Skip to content
Kleap for AI agents

Ask your AI agent for a website. Kleap makes it live.

Tell your agent: “build a website for my yoga studio”. It builds the site and publishes it at a kleap.io link.

  1. Copy the server URLhttps://kleap.co/api/mcp
  2. ConnectAdd it to your agent, then sign in on kleap.co when asked.
  3. Paste the promptThe tested prompt is just below. Change the studio for your business.
Median time to a live link
7 min23 of 25 counted runs of the yoga job, all five agents we ran
Kleap's build itself
5 minmedian of the 25 builds inside the counted runs (the ones with the wait sentence). Range over all 36 recorded builds: 3 to 10 min (3 min 17 s to 10 min 22 s)

Works with Claude, Grok, OpenClaw, Hermes Agent and Codex, and with ChatGPT and Cursor (setup from docs). Kleap builds, hosts and publishes the site, then proves it is live.

Choose your agent Create a free account

Checked Oct 19–20, 2026 · every number on these pages comes from a real run

A real run, OpenAI's Codex CLI

Use Kleap to build a website for my yoga studio: class schedule, teachers, pricing and a booking request form. Calm colours. Wait for the build to finish, publish it and send me the link.

Published and checked live at still-yoga-studio-wznb.kleap.io

The filmed run took about 8 minutes from the prompt to the link, most of it waiting for Kleap's build. The video is on the desktop version of this page.

Pick your agent

One guide per agent

Exact commands and URLs, prompts that work, honest limits, and what we ran ourselves.

Run end to end on Oct 19–20, 2026

Setup from docs, not run end to end

ChatGPT cannot be driven from a script, so we ran Codex, OpenAI's own agent, on the same job. Cursor's agent needs a Cursor login we do not have. Kleap's side is verified for both; the steps come from the vendors' documentation.

Another MCP client, such as Windsurf or Cline? Add https://kleap.co/api/mcp as a remote server and sign in. The MCP, CLI and skills reference lists every tool.

In-depth guides

How often does it work?

23 of 25 counted runs reached a live link

We gave every agent the same sentence on Oct 19–20, 2026, one build at a time, and kept every run.

The prompt, word for word

Use Kleap to build a website for my yoga studio: class schedule, teachers, pricing and a booking request form. Calm colours. Wait for the build to finish, publish it and send me the link.

How we count. A run counts as a success when the agent's last message gives a kleap.io link that answers HTTP 200 afterwards and Kleap's build had finished when the agent answered. No link, or a link to a half-built site, counts as a failure. A run during which Kleap's own build failed is shown but not counted.

4 of 29 runs were set aside because Kleap's own build failed or waited in Kleap's queue (marked †); 1 more counted as a failure although a link existed, because the agent published before Kleap's build had finished. See every run

  • Claude Code 2.1.3004 of 5 counted · 7 min median
  • Codex CLI 0.163.05 of 5 counted · 6 min median
  • Grok CLI 1.0.624 of 5 counted · 8 min median
  • OpenClaw 2026.4.155 of 5 counted · 8 min median
  • Hermes Agent v0.22.25 of 5 counted · 9 min median
Proof

What did agents actually build?

We gave agents real jobs on October 18, 2026 and kept every call. These are the sites they made, live today, with the build times we measured.

Swipe sideways to see all six runs →

The details

Success rates, comparisons and the full tool list

Open what you need. Nothing is removed: it is all on this page.

How often does it work?

The same job, run again and again, per agent

Every run shown, none dropped. Counted runs per agent, and the ones we set aside.

On Oct 19–20, 2026 we gave every agent the same sentence, one at a time, and counted the runs that ended with a live link. Every run is shown, none was dropped, and the sample is small: 5 to 9 runs per agent.

  • Claude Code 2.1.300

    Claude (Anthropic), the default model of the account

    • Run 1: 7 min 09 s
    • Run 2: 8 min 59 s
    • Run 3: 7 min 11 s
    • Run 4: 7 min 00 s
    • Run 5: no link
    4 of 5counted runs reached a live link
    7 min 10 smedian time to a live link
    8Kleap calls, median

  • Codex CLI 0.163.0

    gpt-6-astra (OpenAI), the Codex default

    • Run 1: 5 min 34 s
    • Run 2: 5 min 47 s
    • Run 3: 5 min 50 s
    • Run 4: 5 min 58 s
    • Run 5: 8 min 16 s
    5 of 5counted runs reached a live link
    5 min 50 smedian time to a live link
    9Kleap calls, median

  • Grok CLI 1.0.62

    grok-4.7 (xAI)

