Install GEAK

Install GEAK#

2026-07-17

2 min read time

Applies to Linux

GEAK 4.0.0 is not a Python package. It is a set of Workflows (e2e_workflow.js / kernel_workflow.js) that run inside Claude Code. “Installing” means: get the repo, get a recent Claude Code, and have a working ROCm environment (plus a serving backend for E2E). For a first run, see Run a workflow.

Prerequisites#

GEAK 4.0.0 requires the following software and hardware.

Requirement

Detail

AMD Instinct™ MI GPU

CDNA, gfx942 (MI300X) / gfx950 (MI350X/MI355X). Auto-detected.

ROCm 6+

rocminfo / rocm-smi must work.

A profiler

One of rocprof-compute, rocprofv3, rocprof (also omniperf or metrix). Auto-detected.

Python 3.8+

Tested on 3.12.

Claude Code ≥ 2.1.177

Required for the dynamic Workflow feature. Check claude --version.

Anthropic API key

Set as ANTHROPIC_API_KEY. Get one at console.anthropic.com.

Serving backend (E2E)

A running-capable sglang or vllm, plus model weights on disk.

Set up GEAK#

Clone the repository and run the setup script.

Installing GEAK installs the geak Python package + deps, clones the GEAK repo, and installs the Claude Code CLI. By default the repo lands in ./GEAK under the directory you run the command from (override with GEAK_HOME). Pick either method — both end up the same:

A. One-liner — run it in the directory where you want GEAK to live:

pip install "git+https://github.com/AMD-AGI/GEAK"

B. Clone first — if you’d rather have the checkout up front (e.g. to work on a branch):

git clone https://github.com/AMD-AGI/GEAK.git
cd GEAK
pip install .

It leaves PATH and API access configuration to you. Follow its printed next-steps to add ~/.local/bin to PATH, then set your Anthropic API key:

export ANTHROPIC_API_KEY=<your-key>

Get a key from console.anthropic.com if you don’t have one. Add the export to your shell profile (~/.bashrc or ~/.profile) to avoid setting it each session.

Launch GEAK:

IS_SANDBOX=1 claude --dangerously-skip-permissions

Nothing is compiled at clone time — the workflow .js files and their roles/, knowledge/, scripts/ are used directly. Sandbox mode auto-approves the permissions the workflows need.

Verify the environment#

Run these checks before starting a workflow. A misconfigured environment fails deep into a multi-hour run.

# Claude Code version (must be ≥ 2.1.177)
claude --version

# GPU is visible to ROCm
rocminfo | grep -E "Name:|gfx"

# At least one profiler is on PATH
command -v rocprof-compute || command -v rocprofv3 || command -v rocprof

Expected output:

  • claude --version prints 2.1.177 or higher.

  • rocminfo lists your GPU name and a gfx942 or gfx950 target.

  • At least one profiler command resolves without error.

If rocminfo fails, your ROCm stack is not installed or not on PATH. If no profiler resolves, install rocprof-compute (preferred) or rocprofv3.

Contents