CodeNavi's code embedder needs your NVIDIA card to index. On Windows you install Docker Desktop and run two short PowerShell lines — that's it. Docker Desktop's WSL2 backend passes your GPU through for you; you never open a Linux shell. No Linux to live in. No code leaves the machine.
WSL2 is the lightweight Linux layer Windows uses to run containers. It's the only path to GPU acceleration in Docker on Windows — the Hyper-V backend can't pass through your card, so it can't index. The good news: Docker Desktop sets up and manages WSL2 for you. You never run a WSL command or open a Linux prompt.
Install Docker Desktop, paste two PowerShell lines — then never touch a Linux prompt again. Day-to-day, CodeNavi is a URL your Windows IDE talks to.
On Windows 11 its installer sets up the WSL2 backend automatically. If it ever isn't, the CodeNavi installer prints the one setting to flip and stops — it never silently runs slow.
CodeNavi lives in Docker (WSL2 backend) and stays up. You start it once; it auto-restarts when you sign in.
Claude Code, Antigravity, Windsurf, VS Code — they call CodeNavi over http://localhost:8765. No Linux in the loop, no port proxy to configure.
The embedder, the index, and your code all stay on your machine. This is fully local — no cloud, no upload.
Docker Desktop's WSL2 backend ships ready to go. (Windows 10 21H2+ works too, the same way.)
24 GiB (RTX 4090 / 3090-class) indexes at full speed; 11–16 GiB still runs Tier A, just slower. Pre-RTX cards (GTX 10-series) aren't supported — the installer checks and tells you.
It carries the WSL2 CUDA runtime (driver ≥ 570.65 for this build). If your card runs games, you're basically set — just keep it current.
For Docker Desktop, the CodeNavi image, its update cache, and your index. No admin rights needed to run the installer.
Download Docker Desktop, install it, launch it, and wait for the whale icon to read "Engine running". On Windows 11 it sets up the WSL2 backend for you — that's the whole "WSL" part, and you don't touch it.
Free for individuals and orgs under 250 employees / $10M revenue; larger orgs can use Podman Desktop instead. This is the only thing you install by hand.
Open PowerShell (no admin needed) and paste the two lines from your license email — the first saves a script you can read, the second runs it:
curl.exe -fsSL "https://codenavi.com/trial/install/<your-token>?os=windows" -o codenavi-install.ps1 powershell -ExecutionPolicy Bypass -File .\codenavi-install.ps1
It downloads CodeNavi, verifies every file against a checksum published by codenavi.com, loads the image, and installs into %USERPROFILE%\codenavi. It asks one question — the absolute path of the repo to index (Explorer's "Copy as path" pastes it straight in).
It doesn't just look for a card — it runs a real CUDA operation inside the container. If the GPU can be used you'll see GPU: enabled and Tier A auto-tunes to your VRAM. If it can't (old driver, wrong Docker backend, or a pre-RTX card) it stops with the exact fix — it never silently falls back to CPU, which would be ~100× slower.
The fixes it prints, in order: update the NVIDIA Windows driver; run wsl --update; make sure Docker Desktop → Settings → General → "Use the WSL 2 based engine" is on. More in snags & fixes; the full reference is docs → Windows.
When it finishes, open http://localhost:8765/connect and paste the token it printed. Click "Build it first →" to index, then pick your editor — the wizard hands you a ready-made config with the token filled in:
# Claude Code — .mcp.json in your project (Antigravity / Windsurf / VS Code take the same URL)
{
"mcpServers": {
"codenavi": {
"type": "streamable-http",
"url": "http://localhost:8765/mcp",
"headers": { "Authorization": "Bearer <YOUR_TOKEN>" }
}
}
}
The Connect wizard fills the token in for you — that's the easy path. Your install always issues an auth token, so the Authorization header above is required. Docker Desktop publishes the port to Windows, so localhost:8765 works from your IDE with no port proxy and no netsh — nothing extra to configure. Reaching it from another machine? Put Tailscale in front — see the README in your install folder.
CodeNavi auto-tunes to your GPU at startup — you don't configure anything. Here's how the common tiers land.
RTX 4090 / 3090, A5000-class. Encode batch 16 + reranker on GPU — index a 30k-file repo in minutes. This is the sweet spot.
Autotune drops to batch 2–4 and moves the reranker to CPU so everything fits. Slower indexing, identical search quality.
The image's CUDA build (cu128) covers RTX 20-, 30-, 40- & 50-series — sm_75 through sm_120, Turing to Blackwell. Pre-RTX cards (GTX 10-series, Pascal) aren't supported; the installer detects that and stops rather than fail mid-index.
| Symptom | Fix |
|---|---|
| "GPU present but unusable by the container" | The installer's CUDA test failed. In order: update the NVIDIA Windows driver (≥ 570.65 — the CUDA 12.8 runtime this build needs), run wsl --update, and confirm Docker Desktop → Settings → General → "Use the WSL 2 based engine" is on. Never install a Linux GPU driver inside WSL — it breaks passthrough. |
| "compute capability < 7.5" | A pre-RTX card (GTX 10-series / Pascal). This build needs RTX 20-series or newer. To index on CPU anyway (much slower, needs lots of RAM): re-run with $env:CODENAVI_GPU_CPU_OK='1' set first. |
| "virtualization is DISABLED" | Docker Desktop can't start its VM. Enable Intel VT-x / AMD-V in BIOS/UEFI, then in an admin PowerShell run bcdedit /set hypervisorlaunchtype auto and reboot. |
| Installer warns Docker's VM has too little RAM | Docker Desktop defaults to ~half your RAM. Create %USERPROFILE%\.wslconfig with a [wsl2] line then memory=12GB, and restart Docker Desktop. The installer prints this exact recipe when it applies. |
IDE can't reach localhost:8765 | Make sure the container is up (docker compose ps in %USERPROFILE%\codenavi) and Docker Desktop is running. No port proxy is needed on the Docker Desktop path — the port is published to Windows directly. |
Once /health is green and the first index has built, point your agent at a real question and let it search instead of grep. Everything below runs entirely on your GPU, on your machine.
# in your Windows IDE, ask the agent:
"Use semantic_search to find where requests are authenticated."