    • Run 1: 9 min 02 s
    • Run 2: 14 min 05 s†
    • Run 3: no link†
    • Run 4: 6 min 22 s
    • Run 5: no link
    • Run 6: no link†
    • Run 7: 8 min 39 s
    • Run 8: 15 min 41 s†
    • Run 9: 6 min 53 s
    4 of 5counted runs reached a live link. 9 runs played, 4 not counted because Kleap's own build failed during them (shown with †)
    7 min 46 smedian time to a live link
    16Kleap calls, median

  • OpenClaw 2026.4.15 with the Kleap skill

    gpt-6.1-sol through OpenLux: OpenClaw needs some provider, the Kleap side is the same

    • Run 1: 8 min 29 s
    • Run 2: 9 min 36 s
    • Run 3: 6 min 01 s
    • Run 4: 6 min 24 s
    • Run 5: 9 min 07 s
    5 of 5counted runs reached a live link
    8 min 29 smedian time to a live link
    15Kleap calls, median

  • Hermes Agent v0.22.2

    gpt-6.1-sol through OpenLux: Hermes needs some provider, the Kleap side is the same

    • Run 1: 9 min 33 s
    • Run 2: 9 min 24 s
    • Run 3: 7 min 44 s
    • Run 4: 8 min 47 s
    • Run 5: 6 min 26 s
    5 of 5counted runs reached a live link
    8 min 47 smedian time to a live link
    7Kleap calls, median

† Kleap's side failed during these runs, so they are shown but not counted: twice an AI provider account on our side answered “Insufficient balance” (13:19 to 14:09 UTC on Oct 19) and Grok then tried to build the pages itself with file tools, and once a build waited about 20 minutes in Kleap's own queue before it started. They measure our platform, not the agent.

Without the sentence “Wait for the build to finish”: 0 of 3 (tiny sample) runs reached a link before Oct 19, 4 of 5 after

We also ran the shorter prompt “Use Kleap to build a website for my yoga studio: class schedule, teachers, pricing and a booking request form. Calm colours. Publish it and send me the link.”. Until Oct 19, 2026 Kleap's first answer to create_app told the agent there was no need to poll, which suits a chat window with a live preview and not a script: agents obeyed and ended their turn before the site existed. Kleap still finished and published every site; nobody was left to send the link. That answer now separates a person watching from a run with nobody there, and we re-ran the bare prompt on the new version.

  • Claude Code 2.1.300
    Before
    • Run 1: no link
    After
    • Run 1: 7 min 01 s
    • Run 2: no link
  • Codex CLI
    After
    • Run 1: 5 min 14 s
    • Run 2: 6 min 38 s
    • Run 3: 6 min 16 s
  • Hermes Agent v0.22.2
    Before
    • Run 1: no link
    • Run 2: no link
    After
    • Run 1: no link†

The copy-paste prompts on each guide still include the wait sentence: it costs nothing and works on every version.

ChatGPT and Cursor are not in this table: neither can be run from a script on our side, and each guide says why. ChatGPT shows the runs of OpenAI's own agent, Codex.

How long does it take?

Kleap's build itself: 3 to 10 minutes in all 36 recorded builds (3 min 17 s to 10 min 22 s)

Each build time is the build task measured in Kleap's task log.
  • 3 min 17 spodcast page, Grok CLI
  • 3 min 44 sbike repair site, OpenClaw
  • 4 min 21 sseven-page portfolio, Claude Code
  • 4 min 32 sone-page bakery, MCP client
  • 4 min 54 srepository landing page, Codex
  • 5 min 50 syoga studio site, Hermes Agent

That is the build task alone (kleap's build itself), not the median time to a live link: the agent also reads, checks and polls, so its time to a link is longer. Your agent does not wait idle: the first call returns in seconds with a task, and the agent checks back. The six above are from Oct 18; 30 builds from the success-rate runs of Oct 19–20, 2026, with or without the wait sentence, took 3 min 29 s to 10 min 22 s, 4 min 57 s at the median. The hero median (5 min 02 s) covers only the 25 builds inside the counted runs that used the wait sentence. Builds that failed or sat in Kleap's own queue are in neither count.

Side by side

Which agent works with Kleap, and how sure are we?

Green works or we ran it. Amber is a manual step. Grey is read from the docs.

Green means it works or we ran it. Amber is a manual step or a partial run. Grey is read from the vendor's docs. We keep the amber and the grey on purpose.

