Keep your frontier model. Put your local GPU to work.

Pair Claude Code, Cursor, or Gemini with your own GPU — or run 100% local and air-gapped. Open-source engine, plus turnkey GPU & vLLM setup for teams.

$ npx -y mcp-castor install --client all

Running a team? We set up your GPUs, vLLM, and agent plumbing in 48 hours →

SESSION KV CACHE · 93.8% HIT · RTX 3090 (W4A16)

Effective Prefill

3,247 tok/swarm + cold

Decode (n=4 / n=1)

260.0 / 70.4tok/s aggregate / sustained

Session KV Hit

93.8%across 2.77B tokens

Local GPU Cost

$0.0000~$0.001 power / token-loop

RTX 3090 (24 GB) · Qwen3.8-27B-W4A16 · vLLM 0.33.0 · 244.6 W · 56 °C

78-Day RTX 3090 Benchmarks →
01 // Architecture & Services01 / 04

Your frontier model shouldn't grep.

Keep cloud models for high-level planning. Run the heavy tool loop on your own GPUs — or go 100% local.

40-Turn Coding Task Trace(18 Files · Auth Refactor & Test Suite)

Same task · three execution modes

Turn 1 (cloud)

~3.2K tokens ≈ $0.01

Turns 2–40 (local)

260 tok/s (n=4) on your GPU

Session KV cache

93.8% hit across the loop

Total cost

~$0.01 · 99.6% under cloud

The frontier model issues one ~3.2K-token planning brief (≈$0.01). The remaining 39 file, AST, and test turns execute on your local GPU at ~$0.0014 electricity each, reusing the session KV cache at a 93.8% hit — power cost, not tokens billed.

01 / Cloud + Local Hybrid

One command to pair

Pair Claude Code, Cursor, or Gemini with local vLLM, llama.cpp, or Ollama in one command.

02 / 100% Local & Air-Gapped

Zero cloud dependencies

Zero cloud dependencies, compile-time offline binary, and 5-layer OS sandboxing for sensitive repositories.

03 / Turnkey Setup & Plumbing

48-hour GPU & vLLM setup

We size your GPUs for the workload, configure vLLM 0.33.0 continuous batching and prefix caching, and wire your team’s workflow in 48 hours.

effective prefill

3,247 tok/s

n=4 concurrent decode

260 tok/s

session KV cache hit

93.8%

local power cost

$10.12

vs 6,044.48 cloud API

02 // Team & Hardware Economics02 / 04

Run the math on your team.

Estimate your monthly token burn, hybrid savings, and GPU hardware break-even.

Workload

Frontier lead

Engineers5 engineers
Monthly volume / engineer150M tokens

Annual net savings

$6,481

Cloud-Only / Mo

$638

Castor Hybrid / Mo

$98

1× RTX 5090 Workstation (~$4,500)

≈ 8.3 months

Cloud-Only Monthly Burn$638
Castor Hybrid (Cloud + $0 Local)$98
Hardware Payback · RTX 5090 Workstation (~$4,500)≈ 8.3 months · Year-1 ROI 1.4×
03 // Self-Serve & Team Engagements03 / 04

Run it yourself today. Or bring us in to wire your stack.

Open-source for individual engineers. Fixed-scope GPU setup, vLLM tuning, and custom plumbing for teams.

Agents

  • Claude Code
  • Cursor
  • Gemini CLI
  • Antigravity
  • Codex

Engines

  • vLLM 0.33
  • llama.cpp
  • SGLang
  • Ollama
  • LM Studio

Models

  • Qwen3.8-27B
  • Ternary-Bonsai
  • any 7B–70B W4A16

Hardware

  • RTX 3090 · 24 GB
  • RTX 5090 · 32 GB
  • multi-GPU vLLM node

Front 01

Open-Source Engine

$0 · AGPL-3.0

The Castor MCP engine, one command.

$ npx -y mcp-castor install --client all
  • 3 MCP tools, 5-layer sandbox
  • 27-grammar AST syntax gate
  • vLLM, llama.cpp, Ollama profiles
78-Day RTX 3090 Benchmarks →

Front 02

Turnkey GPU & vLLM Setup

$1,500 · 48-Hour Sprint

GPU sizing & vLLM setup in 48 hours.

  • Hardware & W4A16 quantization sizing
  • vLLM continuous batching, n=4
  • Team routing for Claude Code

Front 03

Custom Agent & Workflow Plumbing

Scoped Build / Retainer

Your stack, wired to your agents.

  • Shared multi-seat GPU pool
  • Custom sandboxed agent tooling
  • CI pipelines for agent work

Front 04

Private & Air-Gapped On-Prem

Regulated Pilot

Zero-egress, audit-evidenced, on your hardware.

  • 100% local, zero-egress deployment
  • 5-layer OS sandbox, audit ledger
  • Audit logs designed for NIST 800-171 / CMMC L2

Scope & Engagement Brief Builder

Pick your front, team size, and GPU status — copy the brief or launch a pre-filled email.

Front

Team size

GPU status

Launch Pre-Filled Email →