> /inferhaven
Your own private AI coding server — local models, your editor, your terminal, on hardware you control. No per-token meter, no code leaving your perimeter. Self-host it, or boot one in our cloud.
/TWO JOURNEYS, ONE CHARTED COURSE
Both stacks run the same workspace, the same models, the same tools. The difference is who owns the ship, you're still the captain.
/core · self-host
The full self-hostable stack. Ollama, terminal-first workspace, code-server IDE, OpenAI-compatible API. Your code never leaves your machine.
/cloud · personal workstation
A persistent private dev environment with secure access, durable storage, and optional GPU power on demand. Boots in seconds, picks up exactly where you left off — no session roulette, no token meter.
/WHY INFERHAVEN
/privacy by default
Inference can run entirely on your hardware — local on Core, your private workstation on Cloud. Nothing leaves your perimeter unless you opt it in. Zero telemetry. No public listeners.
0 bytes shipped to vendor APIs unless you explicitly say so.
/YOUR TOOLS, YOUR METHODS
OpenAI-compatible gateway on localhost. Point Cursor, Cline, Continue, Claude Code, Aider, OpenCode, Goose, Avante, Qwen Code, Pi — all at the same local models. InferHaven isn't competing with your assistant; it's the harbor every one of them docks in.
Up to 10 assistants installed and pre-pointed at your models; 7 re-synced automatically on every model pull.
/PERSISTENCE
Flat monthly. Persistent home, persistent secrets, persistent shell sessions. No per-token meter, no session timeouts, no "your container was reclaimed" at 2am. The closest thing to owning the machine without buying the rack.
Subscription that feels like ownership — not a metered tab.
/WORKS WITH
InferHaven speaks the OpenAI schema natively, so the IDE plugins, CLIs, and agents you already use talk to it with a one-line endpoint swap. It isn't competing for your editor — it's the harbor they all dock in: your models, your machine, every tool pointed at the same local gateway. Privacy by default, security baked in.
/swap one line
# point any OpenAI SDK at your local InferHaven
export OPENAI_BASE_URL="http://localhost/v1"
export OPENAI_API_KEY="haven" # local server, any string works/WHERE IT FITS
Most things you'd compare it to actually run on top of it. The only real question is whether you build the stack yourself, or boot it whole.
/THE STACK, ASSEMBLED
Each of these is something people self-host on its own. InferHaven is all of them — assembled, wired together, and kept in sync.
/SEE IT RUN
From git clone to the first model response in ~3 minutes.
Point any OpenAI-compatible tool at localhost/v1. Your PC / laptop is the model server.
/TRY IT
No account, no credit card, no email gate. Pick the path that's least friction, and be inferencing in under two minutes.
Clone the repo, run docker compose up — or reopen the same devcontainer in VS Code. Runs on your machine, uses your own GPU.
/RESERVE YOUR WORKSTATION
A persistent dev environment with optional GPU, in the region you choose. No credit card to reserve. We email you when your tier opens — founding members get our best early-adopter pricing, and we'll always give advance notice before any change.
See tentative tier breakdown ↓
/PRICING · TENTATIVE
Start free, then pick the GPU power you want. No per-token bills, no minute-by-minute rental — your workspace stays yours.
Pricing below is tentative — final tier names and numbers may shift before billing opens in beta. Reserve a spot for founding-member pricing: we'll honor the most generous rate we can sustain and always give advance notice before any change.
Try a real workstation with a tracked GPU-hour budget. We verify your card to prevent abuse but never bill until you choose to upgrade. Convert any time to keep your work; let it expire and your workspace archives for 30 days.
A persistent dev environment with secure remote access, durable storage, and a gateway for your own LLM keys. No GPU included.
More room and a second seat for solo devs with bigger projects or small teams sharing a single environment. Same persistent dev workstation, more headroom.
Modest GPU access for local coding models, code assistants, embeddings, and occasional inference. Generous fair-use.
Stronger GPU access with priority scheduling for larger local models and heavier coding assistant workflows.
Reserved single-tenant GPU — 20 GB VRAM, your machine alone. Comfortable home for 14B local models or 32B with Q4 quant. No sharing, no queue, no per-token meter.
Reserved single-tenant GPU with 48 GB VRAM. Run 32B at Q8, 70B at Q4, large context windows. The ceiling for serious local-model AI dev work.
For teams with specific compliance, data-residency, or hardware needs — regional single-tenant placement, bare-metal isolation, 80 GB+ class GPUs, and negotiated SLAs. We scope it to you.
