BosunCadre BOSUNCADRE
BosunCadre compass mark
In active development · self-hosted today

Your hardware. Your models. Your control plane.

BosunCadre coordinates a mixed pool of local LLMs and GPU machines from one place — routing jobs, render and training work, and chat to whichever box can handle them. It runs on your own network, so your work, prompts, and files never leave.

How it works

A coordinator, a fleet of runners, and smart routing.

One passive coordinator holds the work; runners on your machines claim what they're capable of; GPU-only boxes are driven over SSH — no agent to install.

01 — Coordinator

Holds the work

A single process keeps the task queue, worker registry, file store, and a cost ledger. It routes work but doesn't run inference itself.

02 — Runners

Claim by capability

Daemons on machines with a local model register, claim tasks they can handle, execute, and report back. They self-update and stay host-guarded.

03 — SSH drive

Reach every GPU

GPU boxes without a local model are driven directly over SSH for render and training jobs — nothing extra installed on the target.

Built and verified

What it already does.

Every capability below has been exercised end-to-end on real hardware — real local models, a real GPU box over SSH, a real repo, real files.

Capability-based routing

Each task is matched to a capable worker and model across a mixed pool, with per-task model routing and auto-detection of a box's installed models.

Agentic code executor

Produces a tested change on a branch — tool loop, apply, build, test — behind a review step and a human land gate. Nothing auto-merges.

GPU render & training

Drives render and training jobs on GPU machines over SSH, with file transfer to endpoints even without git — staged out, outputs captured back.

Cost ledger & benchmarks

Records local token usage against a measured frontier-equivalent, with a benchmark harness that runs real local-vs-frontier A/B comparisons.

Secrets at rest

Recoverable secrets encrypted with AES-256-GCM, with the key held outside the database and its backups.

Review tooling

A configurable review chain lets a capable AI session or a human review the executor's diffs and post a verdict through the same designed path.

Shared-GPU arbitration

Local LLMs and image tools share a single GPU without colliding — VRAM is arbitrated between them so one card can serve multiple workloads instead of sitting idle or crashing.

Offline licensing

Offline-validated, tiered, time-limited keys node-locked to a single machine, applied to a running coordinator without a restart.

Cost & measurement

Draft locally. Let the frontier review.

On a frontier API, producing output costs roughly 5× as much as reading input. BosunCadre does the expensive part — generating the draft — on a free local model, then has the frontier only review the result, a near-zero-output approve or reject cost. The costly generation shifts to $0 local inference while a frontier model still gates quality.

up to 81%
lower LLM cost with a qwen3.6:27b local model, at equal quality — on the tasks it handles well

This isn't a projection — it's a measured A/B against a funded frontier key (Opus 4.8) on real tasks, with quality graded by an objective pass/fail check on each result. These numbers were measured with a qwen3.6:27b local model; a stronger local model raises the local pass-rate and pushes the savings higher.

The 81% is a best case, not a blanket average — it’s the saving measured on the tasks the local model solved at equal quality (80.9% and 81.0%). Work it can’t carry routes straight to the frontier at full price, with no saving there, so your blended savings depend on how much of your workload the local model can take. Better local hardware handles more and pushes the savings higher.

On one representative task, the frontier produced ~5,200 output tokens doing the work directly, versus ~4 as a reviewer of the finished draft. That gap is the whole mechanism.
TaskFrontier-directOffloadSaved
light task$0.116$0.02280.9%
moderate task$0.159$0.03081.0%
complex task$0.350routes up

Honest caveats: the sample is representative but small, and the saving is realized only when the local draft is correct — the frontier review gates quality either way. The trade-off is latency (local inference is markedly slower; a stronger GPU raises the local pass-rate and closes the gap). More complex work routes to the frontier; task size is handled separately by decomposition, which breaks a large job into smaller subtasks the local model can carry. Extrapolating from the validated mechanism: roughly 50–65% on typical tasks, up to an ~80% ceiling on write-heavy ones.

Editions

From a single box to a dedicated fleet.

Feature placement and pricing aren't finalized — this is the current design direction.

← swipe to see all editions →

CapabilityCommunitySoloTeamBusinessEnterprise
Machines · seats · projects1·1·13·2·515·15·∞40·50·∞Unlimited
Routing · executor · land gate · cost ledger
Continuation · decomposition · frontier switch
Review chainsinglemultimultimultimulti
Remote access (direct P2P)
Teams · NAS · work-tracker integrations
Governance (notify · budgets · policies)
SSO (OIDC + SAML) · SCIM · multi-tenancy

Self-hosted perpetual licences with optional annual maintenance. See the full pricing & feature comparison.

Your network is the boundary.

BosunCadre runs on your own hardware. Your work, prompts, and files stay on your network — the passive coordinator routes work, it doesn't ship your data anywhere.

Remote access without exposure.

Planned direct peer-to-peer access reaches a self-hosted server from anywhere via a rendezvous handshake — no port forwarding, no public server, and data does not transit the vendor.

Request early access.

BosunCadre is operationally live on its author's own machines and moving toward a wider release. Leave your email to hear when early access opens, or to talk about a self-hosted deployment.

A product of Acribayik Synapsis LLC. We'll only use your email to talk to you about BosunCadre.