Install Ivy

One binary, no runtime, no package manager. The installer downloads the file for your platform and puts it on your path. Every release ships with a SHA-256 manifest you can check by hand.

macOS

Apple silicon and Intel. The installer picks the right one.

curl -fsSL https://ivy.fightclub.pro/install.sh | sh

It installs to /usr/local/bin/ivy and asks for sudo only if that directory is not writable. Homebrew users can keep it there too; there is no formula yet.

By hand

curl -fsSL https://ivy.fightclub.pro/releases/1.0.0/ivy-darwin-arm64 -o ivy   # or ivy-darwin-amd64
chmod +x ivy && sudo mv ivy /usr/local/bin/ivy
ivy version

If macOS refuses to run it ("cannot be opened because the developer cannot be verified"), the binary is not notarised. Allow it under System Settings, Privacy & Security, or run xattr -d com.apple.quarantine /usr/local/bin/ivy.

Linux

amd64, arm64, arm and 386. Works on any distribution; there are no shared library dependencies.

curl -fsSL https://ivy.fightclub.pro/install.sh | sh

By hand

curl -fsSL https://ivy.fightclub.pro/releases/1.0.0/ivy-linux-amd64 -o ivy   # or -arm64, -arm, -386
chmod +x ivy && sudo mv ivy /usr/local/bin/ivy
ivy version

Without sudo, put it anywhere on your path, for example ~/.local/bin.

Windows

Windows 10 or later, amd64 or arm64. Per-user install, no administrator prompt.

irm https://ivy.fightclub.pro/install.ps1 | iex

It drops ivy.exe in %LOCALAPPDATA%\Programs\ivy and adds that folder to your user PATH. Open a new terminal afterwards so the PATH change is picked up. WSL2 users can follow the Linux steps inside their distribution instead.

By hand

Invoke-WebRequest https://ivy.fightclub.pro/releases/1.0.0/ivy-windows-amd64.exe -OutFile ivy.exe
.\ivy.exe version

Verify the download

Every release publishes checksums.txt next to the binaries. Compare before you trust a file, especially one you are carrying onto a machine that cannot fetch it itself.

curl -fsSL https://ivy.fightclub.pro/releases/1.0.0/checksums.txt
shasum -a 256 ivy-darwin-arm64        # macOS, Linux
Get-FileHash .\ivy-windows-amd64.exe   # Windows

Get a model

Local mode needs an OpenAI-compatible endpoint serving a model that supports tool calling. Two routes.

Ollama, for a laptop or a single card

ollama pull qwen2.5-coder:32b     # or qwen3:8b on a smaller card

Raise Ollama's context before you use it with Ivy. The default is 4096 tokens and the agent's prompt is about 6,500, so the model silently loses part of its tool contract and starts improvising. Either derive a model:

printf 'FROM qwen2.5-coder:32b\nPARAMETER num_ctx 32768\n' > Modelfile
ollama create qwen2.5-coder-32k -f Modelfile

or set it on the server:

OLLAMA_CONTEXT_LENGTH=32768 ollama serve

vLLM, for a GPU box

vllm serve Qwen/Qwen2.5-Coder-32B-Instruct \
  --enable-auto-tool-choice --tool-call-parser hermes \
  --host 0.0.0.0 --port 8000

Tool parsers: hermes for Qwen, llama3_json for Llama 3.x, mistral for Mistral. vLLM serves the full native context, so no context change is needed. NIM containers, llama.cpp server, LM Studio, TGI and SGLang expose the same endpoint shape.

First run

Check the model can drive tools, then work.

ivy doctor --model qwen2.5-coder-32k                      # Ollama on this machine
ivy doctor --endpoint http://gpu.internal:8000/v1 --model Qwen/Qwen2.5-Coder-32B-Instruct

cd your-project
ivy --local --model qwen2.5-coder-32k

The first request after a model loads can take a minute while the weights go into VRAM; Ivy waits up to five minutes for it. Everything after that is normal speed.

Hosted mode

The same binary with no flag runs against hosted models through a Knockout account. Create one from the Ivy Dashboard, then:

ivy auth login
ivy "add rate limiting to the login endpoint"

Air-gapped machines

Download the binary and checksums.txt on a connected machine, verify the hash, copy the binary across, and point it at an endpoint on the LAN. Local mode makes no connection of its own except to that endpoint. The agent's WebFetch and WebSearch tools will fail there and the agent carries on.

ivy --endpoint http://10.0.0.5:8000/v1 --model my-model --ca-cert /etc/ssl/corp-ca.pem

Updating

ivy update

Checks ivy.fightclub.pro/latest-version.txt, downloads the binary for your platform, verifies it against the release manifest and replaces the running executable. Nothing checks for updates on its own; this is the only time Ivy contacts the site.

Uninstall

sudo rm /usr/local/bin/ivy          # macOS, Linux
rm -rf ~/.ivy                        # settings and memory, optional
# Windows: delete %LOCALAPPDATA%\Programs\ivy and remove it from your user PATH

Project-level state lives under .ivy inside each project you used it in: sessions, memory files and edit snapshots. Delete those directories if you want them gone.