For teams that already have AWS, GCP, Hetzner, or on-prem GPU capacity. We install our agent, run your InferHaven dashboard, monitor and update. You pay your own cloud bills direct — we just manage.
All cloud tiers include daily backups, encrypted storage, your data stays in the region you pick (GPU regions roll out through beta), and direct support is included on every tier. Fair-use means no per-token bills and no session timeouts — generous, with throttling only for clearly non-interactive abuse. If you outgrow Studio, Dedicated is built for you. common questions ↓
/FAQ
A persistent private workstation — like a dev machine in the cloud that's yours. SSH in, write code, run your tools, pick up the same tmux session tomorrow. Optional GPU tiers add local model inference on top.
Sign up with no charge — we verify your card to prevent abuse but don't bill until you upgrade. You get 14 days with a full Workspace plus 7 GPU-hours of compute, tracked live in your dashboard so you always know what's left. Upgrade any time to convert your environment to a paid tier — your code, secrets, and dotfiles carry over untouched. If you don't upgrade by day 14, the workspace archives for 30 days; resume in that window or it's purged. Note: hourly GPU limits and trial allotments are tentative and may change before beta release.
Shared GPU tiers run on a pool with generous fair-use — no per-token bills and no session timeouts, unlike metered or session-limited services. We keep it fast by fair-scheduling under load, and only throttle workloads that are clearly non-interactive (24/7 max-throughput batch or training). Dedicated tiers are single-tenant — your card, genuinely unlimited. Workspace has no GPU, for cloud-model users.
No. Flat monthly subscription. You're paying for the workstation, not the compute meter. The only variable cost is anything you opt into yourself, like an OpenAI or Anthropic API key you choose to wire up.
Paid tiers: your home directory, secrets, and storage are preserved for 90 days while paused — life happens, and we don't want a hard month to cost you your environment. Resume anytime in that window and pick up exactly where you left off. Free trial workspaces archive for 30 days after expiry. After the retention window, data is purged unless you've exported it.
You pick the region at signup. Storage is encrypted at rest; the network is inside our VPC and never exposed to the public internet by default. GPU availability expands by region through beta, so some GPU tiers start in a subset of regions. Your code is yours — we don't train on it, scan it, or share it.
The IDE is just one surface. Underneath, you get a real persistent workstation: SSH, tmux, root, your dotfiles, your tools, your shell. You're not in someone else's sandbox — you're in your own machine.
You absolutely can — and if you do, you've built the first two rooms of what InferHaven ships whole. We're Ollama plus a terminal-first workspace, a browser IDE, auto-HTTPS, SSH key management, and up to ten coding assistants installed and pre-pointed at your local models — re-synced automatically every time you pull a new one. The DIY stack is a weekend of wiring and a maintenance tab that never closes; this is `docker compose up -d`. Same parts, assembled, kept in tune.
Tabby and friends are excellent at one job: code completion from a self-hosted model. InferHaven isn't trying to be a better completion engine — it's the whole machine that engine runs on. You get the model server, the workspace, the IDE, and your pick of assistants (Continue, Aider, Cline, Goose, and more) wired up at once, so you can switch between them or run several at once. If all you want is autocomplete, Tabby is a clean choice. If you want a private coding box you live in, that's us.
Honestly? For the hardest reasoning, the frontier cloud models are still ahead — usually by a few months. But for day-to-day coding, Qwen 2.5 Coder, DeepSeek, and Llama on right-sized hardware are genuinely good, and the gap closes every release. And InferHaven isn't local-only: wire up your own Anthropic or OpenAI key and route the gnarly refactor to the cloud per-task, while everything sensitive stays on your box. You pick what leaves, and when.
Your hardware tier, your call. Pull any model the GPU class supports (Qwen 2.5 Coder, Llama 3, DeepSeek, etc.). The OpenAI-compatible gateway speaks the same protocol so every harness -- Cursor, Claude Code, Continue, Aider, OpenCode, and more -- works without reconfig.
InferHaven Core is the same stack, self-hostable on any machine with Docker. The Cloud product is built ON Core — you can move between them without rewriting anything. Many users will run Core on a home GPU and the Cloud workstation for roaming or specific access needs.
InferHaven Core ships under FSL-1.1-Apache-2.0 (Functional Source License). The code is on GitHub and you can read, run, modify, and self-host it today — but it's source-available, not OSI open source. Each release converts to Apache-2.0 automatically two years after publication. FSL restricts one thing: building a managed-hosting service that competes head-on with InferHaven Cloud during that 2-year window. Everything else — personal use, internal commercial use, modification, consulting — is permitted.