// agentic coding · windows x64 · free & offline
A complete agentic coding IDE that runs open models on your own GPU with a built-in llama.cpp engine. No account, no cloud, no telemetry, and no server of ours anywhere in the path. Prefer a cloud model? Paste your own key and it goes straight to the provider — never through us.
// choosing a model
Most AI IDEs start in the cloud and treat local as an afterthought. OrangeIDE is built the other way up — and never puts itself between you and a provider.
A bundled llama.cpp engine runs open models — DeepSeek, Qwen, Llama, GPT-OSS — straight from Hugging Face. Pick a quant, and the autotuner splits layers across GPU and RAM for your exact hardware.
Chat, agent mode, RAG, terminal: all of it works with the network cable unplugged.
free forever · offline foreverPaste your own API key for OpenAI, Anthropic, Google, DeepSeek, Grok, and more. Calls go straight from your machine to the provider — we never proxy them, never see them, never charge for them.
Bring-your-own-key is a feature here, not a business model.
free feature · direct connectionPoint OrangeIDE at an aggregator like OpenRouter and a single key reaches hundreds of models. Billing stays between you and them — we take no cut and never see the traffic.
There is no OrangeIDE account, balance, or gateway. There never will be.
optional · your key, your billing// hardware fit
Pick your specs for an honest starting point — the same logic the in-app autotuner uses. Mixture-of-experts models punch well above their VRAM tier, which is why a 12 GB card can run a 35B.
# real GGUF file sizes · speeds measured where a public benchmark exists, estimated otherwise · checked Aug 2026
Qwen3.5-9B
UD-Q4_K_XL · 5.97 GB download
The sweet spot for an 8 GB card: fits entirely in VRAM and is the strongest tool-caller available in single-digit billions.
⚠ 8 GB system RAM is below the 16 GB minimum — expect heavy swapping.
// the ide
Editor, terminal, git, and an agent that uses all three — in one portable install.
An autonomous loop that plans, edits, runs your tests, and verifies its own work — with an autonomy dial from read-only to hands-off and a live plan you can steer.
Search Hugging Face in-app, compare quantizations, download multi-part GGUFs, and let the autotuner pick GPU layers and context for your card.
Index your own PDFs, DOCX, and docs into a local knowledge base the agent can search. Your documentation never leaves the disk it lives on.
An integrated shell that follows your file explorer, plus commit, branch, and PR panels the agent can drive with your permission.
Extend the agent with Model Context Protocol tools — browsers, databases, custom scripts — managed from a built-in MCP panel.
Drop the folder on an external SSD and the whole IDE travels — models, config, sessions, keys. It even heals drive-letter changes between machines.
// what it costs
Free for individual use — no account, no licence key, and nothing held back for a Pro edition. That's sustainable rather than a bait-and-switch, for a fairly boring reason:
Inference runs on your GPU and your electricity. An extra user costs us nothing, so there's no meter to run and no reason to install one.
Nothing to host, nothing to keep online, nothing to bill for. The only running costs are a domain and the bandwidth to hand you the download.
No licence check, no activation server, no phone-home. The copy you download keeps working whatever happens to us, so there is no lock-in to price against later.
// for organisations
Deploying OrangeIDE across a team, or somewhere source code legally cannot leave the building? We can help with larger deployments — licensing, setup, and ongoing support. Reach out and we'll put together a quote.
// in writing
Trust claims are cheap. These are specific enough that you could catch us breaking them.
// download
One download, batteries included: the inference engine, a Python and Node runtime, and the agent's whole toolbelt ship in the box. That's why it's big.
OrangeIDE-Portable-0.3.2.zip · 1.87 GB
Unzip anywhere — including an external SSD — and run OrangeIDE.exe. No installer, no registry entries, no admin rights. Config, models, and sessions all live inside the folder you unzipped.
Download · 1.87 GBSmartScreen will say the publisher is unknown, because the build isn't code-signed yet. Click More info, then Run anyway. This is expected — not a sign anything is wrong with your download.
Signing is on the list. Worth knowing it wouldn't remove that prompt immediately either: Windows builds trust per-file over many clean installs, so a new signed app gets warned about too.
SHA-256
f468d062219451d7ce8f2d8b61387f1eb1f21c2e28a36009cc7927e6a73b4a28
Get-FileHash .\OrangeIDE-Portable-0.3.2.zip
# no installer yet — the portable build is the whole product · macOS & Linux are not supported today
// get in touch
Questions, bug reports, and commercial licensing enquiries all land in the same inbox. Every message is read; please allow a little time for a reply.
// faq
Yes. Download a model while you have a connection, then unplug — chat, agent mode, RAG, terminal, and git all run against the bundled llama.cpp engine on your own hardware. The only things that inherently need a network are cloud models and searching Hugging Face for new models.
There isn't one, and the reason is unglamorous: inference runs on your GPU, so an additional user costs almost nothing to serve. There's no hosted service behind it to fund, which is exactly why individual use can stay free without a countdown attached. Revenue comes from commercial and air-gapped licensing for organisations — see what it costs.
No — not to download, not to install, not for any feature. There is no sign-up anywhere in the product.
On your machine, in OrangeIDE's local config folder — in portable mode, inside the folder you carry. Each key is sent only to the provider it belongs to, directly from your computer. They are never transmitted to us, because there is no "us" in the network path.
Never. There is no OrangeIDE server anywhere — not for local models, not for cloud models via your own key. Nothing to see it with.
Because the build isn't code-signed yet. A signing certificate requires a registered business entity and an annual fee, and it's on the list rather than done. Until then, SmartScreen will flag it as an unknown publisher — verify the published SHA-256 against your download if you want certainty about what you got.
Your copy keeps working, permanently. There's no licence check, no activation server, and no phone-home — nothing that can be switched off remotely. Whatever version you have on disk runs until Windows itself stops supporting it.
Windows 10/11 x64 with 16 GB of RAM minimum. An NVIDIA GPU makes local models substantially faster, but CPU-only works if you're patient. Use the "What can your machine run?" picker above for an honest starting point.
No. Windows x64 only today. The bundled runtime — llama.cpp binaries, Python, Node — is all built for Windows, so ports aren't a recompile away.