Analytics

Tool	Description	Extra parameters
get_web_summary	Summary KPIs — visitors, pageviews, bounce rate, average duration	filters
get_web_timeline	Visitors/pageviews as a time series	interval (hour/day/week/month), filters
get_web_breakdown	Break down by dimension (COUNTRY, BROWSER, OS, DEVICE, PATH, REFERRER, …)	dimension (required), page, size, filters

All three also take the period comparison parameters, so "how does this week compare to last?" is one call rather than two.

dimension is a closed set, and the tool schema lists it, so an assistant picks from it rather than guessing. There is no page dimension — pages are PATH, ENTRY_PAGE, EXIT_PAGE and TOP_PAGE. The full set:

DEVICE, BROWSER, BROWSER_VERSION, OS, OS_VERSION, BRAND, MODEL, COUNTRY, REGION, CITY, UTM_SOURCE, UTM_MEDIUM, UTM_CAMPAIGN, PATH, ENTRY_PAGE, EXIT_PAGE, TOP_PAGE, REFERRER, TOP_REFERRER, OUTBOUND_LINK, CUSTOM_EVENTS, EVENT_NAME, USER.

The REST API takes the same dimensions in lower case (top_page); over MCP either spelling is accepted.

Custom Events

Tool	Description	Extra parameters
list_events	List raw tracked events	event, channel, userId, filters, page, size, sort (desc default, asc)
get_event_stats	Grouped event counts	groupBy (event default, channel, user), channel
get_event_timeline	Event counts over time	interval, event
get_event_breakdown	Break events down by dimension (EVENT_NAME, USER, COUNTRY, …)	dimension (required), limit

Users

Tool	Description	Extra parameters
list_users	List/search aggregated user analytics	type (all/identified/anonymous), search, filters, page, size
get_user_summary	Comprehensive single-user summary	userId (required)
get_user_activity	Daily activity heatmap for one user	userId (required), months (default 4); no date params
get_user_events	Paginated event timeline for one user	userId (required), page, size; no date params

Sessions

Tool	Description	Extra parameters
list_sessions	List aggregated sessions	search, filters, page, size
get_session_detail	Single session details (metadata only)	sessionId (required); no date params
get_session_events	Events inside one session	sessionId (required), page, size; no date params

Funnels & Journeys

Tool	Description	Extra parameters
analyze_funnel	Ad-hoc funnel: conversion + drop-off per step	steps (required — ordered list of {type, value, name?, filters?}), analysisMode (session/user)
analyze_journey	Most common multi-step event sequences	steps (count, default 3), limit (default 100)

Frustration Signals

Tool	Description	Extra parameters
get_frustration_signals	Rage clicks and dead clicks, with top pages/elements	limit

Retention

Tool	Description	Extra parameters
get_retention	Cohort retention: visitors grouped by the week/month of their first visit, with the share still active in each following period	interval (week default, or month)

Cohorts are anchored to each visitor's first visit ever, not the first visit inside the requested window, so long-standing visitors are never counted as new. Offset 0 is the cohort itself (100%). The date range selects which cohorts are returned.

Realtime

Tool	Description	Extra parameters
get_realtime_visitors	Distinct visitors active right now	windowMinutes (1–60, default 5); no date params

AI Visibility

How often your brand appears in AI search answers (ChatGPT, Gemini, Perplexity, Claude), which domains those answers cite, and how you stack up against competitors.

Most of these tools accept a providers filter — a list of platforms such as ["CHATGPT", "GEMINI"], where omitting it means every platform — and a jobs parameter, how many of the most recent scan jobs to aggregate when no date window is given (default 10). Six of them also take the period comparison parameters, which need a date window rather than a jobs count.

Tool	Description	Extra parameters
list_ai_brands	Configured brands (yours + competitors)	none
list_ai_prompts	Prompts tracked against AI platforms	none
list_ai_topics	Prompt topics (thematic groupings of prompts)	includeArchived (default false); no date params
get_ai_visibility_series	Visibility score over time	providers, limit (default 30)
get_ai_visibility_domains	Domains AI answers cite, ranked by citation share	providers, jobs, limit (default 25)
get_ai_cited_pages	The specific pages (URLs, with titles) AI answers cite, ranked by citation count	providers, mentioningBrand, topicIds, jobs, limit (default 25)
get_ai_competitors	Your brand vs competitors (mentions, sentiment)	providers, jobs, limit
get_ai_share_of_voice	Your share of all brand mentions over time	providers, jobs
get_ai_prompts_breakdown	Per-prompt: did the brand appear in each provider's answer	jobs

Bot & AI-Crawler Traffic

Which bots and AI crawlers (GPTBot, ClaudeBot, PerplexityBot, Googlebot, …) hit your site, from where, and what they read.

