AI and agents

Your model, your costs, your agent.

Caelo works with the AI you choose. Use its assistant with your own key, or connect your own agent and keep the thinking on the plan you already pay for.

Connect your agentLive editing

Three ways

Pick where the thinking happens.

All three work on the same site with the same tools and the same rules. They differ in who runs the model and who pays for it.

In Caelo

Caelo's assistant

Chat in the admin or click an element on your live page. Caelo calls the model with your own API key, and every call lands in the cost dashboard. This is how your team edits day to day.

  • Live edit
  • Chat
  • Your API key

From your client

caelo_chat over MCP

Talk to Caelo's assistant from Claude Code, Claude Desktop or another MCP client that runs local servers, with nothing but a token. Do research and drafting in your own tool, and let Caelo build the page in your design. ChatGPT and web connectors are not supported yet.

  • MCP
  • Token

For builds and migrations

Your own agent: Power-MCP

Connect Claude Code or a comparable agent. It gets the tools Caelo's assistant uses, and the thinking runs on your Claude subscription or another plan you already pay for. Caelo makes no AI calls of its own on that connection; tasks like translations come back to your agent. We build and maintain caelo-cms.com this way.

  • Claude Code
  • Claude Desktop
  • Admin token

$ claude mcp add caelo-admin \
  --env CAELO_ADMIN_URL=https://admin.example.com \
  --env CAELO_MCP_TOKEN=mcp_… \
  -- bunx --package @caelo-cms/mcp-server@latest caelo-admin-mcp
# copy the exact command from Settings, MCP in your admin; the server switches to your install's version on its own

Caelo on Caelo

This website is built by an agent. A person publishes it.

caelo-cms.com runs on Caelo. Most of it was built from Claude Code over the Power-MCP. The agent drafts and stages; going live stays a person's click.

  1. Settings, External AI tools: an admin-scope MCP token for Claude Code with its expiry date and a Revoke button.
  2. A decision the agent requested, waiting in the Caelo admin for an editor to click Approve or Reject, with the agent's reason.

Screenshots from the admin of this website, October 2026.

Transcript

The session that put these screenshots here

This is the agent's side of the session that added the product tour to this website, shortened to the tool calls. The request came from a person in Claude Code; every line below is a call to Caelo over the Power-MCP.

> Put the product tour on the product pages, with the
  screenshots from the demo site.

● caelo_open_session        resumed "Website-Redesign"
● caelo_get_context         site model, playbook, skills index
● find_media "tour-"        19 screenshots with alt text
● add_module                new modules: Product tour, Tour step
● create_content_instances  19 steps in one transaction
● add_module                tours on /product/control, /product/system,
                            /migrate and /ai
● screenshot_page           desktop and phone, checked
● stage_changes             staging rebuilt, quality check running
◐ Publish live              waiting for a person

The thinking ran in Claude Code, on a subscription we already pay for. Caelo made no AI calls of its own for this work; the cost dashboard in the admin showed $0.00 for the week.

Providers

Bring the provider you trust.

You choose a provider and a model when you set Caelo up, and can change both later. Every provider sits behind the same layer, so switching does not change how your site is built.

AnthropicClaude models with your Anthropic API key.
OpenAIGPT models with your OpenAI API key.
GoogleGemini models with your Google API key.
LocalAny OpenAI-compatible endpoint, for example a model on your own hardware.
ImagesImage generation through Google or OpenAI, when you want it.

Same rules

An agent gets the tools, not the keys to production.

Your own agent is not a back door. It works under the permissions of the token you give it and stops at the same gates as Caelo's assistant.

Drafts and approvals
Draft onlyEvery agent writes into the shared draft, like the people on your team.
ApprovalsActions that are hard to undo wait for a click in the Caelo admin, whoever asked for them.
PublishingAn agent can stage and read the quality check. Publish live stays a person's click.
ContextOne call hands the agent the site model, the playbook, the brand voice and the skills, so it works the way Caelo's assistant does.

Costs

See every cent the AI costs.

Caelo itself is free and makes no money on your AI usage. When your team uses Caelo's own AI, in the chat or in live edit, each call runs on your own API key. You pay your provider directly, every amount shows in the cost dashboard, and budgets cap it. For scale: our automated real-AI test builds a small branded site and edits it across 14 scenarios with Claude Sonnet, and a run typically costs a few dollars. Building a large site from scratch or moving one with many pages takes far more tokens. For that we recommend the Power-MCP with Claude Code or a comparable agent, on a subscription you already pay for. The thinking runs there, and Caelo makes no AI calls of its own. Your team then edits day to day in the admin.

Hosting costs
Cost dashboardEvery AI call with its model and cost, under Settings, Costs.
BudgetsLimits under Settings, AI budgets. A budget warns or blocks further calls when it is used up.
EditorsYour team sees "AI", not a vendor. The provider shows up only in the owner settings and the cost dashboard.

Start

Connect your agent in a few minutes.

Create a token, run one command, and your agent can work on your site.

MCP and Power-MCPQuickstart