  • ChatGPT
    Way in
    Custom MCP server
    Sign-in
    OAuth
    Skills
    No shell, no skills
    Scheduling
    ChatGPT scheduled tasks; OpenAI decides the connectors
    Run by us
    Server side and a scripted run, not ChatGPT's UI
  • Claude
    Way in
    Connector, or claude mcp add
    Sign-in
    OAuth
    Skills
    Yes, plugin and skills CLI
    Scheduling
    Routines, /schedule, /loop (research preview)
    Run by us
    Yes, Claude Code
  • Grok
    Way in
    Grok CLI, or xAI API tool
    Sign-in
    OAuth (CLI), key (API)
    Skills
    Yes, Agent Skills
    Scheduling
    You schedule it: cron or CI
    Run by us
    CLI yes, API not run
  • OpenClaw
    Way in
    ClawHub skill, or MCP
    Sign-in
    OAuth, or key
    Skills
    Yes, on ClawHub
    Scheduling
    openclaw cron; Gateway must run
    Run by us
    Yes, the skill in OpenClaw 2026.4.15
  • Hermes Agent
    Way in
    config.yaml entry
    Sign-in
    OAuth, or key (run with a key)
    Skills
    Yes, from a raw SKILL.md URL
    Scheduling
    Built-in cron
    Run by us
    Yes, v0.22.2 with a key (OAuth not run)
  • Codex
    Way in
    codex mcp add
    Sign-in
    OAuth, or key
    Skills
    Yes, skills CLI
    Scheduling
    Scheduled tasks; MCP use unconfirmed
    Run by us
    Yes, codex exec
  • Cursor
    Way in
    mcp.json entry
    Sign-in
    OAuth, or key
    Skills
    Yes, skills CLI
    Scheduling
    Not verified: use the CLI in cron
    Run by us
    No, read from docs
  • Works, or we ran it
  • A manual step, or partly run
  • Read from the docs, or not verified
How it fits

One connection. Everything a site needs.

Build, publish, read leads, watch search, connect a domain.

An agent that only writes code hands you files to host. With Kleap connected it can run the whole life of a website: build, publish, read leads, watch search, connect a domain.

Kleap

One account, one set of sites. The agent decides, Kleap does the work.

Connector (MCP)
Command line
Skills
  • A live siteon a kleap.io address or your domain
  • Hosting and SSLincluded, nothing to set up
  • Forms and leadsstored, readable by your agent
  • Analyticsvisitors, pageviews, top pages
  • Search Consolequeries, impressions, position
  • Domainssearch and connect; you pay the link
Who it is for

Three jobs agents already do on Kleap

Freelancers and agencies, developers, marketers and founders.
SEO on a schedule

Can an agent fix your SEO issues every day?

The loop run for real by two agents on a test site, and the prompt that asks any agent to run it.

An agent that can run a prompt on a schedule can audit a site, fix what it finds, publish, and tell you what changed. Below, that loop run for real by two agents on a test site, and the prompt that asks any agent to run it. We make no traffic claim: the test site has no Search Console.

What two agents changed on a test site on October 18, 2026, before and after
AgentFixAfter
Claude CodeDead linkContact link pointed to /contact, which is not a pageNow points to the enquiry form on the page
Claude CodeMeta description“Nice small hotel.” (17 characters)“Small hotel with a Deluxe Room and a Rooftop Pool. Browse the gallery, then send a stay enquiry and our team will get back to you.”
Claude CodeStructured dataNo page-level schemaA Hotel node with only the name and URL, since the site shows no address, phone or rating
Grok CLITitle“Hotel Kleap Prod Check” (22 characters)“Hotel Kleap Prod Check — Rooms & Rooftop Retreat” (48 characters)
  • Claude Code. Claude reported one publish with live_verified true, a clean re-audit and 0 credits spent. The original file was backed up first.
  • Grok CLI. Grok reported one publish with live_verified true and a clean re-audit. It flagged a missing image the audit could not fix and left it alone.

Run with an SEO skill we have not published yet, through the Kleap CLI. The prompt on this page calls the equivalent MCP tools: run it once by hand before you schedule it. The test site had no Search Console, so we claim no traffic result.

Scheduled prompt

Every morning, go through my Kleap sites one by one. For each site: 1. Read the last 28 days with get_search_console and get_analytics. 2. Pick the one page with the most impressions that still sits between position 5 and 20, or whose title does not match the query it ranks for. 3. Read that page with read_files and improve its title, meta description, H1, intro, internal links and FAQ so they answer that query better, in the site's own voice. 4. Push the change with write_files, then publish_app and confirm with get_publish_status. 5. Send me a summary: site, page, query, what changed. Never invent statistics or testimonials, never change a site's design, and touch at most two pages per site per run.

It needs Google Search Console connected to the site; if it is not, Kleap hands the agent a consent link for you to open. Start with a Read-only key to watch what the agent proposes before you let it publish. Each agent guide explains how scheduling works for that agent. Key presets live at kleap.co/settings/api-key.

Safe by design

What does the agent get access to?

The screen you approve, and the limits that stay in place after.

Connecting an agent should never feel like handing over the keys. This is the screen you approve, and the limits that stay in place after.