Tool	Description	Extra parameters
get_bot_timeline	Bot requests over time, grouped by vendor	interval
get_bot_countries	Per-country bot traffic (requests + distinct bots)	page, size (default 8)
get_top_bots	Top bots ranked by request count	size (default 10)
get_bot_top_paths	Pages most requested by bots	size (default 10), vendor (e.g. GPTBot)

All four also take the period comparison parameters, which is how "is ChatGPT crawling us more than last month?" gets answered in one call.

Backlinks

Referring domains detected from the project's real traffic — search engines, social networks, AI assistants, and self-referrals are excluded. These tools use a days window instead of the standard date parameters.

Tool	Description	Extra parameters
list_backlinks	Referring domains with referred-visit counts	status (all default, active, ignored, lost, untapped), sort (first_seen default, sessions), q (domain search), days (default 30), page, size
get_backlink_detail	One referring domain: daily visit timeline, landing pages, and the linking-page URLs	domain (required), days (default 30)
get_backlink_reclamation	Dead pages that still receive referral traffic — backlinks recoverable with a redirect	days (default 90)
get_backlink_opportunities	Domains AI answers cite where your brand is absent — sites worth earning a link from; requires AI Visibility to be configured	jobs, limit (default 25)

Content Recommendations

Tool	Description	Extra parameters
get_content_recommendations	AI-generated create/update-page recommendations, each with the analytics, AI-visibility, and SERP facts it is grounded in	status (SUGGESTED/ACCEPTED/DONE/DISMISSED, default all), page, size; no date params

Workflow Reads

Workflows are the automations built on the dashboard's workflow canvas. These reads require workflow.view; the tools that change a workflow are under Workflows. None accept date parameters.

Tool	Description	Extra parameters
list_workflows	The project's workflows with their status (DRAFT, ACTIVE, PAUSED, ARCHIVED); graphs omitted	page, size (default 20)
get_workflow	One workflow with its full graph — every step, its config, and the connections between them	workflowId (required)
list_workflow_step_types	The catalog of step types: what each does, the payload types its inputs accept and its outputs emit, whether it costs AI credits, and every config key it takes	none
list_workflow_runs	A workflow's runs, most recent first, with status, trigger, cost and any error	workflowId (required), page, size
get_workflow_run	One run in detail: every step's status, what arrived on its inputs, what it emitted, and why it failed or was skipped	workflowId, runId (required)

Site Audit Reads

Everything a site audit produced. All of them require siteaudit.view, and none accept date parameters.

Each report tool takes an optional auditId and falls back to the project's most recent audit of its own website, so an assistant that has just been told about an audit does not need a second call to name it. An audit can be pointed at any address, and a competitor's report is never what an unqualified question means — so reading one means naming its id, which list_site_audits returns. There is no living page inventory behind these — every answer comes from one run's snapshot, so two figures read together were measured by the same crawl.

