[ edition / self-hosted ]

Works where you work.

Sidekick Self-Hosted runs inside your perimeter — on your hardware, against your models. Binaries, prompts, findings, and notebooks never cross the boundary.

  • 01

    Air-gapped work

    Classified, isolated, or export-controlled environments where no outbound connection is permitted.

  • 02

    IP-sensitive reversing

    Artifacts that must not be uploaded anywhere.

  • 03

    Regulated industries

    Finance, healthcare, defense, and government engagements with contractual limits on data residency and processing.

❯ perimeter

Nothing leaves the boundary.

Cloud deployments route prompts and context to hosted model providers. Self-Hosted never does. Every artifact Sidekick touches stays on infrastructure you control.

⟦ your perimeter ⟧
client Binary Ninja + Sidekick plugin
data Binaries · Notebooks · Findings
service Sidekick server
inference Your models · Your GPUs
❯ parity

Same workflow, either way.

The deployment boundary changes. The product does not.

  • 01

    Same agents

    Search, simplify, surface, and verify behave identically to cloud. No feature flags, no reduced capability tier.

  • 02

    Same notebook

    Findings accumulate across sessions with evidence links intact. The artifact you build here travels with you.

  • 03

    Same verification

    Background validation cross-checks every finding against the binary. Runs locally, on your hardware.

  • 04

    Same API

    Script agents, chain workflows, embed Sidekick in pipelines. One surface across deployment models.

❯ runs on your stack

Your hardware, your models.

Sidekick routes different operations to different models. Hardware sizing is driven by concurrency, context window, and tokens-per-second — how many analysts, how much history they carry, how much background verification — not by raw model count.

Hardware tiers

  • Floor
    96 GB VRAM · 3 concurrent
    e.g. 1x RTX 6000 Pro Blackwell — comfortably hosts both GPT-OSS 120B and GPT-OSS 20B at the same time, routing tasks between them. The higher GPU memory bandwidth also helps keep interactive workloads feeling faster. Fits a single analyst or a small shared team.
  • Team
    192–320 GB · 4–10 concurrent
    e.g. 2× RTX 6000 Pro Blackwell, or 2–4× H100. The added memory bandwidth gives you more headroom for longer context windows under load and for background verification to run alongside interactive chat.
  • Fleet
    640 GB+ · 20+ concurrent
    Single dense node — e.g. 8× H100 or H200. Sized for headless agent pipelines running alongside interactive analysts: long contexts, long histories, batch triage at throughput.

Open-weight models

Sidekick can route work across modern open-weight model families that you choose to host on the inference server you already trust — vLLM, TGI, or any other OpenAI-compatible endpoint. Different tasks route to different models: generation, verification, and analysis each pick the model best suited to the job.

  • GPT-OSS 20B / 120B
  • Gemma 4 4B – 31B
  • Qwen 3.6 35B-A3B
  • Nemotron 3 – 500B

The families listed above are bring-your-own options for customers who want to host their own models. Our deployment package currently ships with recommended GPT-OSS configurations, and you can also connect your own weights or fine-tuned models.

❯ see it live

Contact us for a demo.

Every environment is different — the hardware you have, the models you trust, the boundary you enforce. We set up Self-Hosted trials two ways.

  • in-person

    On-site demo

    We come to you. Bring your hardware, your analysts, and a binary you actually care about.

  • virtual

    Hosted Self-Hosted demo

    We spin up an inference server in a neutral environment so you can evaluate the on-prem experience without shipping GPUs.

Request a demo