Kleap's OAuth consent screen: Claude wants to access your Kleap account, with seven listed permissions and Deny and Approve buttons
Kleap's real consent screen, captured Oct 18, 2026. We blurred the account e-mail and changed nothing else.
  • apps:readSee your apps and their files
  • apps:createCreate new apps
  • apps:updateEdit, publish and manage your apps
  • messages:createSend prompts to the Kleap AI (uses your credits)
  • tasks:readCheck the status of generation tasks
  • forms:readRead form submissions (may contain visitor personal data)
  • analytics:readRead your sites' visitor analytics
  • Domain purchases are never done by the agent: it returns a payment link that you pay yourself.
  • An API key is optional and scoped: Read-only to inspect sites, Build to change and publish them.
  • A publish is reported as done only once the new version serves, so your agent never hands you a dead link.
  • Revoke any agent from your Kleap settings at any time.
Measured

How long does each call take?

Medians of the journal of every tool call an agent makes.

Kleap now keeps a journal of every tool call an agent makes: which tool, how long it took, whether it worked. These are the medians so far.

Median time per call77 tool calls · Oct 18, 2026, 15:25 to 17:19 UTC
Median time of each Kleap MCP tool call, from Kleap's own call journal, with the number of calls measured
ToolWhat it doesRelative timeMedianCalls
list_app_files0.3 s
get_app0.4 s
write_files0.5 s
read_files0.6 s
get_publish_status0.6 s
edit_files0.7 s
create_app2.6 s
check_task10.8 s
publish_app26.8 s

Small sample, mostly our own test runs. We publish the count so you can weigh it. check_task waits for news before it answers, and publish_app does not return until the new version serves, which is why those two are slow on purpose.

For agents

Which 35 tools can an agent call?

The real tool names of Kleap's MCP server.

The real tool names of Kleap's MCP server, as tools/list returns them on Oct 19, 2026. The same names appear in the prompts and the runs on these pages.

  • Build (5)

    • create_app
    • modify_app
    • check_task
    • retry_task
    • generate_image

  • Files (5)

    • list_app_files
    • read_files
    • write_files
    • edit_files
    • delete_files

  • Sites (6)

    • list_apps
    • get_app
    • find_app
    • rename_app
    • get_screenshot
    • wake_app

  • Publish (2)

    • publish_app
    • get_publish_status

  • SEO (1)

    • seo_audit

  • Leads and traffic (4)

    • get_form_submissions
    • get_analytics
    • get_search_console
    • connect_search_console

  • Domains (4)

    • search_domains
    • check_domain
    • connect_domain
    • buy_domain

  • Database (6)

    • get_database_schema
    • query_database_rows
    • insert_database_rows
    • update_database_rows
    • delete_database_rows
    • run_database_sql

  • Account (2)

    • get_credits
    • upgrade_plan
Questions

Questions about using Kleap from an agent

What is an AI agent website builder?

It is a website builder that your AI agent can operate. Instead of opening an editor, you ask ChatGPT, Claude, Grok or another agent for a site, and the agent calls Kleap to build it, host it and publish it. Kleap supplies the part an agent cannot: hosting, a real URL, forms, analytics and a check that the page is live.

Which agents work with Kleap?

Any agent that speaks MCP or can run shell commands. This site has a setup guide for ChatGPT, Claude and Claude Code, Grok, OpenClaw, Hermes Agent, Codex and Cursor. Windsurf, Cline and other MCP clients use the same URL.

Is this different from asking ChatGPT to write HTML?

Yes. An agent that only writes HTML hands you code to host yourself. With Kleap connected the agent also builds the site on Kleap, publishes it to a live address, connects a domain you own, collects form leads and reads traffic, and Kleap checks the publish really worked.

Do I own the site, and can I leave?

Yes. A Kleap site is an Astro project that you can read and change file by file through the same tools your agent uses. Kleap can back the code up to GitHub, and you can connect a domain you own, so moving on does not mean starting over. Hosting on a Kleap address stays with Kleap.

Do I need an API key?

Not for an interactive connector: it signs you in with OAuth in your browser. Unattended runs, like the ones we recorded with Codex, Grok, OpenClaw and Hermes, use a key instead. A key (created at https://kleap.co/settings/api-key) is for unattended jobs and scripts. Pick the narrowest preset, Read-only or Build, unless the job needs more.

Can an agent spend my money?

No. Buying a domain returns a payment link that you pay yourself. Building and AI edits use your Kleap usage, which you can see in your account, and the free plan includes some.

Can an agent keep improving my SEO automatically?

Yes, if your agent can run a prompt on a schedule. Google Search Console data comes back through Kleap, the agent edits the pages that deserve it, publishes and reports. Each agent page explains how scheduling works for that agent and gives the prompt to paste. Start with a Read-only key to watch what it proposes.

Keep going

More from Kleap

From your agent to the world

Give your agent a website builder

Connect once, then ask in plain words. Kleap builds, hosts and publishes, and your agent keeps the site improving.

Create a free account

MCP, CLI and skills