Tool	Description	Extra parameters
list_site_audits	Recent audits, most recent first	limit (default 10)
get_site_audit_worklist	The whole audit as one table to work through, worst first: every failed rule with why it matters, how to fix it, how many pages carry it, and the first few of those URLs. list_site_audit_issues plus the first page of list_site_audit_issue_pages for every rule, in one call — the place to start when the task is "fix what the audit found"	auditId, limit (default 20, max 100), examplesPerIssue (default 3, max 10)
get_latest_site_audit	The most recent audit of the project's own website: weighted AI-readiness score, the 16-item scorecard, headline crawl statistics	includeExternal (default false)
get_site_audit	One audit in full, by id: which site it ran against and whether that is your own, the score and its category scores, the scorecard's verdicts, which checks the run was narrowed to, headline statistics	auditId
get_site_audit_summary	One audit's headline statistics: pages crawled, issues by severity, average response time, and how many pages came back blocked, broken or erroring	auditId
list_site_audit_issues	What the audit found, one row per rule, worst first — why it matters, how to fix it, how many pages it affects	auditId
list_site_audit_checks	Every check the audit graded, worst first, including the ones that hold up — each with what this run measured, how to fix it, and the pages it names. The issue tools are issues-only, so this is the only way to tell "passed" from "never ran", and the only place that reports how many checks ran at all	auditId
list_site_audit_performance	Lighthouse measurements, one row per page and device: performance, accessibility, SEO and best-practices scores plus LCP, FCP, CLS, TBT, TTFB and Speed Index	auditId, device, band, failuresOnly, sort, q, page, size
get_site_audit_performance_page	One measurement with the audits it failed, worst first — what Lighthouse objected to (render-blocking resources, unsized images, long main-thread work), not just the score it ended on	resultId (required), auditId
list_site_audit_duplicates	Pages that say the same thing, clustered by meaning and tightest first. NEAR_DUPLICATE is two pages that are nearly the same page — the ones worth merging or canonicalising; RELATED is normal for a site about one subject	auditId
list_site_audit_issue_pages	The pages behind one issue. Pass both source and findingId exactly as list_site_audit_issues returned them — a rule id is only unique within its source	source and findingId (both required), auditId, page, size
list_site_audit_crawled_pages	The page-by-page snapshot of one audit's crawl: status, title, word count, images missing alt text, response time, crawl depth, issue count	auditId, status, sort, q, page, size
list_site_pages	The same page list for the project's latest completed audit of its own website — the shortcut for "what does our site consist of right now"	page, size, sort, filter, q
get_site_audit_page	Everything one audit knows about one page: what the crawl measured, which checks it failed and what each found, how a model read it, and the nearest pages by meaning. Address it by url or by the pageId from list_site_audit_crawled_pages	url or pageId (one required), auditId
get_page_content	The same page detail for the project's own site, addressed the way a person names it — by path	path (required), e.g. /pricing

sort takes URL, STATUS, SPEED, WORDS or DEPTH (default URL). The status filter — status on list_site_audit_crawled_pages, filter on list_site_pages — takes ALL, OK, REDIRECT, BROKEN, SERVER_ERROR, BLOCKED or MISSING (default ALL). q is a case-insensitive substring match on the URL.

On list_site_audit_performance, device takes ALL, MOBILE or DESKTOP, band takes ALL, GOOD, NEEDS_WORK or POOR, and sort takes PERFORMANCE, URL, LCP, CLS or TTFB (default PERFORMANCE, worst first).

Two empty answers mean "did not run" rather than "nothing found": no Lighthouse rows means no measurement happened — no provider configured, or the audit predates it — not that the pages are fast, and list_site_audit_duplicates reporting pagesCompared: 0 means the similarity pass never ran, not that no page repeats another.

A project with no finished audit has no pages

list_site_pages and get_page_content read stored rows only. If the project has never completed an audit of its own website there is nothing to read — start one with start_site_audit first. A path nobody crawled is reported as not found rather than as an empty page, because "we looked and there is nothing there" and "we never looked" are different answers.

Reviewing a Draft Before You Publish

Every tool above reads a crawl, so the earliest they can see an article is after it is live — and by then the expensive mistake is already quotable. audit_draft takes the file instead: your agent reads the draft off disk, passes the text, and gets back the checks the site audit runs on a published page, plus the one only a pre-publish review can make — whether the draft contradicts something your site already says.

Tool	Description	Extra parameters
audit_draft	Review an unpublished article: the prose checks the page report runs, plus the figures, structure words and titles it would contradict on the live site	content (required), format (markdown/html, detected when omitted), title, metaDescription, path, auditId

Requires siteaudit.view. Nothing is stored: the draft is read, graded and forgotten. The dashboard runs the same review on the same endpoint — Website Scan → Review a draft — for the times nobody is at a terminal.

"Audit blog/ai-visibility.md before I publish it"

Pass the text, not a path. The server cannot read your disk — the client opens the file. Markdown front matter is read for title and description, so a normal blog file needs nothing else; path is worth passing when you know where the piece will live, because it enables the check against the page already at that address.

