kibbuFor agents .md
For agents .md

Ship more AI featureswithout worryingabout margins.

Try it on one featureGet an API key

kibbu runs AI workloads on idle laptops your company already owns. You buy less cloud inference, and cloud stays available when a job needs it.

Works best forThe AI jobs nobody is sitting there waiting on, like evals, embeddings, background agents and overnight data jobs.
Why it works

Here’s what changes when your own machines do the work.

  1. Your costs stay flat as you grow

    When your own machines handle the work, you’re mostly paying for electricity instead of per-token fees, so the bill doesn’t have to climb every time you add users.

  2. It only uses idle time

    kibbu only uses a machine while nobody’s on it, and it steps back as soon as someone sits down.

  3. Your code stays the same

    You change the base URL and the API key, and the rest of your code stays the way it is.

How it works

7 min from now, your first AI workload runs locally.

  1. Sign up

    With your work email.

    1 min

  2. Add your first node

    Through the admin console.

    2 min

  3. Get an API key

    Change your current API key on one workflow with kibbu’s.

    2 min

  4. Configure kibbu with your preferences

    Like what models to run and how to show you data.

    2 min
The kibbu workshop

What have you got?

Your machine

18 GB available memory of 24 GB · MacBook Pro M5 Pro

8 of 13 models fit this estimate.

Google · GGUF Q4_K_MFits estimate
≈ 3.43 GB needed≈ 63 tok/s
From 307 GB/s of memory bandwidth and 3.43 GB of weights read per token. A rough single-request figure, not a measurement.

Memory and speed estimates depend on model settings and context length. Speed is a rough single-request figure, not a measurement.

Explore the models13 models
  • AlibabaQwen 3 Embedding - 0.6B≈ 0.64 GB needed≈ 336 tok/sFits estimate
  • GoogleGemma 4 - Compact≈ 3.43 GB needed≈ 63 tok/sFits estimate
  • GoogleGemma 4 - Small≈ 5.34 GB needed≈ 40 tok/sFits estimate
  • GoogleGemma 4 - 12B≈ 7.38 GB needed≈ 29 tok/sFits estimate
  • MetaMuse Glimmer≈ 16.76 GB needed≈ 10 tok/sLimited headroom
  • GoogleGemma 4 - 26B MoE≈ 16.8 GB needed≈ 67 tok/sLimited headroom

Memory and speed estimates depend on model settings and context length. The speed shown is a rough single-request figure derived from memory bandwidth, not a measurement.

Running suitable workloads on machines you already own can reduce the inference you buy from the cloud. Bring your own numbers and we will work through the rest.

Try it on one feature
Built with privacy in mind

We take your data seriously.

Using AI shouldn’t mean giving up control of sensitive information. kibbu helps you decide where your workloads run and how your data is handled.

Talk to us about privacy

  1. Zero data retention

    We don’t store your prompts or responses. Operational metrics are handled separately.

  2. You decide what reaches the cloud

    Choose which workloads can use external providers and which must stay within your environment.

  3. On-prem when you need it

    For stricter requirements, deploy kibbu within your own infrastructure.

Good questions

Questions we get before every rollout.

What happens to our AI costs as we add users?

Work that runs on your own machines costs you electricity instead of a per-token fee, so it doesn’t grow with every new user the way a cloud bill does. Net savings still depend on workload fit, how much idle capacity you have, kibbu fees and operating costs, and anything your fleet can’t take goes to the cloud as usual.

How much can we save?

It depends on your models, usage and hardware. Start with one representative workflow, compare its total cost and results, and use that measurement to estimate the wider opportunity.

Which workloads should we start with?

Recurring, latency-tolerant jobs: document processing, enrichment, embedding batches and background agent steps. Check quality and deadline requirements before moving each one.

What isn’t a good fit?

Anything a user is actively waiting on, like live chat or other real-time features. Those are better off staying on your current setup, and kibbu sends them to the cloud anyway.

Will we need to change models?

Possibly. Local execution uses models supported by kibbu and your hardware. A proprietary cloud model may need to stay with its provider. Evaluate any replacement against your own acceptance criteria.

Can we keep using cloud models?

Yes. Use cloud models for workloads that need them, subject to supported provider integrations and your routing policy. Local execution comes in workflow by workflow.

Do we need new hardware?

Start by assessing what you already own. Macs, Windows and Linux, with or without a GPU. Model size, available memory and job deadlines determine which workloads your fleet can handle.

Does this run on employee laptops?

Only the ones that are idle and plugged in, and only the ones you allow. kibbu watches battery and temperature and steps aside the moment someone sits back down.

Can sensitive workloads stay local?

Choose a deployment and routing configuration that keeps those workloads within your approved environment. Confirm logging, telemetry and fallback behavior as part of the deployment review.

Pick one AI feature and we’ll show you what it could save you.

We’ll run that feature’s inference on your idle machines, then walk you through how the cost compares with what you’re paying now.

Try it on one featureGet an API key