Pricing
Priced on calls, because that is what costs us
Everything is free during preview. Billing is not built — there is no card field and no meter that could charge you. What follows is the model we intend to run, published early so nobody is surprised later. A page arguing that we refuse to ship things that do not work cannot also advertise a billing system that does not exist.
Free
One crawl, and enough calls to prove it works
- Crawls
- 1
- Calls / month
- 1k
- Then
- —
- One published tenant
- Hosted MCP endpoint and WebMCP page
- Every capability probed against the live site before it ships
- Weekly health check
Calls stop at the allowance rather than billing you by surprise.
€29/month
For one business that wants to be reachable by agents
- Crawls
- 3
- Calls / month
- 50k
- Then
- €1.5 / 10k
- Three published tenants
- Daily health check with drift detection
- Model-written tool names and descriptions
- Domain verification, so the tenant is yours on the record
- Email support
Overage is billed per 10,000 calls, and a cap you set is honoured.
€149/month
For groups and chains, and anyone re-crawling often
- Crawls
- 15
- Calls / month
- 500k
- Then
- €1 / 10k
- Fifteen published tenants
- Re-crawl on a schedule, with regression blocked automatically
- Alerts when a capability stops verifying
- Custom vertical recipes
- Priority crawl queue
Volume pricing starts here; talk to us before you outgrow it.
Talk to us
Self-hosted, white label, or an SLA
- Crawls
- —
- Calls / month
- —
- Then
- —
- Run the whole thing on your own infrastructure
- Your domain on the tenant surfaces
- Uptime and response commitments
- A data-processing agreement
The code is MIT; the hosting, the crawler and the support are what you would be buying.
Two units: a crawl, and a tool call
A crawl is work we do each time you ask — a pass to name and describe the tools, and about ninety seconds of headless Chromium — so a plan includes a number of them and a re-crawl spends one. Serving is different: one cached fetch of your own site per call, with a thirty-minute TTL in front of it, which is why calls are metered rather than bundled.
Charging per site would have meant re-reading a site every week for free while somebody who crawled once and left paid the same. Crawls are the work; calls are the use; nothing is charged for a tenant merely existing.
What a call is
- One
tools/callagainst your hosted MCP endpoint, or one tool invocation on your WebMCP page. Both count the same. - A cache hit still counts. It cost us a request either way, and pretending otherwise would make the meter unpredictable in the direction that helps nobody.
tools/listis free. An agent reading what exists is not usage.- A capability that returns nothing is not billed. It also should not have shipped.
Why not per site
A flat fee per site is easier to sell and it decouples the price from what we spend, which means one busy tenant quietly subsidised by everyone else. Metering calls keeps that honest in both directions: a shop that agents hammer pays for it, and a cinema that gets asked about showtimes forty times a week stays on the free tier indefinitely.
The code is free either way
ClawHarbor is MIT-licensed and the repository is the whole system — crawler, manifest schema, binding runtime, both renderers. You can run it yourself and pay nothing. What is for sale is the hosting: the crawl queue, the browser fleet, the health checks that notice when a site changes shape, and somebody to email when it breaks.