The response answers "publish or not" first and gives the reasons under it:

Field	What it says
verdict	BLOCK only when the draft contradicts the published site. A thin description or a missing opening answer is a fix, not a contradiction, and comes back WARN; CLEAR means nothing this review can see would be contradicted
summary	One line: contradictions, what to fix, what is worth a look, and how many checks hold up
checks[]	Every graded line — passes included — with status (FAIL/WARNING/PASS/INFO), what was measured, and what to do about it
comparedSite	The audit the draft was read against and how many pages that comparison saw. null when the project has no completed audit of its own site
comparisonNote	Why the comparison is missing or partial, when it is
onlyAfterPublishing	The checks this review deliberately does not answer

What it grades depends on what you send. A markdown file is the prose you own, so it is graded on the opening answer, question headings, statistics, heading order, lists and tables, length, title, description and image alt text. Structured data, social tags and the byline belong to the template that renders it — asked only of an HTML draft, which is the whole page. Internal versus outbound links need to know which host is yours, so they are graded only when there is an audit supplying the domain.

A CLEAR draft is not a clean audit

onlyAfterPublishing exists because a file cannot answer everything: HTTP status, the canonical tag, sitemap membership, whether AI crawlers are allowed in, and whether anyone has cited the page are all properties of a live URL. Review the draft here, then audit the page once it ships.

Without a completed audit of your own website the contradiction half cannot run — the prose checks still do, and comparisonNote says exactly that. start_site_audit fixes it.

Configuration Reads

Unlike the analytics reads above, these expose org/project configuration, so each requires the same view permission its dashboard page enforces. None accept date parameters.

Tool	Description	Permission	Extra parameters
list_projects	All projects in the connected organization	project.read	none
list_webhooks	The organization's webhooks	webhook.view	page, size (default 50)

Management Tools

These tools change configuration — never analytics data. They work only over an OAuth connection (see Authentication): each checks the listed permission on the signed-in user and fails without it; private-API-key calls are rejected.

AI Visibility Setup & Runs

All of these require the aivisibility.manage permission.

Tool	Description	Parameters
add_ai_brand	Add a brand — your own (isSelf: true, only one allowed) or a competitor	name (required), primaryDomain, isSelf, aliases, excludedTopicIds
add_ai_competitor	Add a competitor brand — shorthand for add_ai_brand without isSelf	name (required), primaryDomain, aliases, excludedTopicIds
update_ai_brand	Edit a brand's name, domain, aliases or topic scope, or (de)activate it	brandId, name (required), primaryDomain, active, aliases, excludedTopicIds
delete_ai_brand	Delete a brand	brandId (required)

Scoping a competitor to certain topics. excludedTopicIds takes topic ids from list_ai_topics and mutes that competitor for them: it is treated as absent from every answer on those topics, so they leave its visibility rate and its share of voice, and it drops out of the leaderboard and charts whenever results are filtered to one. Resolve the topic name to an id with list_ai_topics first, and read a brand's current scope from list_ai_brands.

The set is replaced wholesale, so send the full list. On update_ai_brand, omitting the parameter leaves the current scope untouched — unlike aliases, which clears when omitted — so editing a name can't wipe a scope set in the dashboard. Send an empty array to deliberately track the competitor everywhere again. It is rejected for your own brand. | add_ai_prompt | Add a prompt tracked across answer platforms | text (required), locale (default en-US), topicId, personaId | | update_ai_prompt | Edit a prompt's text or locale, or (de)activate it. It cannot change the topic or persona — use the two tools below | promptId, text (required), locale, active | | move_ai_prompts_to_topic | File prompts under one topic, or take them out of every topic by omitting topicId. Pass one id to move one prompt; all or nothing across the list | promptIds (required, ≤500), topicId | | assign_ai_prompt_persona | Assign one persona to prompts, or unassign them by omitting personaId. Never touches their topic or text | promptIds (required, ≤500), personaId | | delete_ai_prompt | Deactivate (remove) a prompt | promptId (required) | | add_ai_topic | Create a prompt topic (thematic grouping) | name (required) | | update_ai_topic | Rename a topic | topicId, name (required) | | delete_ai_topic | Archive a topic — its prompts become uncategorized; historical runs keep the topic | topicId (required) | | suggest_ai_prompts | Generate prompt ideas for a brand — nothing is saved; pass the keepers to add_ai_prompt | brandName (required), primaryDomain, count (1–15, default 8) | | list_ai_exploration_engines | The platforms Prompt Explorer can run right now. Call it before explore_ai_prompt: this list is not the tracked-analysis platform list, and anything outside it is rejected | none | | explore_ai_prompt | Ask one question of up to 4 platforms at once and read every answer, the brands each named, and the sources it cited. Spends the organization's AI budget on every call — ask the user first, and don't loop over reworded variations: an identical question is free for 7 days, a reworded one is a fresh purchase, and a daily cap will start refusing calls. Nothing is tracked; use add_ai_prompt for a question worth monitoring | promptText, providers (required, 1–4), highlightBrand, region (US/EU), analyze (default true), runFresh | | pause_run, resume_run, cancel_run | Control an in-flight analysis run | jobId (required) |

