Add private AI to your clients without moving their data.

You do. You introduce us, you stay in the room, and the account, the renewal and the first call when something changes stay yours. We do not sell help desk, network or security operations, and we are not building a services arm that will. Our job is to make the managed contract you already hold harder to leave.

Kirk, by default. Tuning, connector work, model upgrades, security patches and day-to-day operations are ours under a subscription, reported against the measure agreed at the start. Your team keeps everything around it: identity, network, backup, monitoring and the client conversation. Partners who would rather operate the platform themselves can; it is the same open-source system either way, and we train your engineers as part of the handover.

Whichever the client approves, running as a file on a GPU they own, not as a service somebody else meters. The default general model today is Google’s Gemma 4 31B, open weights, on one dedicated GPU in the client’s tenant with a quarter-million tokens of context; it is the same model, on the same class of card, in every production system we run. When a job wants something else, we use something else: a small open model where small is enough, specialist speech models for voice, a flagship-class open model such as GLM-5.2 for the hardest reasoning. The platform does not change when the model does.

No. It runs on the cloud the client already buys from you or through you, and the GPU, storage and network consumption sit in their account, under your practice. Copilot stays where it is: it is a per-seat assistant inside Microsoft 365, and FlatClaw is a platform for agents that work across the client’s systems and data at a flat rate. Clients who have both use both, and the cloud line item grows on the bill you already manage.

One workflow, one client, about six weeks to a working deployment. We scope it together and agree the measure before we start. We stand the platform up in the client’s account, take the security review, build the first agent and document what was built. Fixed scope, fixed price, quoted after a short discovery. You bring the next workflow and the next account, and both ride on the tenant the client already pays for.

In writing, before any client conversation. Two shapes so far. Referral: you introduce, we deliver in the client’s account, and you earn on the introduction. Resale: the platform becomes a line item on your agreement, you buy it at a flat cost per tenant and set your own price, standalone or bundled, and we build the workflows and train your engineers. Project and recurring revenue are shared, and the split is agreed per partner. Ask us and we will be specific.

Because an open-weight model is a file. It does nothing until you run it, it runs where you put it, and nobody can change it, meter it or train on what passes through it. That is the privacy claim, and it is mechanical rather than contractual: your client’s documents, prompts and transcripts never leave their account, and you can prove it with a packet capture on the segment.

Yes, and we do. FlatClaw’s agent harness is built on the same minimal open coding-agent core our own engineers work in every day: an agent on a loop that reads a repository, changes code, runs the tests and goes again, on private inference, with the repository never leaving the client’s tenant. Code is where open models have closed the gap fastest. GLM-5.2, MIT-licensed with a million-token context, is the one we keep a build recipe for, and we benchmark candidates against the client’s own repositories before choosing.

One GPU per tenant, about two thousand dollars a month of cloud at published list rates, billed by their cloud provider on their own account, and every workflow on it at that one flat rate. Kirk’s subscription for operations and upgrades sits on top and is quoted per tenant. There are no per-token, per-minute or per-seat charges from us, ever.

We swap it. The platform does not change; the model file does, and model changes are ours to make and report on under the subscription, which includes a twice-yearly inference review whose job is to find exactly those opportunities. A client who standardizes on the platform is not standardizing on this year’s model, and neither are you.