Pricing
MonkeyLLM is open source and self-hosting is unrestricted: there is no paid tier hiding a feature. What you can buy is a counterparty: a commercial license when AGPL-3.0 does not fit, a support plan when uptime is somebody’s job, and paid sessions when you want an architecture answer before you commit.
Free
The engine under Apache-2.0; the Station, Studio and Clipper under AGPL-3.0-only. Every feature of the product, with no seat count and no key to ask us for.
Custom
The copyright is held by a single author, so the AGPL is not the only way to get this software. Embed the Station in a closed product, or offer it as a managed service without opening your stack.
Support & maintenance
Month to month, cancelled whenever you like. Subscribed from the client portal; billing runs through Stripe.
For teams already self-hosting who want a second pair of eyes on upgrades and security.
$490/month
For teams running a forest in production, with a direct channel and hands-on upgrades.
$1,490/month
For forests the business depends on: a response SLA, a named engineer and a say in the roadmap.
$3,900/month
Consulting
$250· 60 minutes
A paid, focused session on your corpus, your hops and your governance constraints. Requested from the client portal: you describe the topic, we price and schedule it, and you pay only once the scope is confirmed.
The standard price is shown above and set per deployment; the portal always shows the exact price before you pay.
FAQ
Yes, and unrestricted. The engine is Apache-2.0; the Station, Studio and Clipper are AGPL-3.0-only. Running an unmodified copy for yourself or your organisation triggers no obligation at all: no notice, no publication, nothing to buy.
Less than you might think. Entry search is BM25 over SQLite: no embeddings, no GPU. For local inference, a quantized 12B (the class our benchmark uses) fits a mid-range 12 GB card like an RTX 3060, which is exactly the card we measured on; 8 GB cards work with tighter quantization or partial CPU offload via llama.cpp. Or skip local hardware entirely and point the chat binding at any OpenAI-compatible endpoint, OpenRouter included. The benchmark rig is that mid-range card, not a requirement above it.
When AGPL-3.0 does not fit your distribution model: embedding the Station or the Studio in a closed product, or offering a modified version to other people over a network without publishing its corresponding source. Running it internally, modified or not, never requires one.
Review of updates and security patches, a defined response time, help planning and executing upgrades on your own infrastructure, and a review of your ingestion pipeline. It is engineering support for a deployment you own. It is not hosting, and it is not access to a feature the open-source build lacks: there is no such feature.
No. They are separate: a support plan buys response time and engineering attention, a corporate license changes what you are allowed to do with the AGPL-3.0 components. Most teams need only one of the two.
Nowhere. You host the Station, the forests live in volumes you own, and you choose the model bindings, including a purely local llama.cpp endpoint. This website’s client portal stores only what you send it: your email, your company, your meetings and your invoices.
Plans are billed monthly in advance through Stripe from the client portal, and you can cancel there at any time; the plan stays active until the end of the period you already paid for. Meetings are one-off payments made only after we confirm the scope.
That is the intended path. Clone the repository, build the fixture forest and run the benchmark: reproducing the 0/11 result takes one command. Buy a counterparty only once you have decided the architecture serves your case.
Reproduce the benchmark first: it is one command. Then talk to us about the parts you should not have to build or operate yourself.