Site, SEO & Content

Tool	Description	Permission	Parameters
start_site_audit	Start a link-following crawl and technical/SEO/AI-readiness audit of the connected project's website	siteaudit.manage	maxPages (defaults to the service setting), followLinks (default true)
update_backlink_status	Set a backlink source to active (shown) or ignored (dismissed as noise)	backlinks.manage	backlinkId, status (required)
update_recommendation_status	Move a content recommendation to ACCEPTED, DONE, or DISMISSED	contentrecs.manage	recommendationId, status (required), dismissedReason

start_site_audit always audits the connected project's own website, and runs every check the server can run. Auditing somebody else's site, or narrowing a run to a subset of the checks, is a dashboard decision — an assistant cannot spend your daily allowance for audits of other websites on its own.

Projects & Organization

Tool	Description	Permission	Parameters
create_project	Create a project in the connected organization	project.create	name, projectType (required), description, websiteDomain
update_project	Update the project's name, description, or website domain	project.update	name (required), description, websiteDomain (empty string clears it)
update_organization	Organization settings: name, logo, 2FA requirement, data-retention notification emails	organization.update	name, logoUrl, require2fa, dataRetentionNotificationsEnabled (all optional)

Webhooks

Webhooks are organization-scoped, not per-project.

Tool	Description	Permission	Parameters
create_webhook	Create a webhook that POSTs subscribed events to a URL	webhook.create	name, url, events (required), secret, enabled (default true)
update_webhook	Update name, URL, events, secret, or enabled state	webhook.update	webhookId (required); the rest optional — events replaces the whole set, empty secret clears it
delete_webhook	Delete a webhook	webhook.delete	webhookId (required)

Workflows

Everything the workflow canvas can do: author a graph, validate it, publish it, run it. Authoring requires workflow.manage and running requires workflow.run — the same permissions the dashboard enforces — and both are held to the plan's workflow entitlements (feature enabled, active-workflow cap, monthly run allowance, AI budget).

Start with list_workflow_step_types: it returns each step type's ports (which decide what can connect to what) and its config keys, so a graph can be built without guessing. Every authoring tool returns the workflow with its updated graph.

Tool	Description	Permission	Parameters
create_workflow	Create a workflow, empty or with its steps and connections in one call. It starts as a DRAFT	workflow.manage	name (required), description, steps ({type, key?, config?}), connections ({fromKey, toKey, fromHandle?, toHandle?})
update_workflow	Rename a workflow or change its description; omitted fields are left alone	workflow.manage	workflowId (required), name, description
duplicate_workflow	Copy a workflow's steps, connections and settings into a new DRAFT	workflow.manage	workflowId (required), name (defaults to "<name> (copy)")
delete_workflow	Delete a workflow and its run history; can report that it is pending schedule teardown — retry shortly	workflow.manage	workflowId (required)
add_workflow_step	Add a step, optionally wiring it to an existing one in the same call	workflow.manage	workflowId, type (required), config, key (defaults to the lowercased type), afterKey
update_workflow_step	Change a step's settings; the given keys are merged into its config, and a null clears one	workflow.manage	workflowId, stepKey, config (required)
remove_workflow_step	Remove a step and every connection that touched it	workflow.manage	workflowId, stepKey (required)
connect_workflow_steps	Feed one step's output into another's input	workflow.manage	workflowId, fromKey, toKey (required), fromHandle (default out), toHandle (default in)
disconnect_workflow_steps	Remove the connection(s) between two steps, leaving both in place	workflow.manage	workflowId, fromKey, toKey (required)
set_workflow_graph	Replace the whole graph — for rebuilding rather than editing. Anything not listed is dropped	workflow.manage	workflowId, steps (required), connections
validate_workflow	Check whether the draft could be published; returns each error pinned to a step or a connection	workflow.manage	workflowId (required)
publish_workflow	Compile the draft, make it the live version, and start its schedule. Fails with the validation errors if it is not runnable	workflow.manage	workflowId (required)
unpublish_workflow	Retire the live version, stop the schedule, and return the workflow to DRAFT	workflow.manage	workflowId (required)
run_workflow	Start a run and return its id; runs are asynchronous, so poll get_workflow_run	workflow.run	workflowId (required), mode (PUBLISHED default, or DRAFT_TEST), upToStepKey (runs only that step and everything feeding it)
cancel_workflow_run	Cancel a run that is still in flight	workflow.run	workflowId, runId (required)

A worked example — a weekly check of the site's busiest pages, emailed out:

list_workflow_step_types                              → what steps exist and what they accept
create_workflow  name: "Weekly page health"
add_workflow_step  type: SCHEDULE_TRIGGER   config: {"frequency": "WEEKLY"}
add_workflow_step  type: MY_SITE_PAGES      config: {"source": "TOP_TRAFFIC", "limit": 10}
                   afterKey: schedule_trigger
add_workflow_step  type: PAGE_HEALTH_CHECK  afterKey: my_site_pages
add_workflow_step  type: EMAIL_DELIVER      config: {"recipients": ["you@company.com"]}
                   afterKey: page_health_check
validate_workflow                                     → fix anything it reports
publish_workflow                                      → live, and running every week

Team

Tool	Description	Permission	Parameters
invite_member	Invite a user to the organization by email, with a role	member.invite	email, roleId (required)
cancel_invitation	Cancel a pending invitation	invitation.cancel	invitationId (required)
change_member_role	Change a member's role	member.update_role	memberId, roleId (required)
remove_member	Remove a member from the organization	member.remove	memberId (required)

Common Parameters

Date Filtering

Most tools accept date filtering via either a preset or an explicit date range:

• preset — one of last_30_minutes, last_hour, today, yesterday, last_7_days, last_30_days, this_month, this_year
• from / to — explicit date range in YYYY-MM-DD format (inclusive)

If both a preset and explicit dates are provided, the preset takes precedence.

Every window is resolved on the server in the project's reporting timezone — the same way the dashboard resolves it. That is why a relative period should be asked for with its preset rather than with dates the assistant computes: the dashboard's Last 7 days is last_7_days (today plus the six days before it), and an assistant that sends from = today − 7, to = today instead reads one day more and reports different numbers for the same funnel. Reserve from/to for explicit calendar dates. analyze_funnel echoes the days it actually covered as from/to in its response, so the period an assistant quotes is the one the numbers came from.

For the AI visibility tools, omitting all date parameters aggregates the last jobs scan jobs instead of a time window.

Period comparison

Thirteen of the read tools answer for an earlier window in the same call, so an assistant can say what changed without running the query twice and subtracting. They do it by default; these parameters change which window:

• comparePreset — one of previous_period, previous_week, previous_month, previous_quarter, previous_year
• compareFrom / compareTo — a custom comparison window in YYYY-MM-DD format (inclusive); send both or neither

comparePreset takes precedence when both are given, and omitting all three compares against the previous period — an assistant gets the change for free and only names a parameter to ask for a different window. previous_period keeps the length of your window and ends where it starts; the calendar presets are instead the week, month, quarter or year running up to it, which are usually longer — so an assistant comparing a day against previous_month is comparing it with the month behind it, not with the same day a month ago. Compare to an earlier period has the full semantics, the one-year guard, and how the extra fields read.

Area	Tools
Web analytics	get_web_summary, get_web_timeline, get_web_breakdown
Bots & crawlers	get_bot_timeline, get_bot_countries, get_top_bots, get_bot_top_paths
AI visibility	get_ai_visibility_series, get_ai_visibility_domains, get_ai_cited_pages, get_ai_competitors, get_ai_share_of_voice, get_ai_prompts_breakdown

AI visibility comparisons need a date window

The AI visibility tools fall back to aggregating the last jobs scans when no dates are given — a count, not a period. A comparison asked for in that mode is refused rather than silently dropped, so pass preset or from/to alongside it.

Pagination

List tools support pagination with page (0-indexed) and size.

Filters

Web analytics and event tools support filters in the format dimension:operator:values:

"country:is:US,UK"
"browser:is:Chrome"
"page:contains:/blog"

Pass multiple filters as an array to combine them.

Example Prompts

Question	Tools Used
"How many visitors are on the site right now?"	get_realtime_visitors
"How many visitors did we get last week?"	get_web_summary
"Compare this week's traffic to last week by country"	get_web_breakdown with comparePreset
"What's our weekly retention looking like this quarter?"	get_retention
"Show me the signup funnel conversion rate"	analyze_funnel
"How visible is our brand in AI search this month?"	get_ai_visibility_series
"Which domains does ChatGPT cite instead of us?"	get_ai_visibility_domains
"What are the most cited pages for our lead-capture topic?"	get_ai_cited_pages
"How do we compare to competitors in AI answers?"	get_ai_competitors
"Is ChatGPT crawling our docs? Which pages?"	get_top_bots + get_bot_top_paths
"Which pages have the most rage clicks this week?"	get_frustration_signals
"Show me a summary of user U-123's activity"	get_user_summary
"Add Vercel as a competitor we track in AI answers"	add_ai_competitor
"Stop counting Blinq on the badge-scanning topic"	list_ai_topics + list_ai_brands + update_ai_brand
"Suggest prompts for my brand and track the good ones"	suggest_ai_prompts + add_ai_prompt
"What does AI say about us right now?"	list_ai_exploration_engines + explore_ai_prompt
"Run a site audit and walk me through the findings"	start_site_audit + get_latest_site_audit + list_site_audit_issues
"Fix everything the last audit found, in my codebase"	get_site_audit_worklist + list_site_audit_issue_pages
"Which pages are slowest, and what is slowing them down?"	list_site_audit_performance + get_site_audit_performance_page
"What's wrong with our pricing page?"	get_page_content
"Check this draft before I publish it"	audit_draft
"Which sites should we try to get links from?"	get_backlink_opportunities

Connect external MCP servers to the assistant

The direction above is Kitbase acting as an MCP server. The project assistant can also act as an MCP client: connect any Streamable HTTP MCP server and its tools become available to the assistant in chat.

Where: Project Settings → Assistant → Connect server.

Field	Rules
Name	2–24 chars, lowercase letters/digits/hyphens. Immutable — it namespaces the server's tools as mcp__<name>__<tool>.
URL	The server's Streamable HTTP endpoint. https only.
Authentication	Auth header or OAuth, chosen at creation and immutable afterwards.
Auth header	Header connections only. Either a full Name: value line, or a bare token sent as Authorization: Bearer <token>. Stored encrypted; write-only — it can be replaced or removed, never read back.
Require approval	On by default: every tool call from this server pauses the chat for your approval, showing the exact arguments. Turn it off only for servers you trust.

OAuth-protected servers

Servers that authenticate with OAuth — Atlassian's https://mcp.atlassian.com/v1/mcp/authv2, for example — need no token pasted in. Create the connection with OAuth, then press Connect: Kitbase asks the server who authorizes it, registers itself as a client, and sends you to that server's own consent screen. Approve there and you land back on the project's settings page with the connection live.

• The grant belongs to the account you approved it with, and Kitbase keeps it alive by refreshing it in the background. Tokens are encrypted at rest and are never shown, exported, or returned by the API.
• Until you complete the flow, an OAuth connection holds no credential and contributes no tools.
• If the server revokes the grant — or you remove Kitbase's access on its side — its tools stop appearing. Press Connect again to re-authorize.
• Servers that do not offer dynamic client registration cannot be connected this way; use an auth header instead.

Behavior and limits:

• Up to 10 connections per project; the first 40 tools per server are offered.
• Tool lists refresh about once a minute; an unreachable server simply contributes no tools until it answers again. Use Test to dial a connection on demand and list its tools.
• External results are treated as untrusted data by the assistant — a page or tool output cannot instruct it to change its behavior.
• Connections are project-scoped, changes are audit-logged, and creating or editing them requires the project.update permission.

Requirements

• A Kitbase account with an active project
• An MCP-compatible client that supports remote servers over Streamable HTTP
• For API-key auth: a private API key (sk_kitbase_) from Project Settings > API Keys

Next steps

• Project Assistant — the same tools, built into the dashboard, with no client to connect.
• CLI — the same operations from your terminal instead of an AI assistant, and where kitbase login lives.
• API reference — the REST API behind every MCP tool.
• Analytics dashboard guide — the data these tools query.