BLADE AI

Languages: 中文 | English

Kubernetes chaos-engineering AI agent — describe a fault in plain English (or Chinese), no need to memorize CLI flags.

BLADE AI is the orchestration layer on top of ChaosBlade: the agent understands intent, runs four-layer safety review, drives the injection, verifies the effect, recovers cleanly, and produces a structured report — so a drill goes from "look up the right flag" to "talk to the agent".

Documentation map

Introduction → docs/INTRODUCTION_en.md — positioning, capability matrix, architecture, safety model, tech stack
Usage → docs/USAGE_en.md — install, TUI, CLI, all 19 fault scenarios, server REST+SSE API, config

The rest of this file is the fastest path to a working install (≈5 min). Read the two longer docs above for "why it's designed this way" and the full capability surface.

Install

The release-blade-ai.yml pipeline publishes self-contained binaries (bundled Python runtime + ChaosBlade binary + skill files; extract-and-run) for four platforms whenever a blade-ai-v* tag is pushed: linux-amd64 / linux-arm64 / darwin-amd64 / darwin-arm64. Windows is not yet supported.

One-liner (recommended)

When no version is given, the script queries the GitHub Releases API and resolves the latest blade-ai-v* tag automatically — no need to edit the script for each new release:

# macOS / Linux — install latest (default; auto-resolves the newest release)
curl -fsSL https://chaosblade.io/install-agent.sh | bash

# Pin a specific version (bare semver, no blade-ai-v prefix)
curl -fsSL https://chaosblade.io/install-agent.sh | bash -s -- --version 0.1.0

# Same via env var (works through irm | iex / docker / CI)
BLADE_AI_VERSION=0.1.0 curl -fsSL https://chaosblade.io/install-agent.sh | bash

Windows: install.ps1 is in place but the current release matrix does not ship a Windows binary; the script prints a clear "not yet supported" message and points you at WSL2 / building from source. Once a Windows matrix entry lands, irm | iex will work immediately with the same latest-version auto-resolution.

If the chaosblade.io redirect is not configured yet, fetch the installer directly from a GitHub Release:

# Fetch the install script straight from the GitHub Release
VERSION=0.1.0
curl -fsSL "https://github.com/chaosblade-io/chaosblade/releases/download/blade-ai-v${VERSION}/install.sh" | bash -s -- --version "${VERSION}"

Download a prebuilt archive manually

Each release uploads 4 archives + checksums.txt to the blade-ai-v<version> Release:

Platform	Archive
Linux x86_64	`blade-ai-linux-amd64.tar.gz`
Linux ARM64	`blade-ai-linux-arm64.tar.gz`
macOS Intel	`blade-ai-darwin-amd64.tar.gz`
macOS Apple Silicon	`blade-ai-darwin-arm64.tar.gz`

VERSION=0.1.0
PLATFORM=darwin-arm64    # match your host
URL="https://github.com/chaosblade-io/chaosblade/releases/download/blade-ai-v${VERSION}/blade-ai-${PLATFORM}.tar.gz"
curl -fSLO "${URL}"
tar -xzf "blade-ai-${PLATFORM}.tar.gz"
./blade-ai/blade-ai version
# Add the blade-ai/ dir to PATH, or symlink blade-ai into /usr/local/bin

Uninstall

uninstall.sh / uninstall.ps1 are uploaded next to install.* in every blade-ai-v<version> release and mirror the install invocation style.

# macOS / Linux — one-liner (recommended; mirrors the install one-liner)
#
# Note: when invoked via curl | bash, stdin is not a tty, so the
# interactive y/N prompt is disabled. Uninstall is a destructive
# operation — you must pass --force to confirm explicitly. Combine
# with --keep-config / --version if you don't want everything gone.
curl -fsSL https://chaosblade.io/uninstall-agent.sh | bash -s -- --force

# Run a dry-run first, then decide whether to actually delete
curl -fsSL https://chaosblade.io/uninstall-agent.sh | bash -s -- --dry-run

# Remove binary + PATH but keep ~/.blade-ai/ (config / memory / skills)
curl -fsSL https://chaosblade.io/uninstall-agent.sh | bash -s -- --force --keep-config

# Remove a single version (other versions and symlink kept in multi-version setups)
curl -fsSL https://chaosblade.io/uninstall-agent.sh | bash -s -- --force --version 0.1.0

If the chaosblade.io redirect is not yet configured, pull the script directly from the GitHub Release:

VERSION=0.1.0
curl -fsSL "https://github.com/chaosblade-io/chaosblade/releases/download/blade-ai-v${VERSION}/uninstall.sh" | bash -s -- --force

If you already have the script locally (e.g. after install):

# Direct terminal invocation: interactive y/N — no --force needed
bash ~/.blade-ai/versions/blade-ai-v0.1.0/scripts/uninstall.sh --dry-run
bash ~/.blade-ai/versions/blade-ai-v0.1.0/scripts/uninstall.sh
bash ~/.blade-ai/versions/blade-ai-v0.1.0/scripts/uninstall.sh --keep-config
bash ~/.blade-ai/versions/blade-ai-v0.1.0/scripts/uninstall.sh --version 0.1.0

# Windows (script is in place — works once install.ps1 ships a real binary)
.\uninstall.ps1                          # full uninstall
.\uninstall.ps1 -KeepConfig              # keep config dir
.\uninstall.ps1 -Version 0.1.0     # safety check: only proceed if manifest matches
.\uninstall.ps1 -DryRun                  # show the plan, no deletion

Every shell-rc / registry edit takes a sibling backup (~/.zshrc.blade-ai-uninstall.bak / ~/.blade-ai/path-backup.txt), so a misclick is recoverable.

Build from source

git clone https://github.com/chaosblade-io/chaosblade.git
cd chaosblade/blade-ai
make dev      # install dev deps
make build    # PyInstaller bundle into dist/blade-ai/

Quick start

First launch

blade-ai

The first run walks you through a 5-step onboarding wizard (similar feel to Claude Code's first launch):

LLM API key — supports Alibaba Cloud Bailian and any OpenAI-compatible endpoint; echo is masked
Model selection — recommend qwen-max-latest, qwq-32b, or anything that supports deep reasoning
Cluster config — auto-scans ~/.kube/ and picks defaults for cluster + namespace
Permission mode — confirm / auto / plan; recommend confirm for daily use
Environment doctor — verifies the Blade binary, K8s connectivity, ChaosBlade Operator install, skill files

The wizard writes ~/.blade-ai/config.json and the agent enters the chat loop immediately — no restart needed.

Your first injection

💬 You: inject 80% CPU pressure into the accounting service in cms-demo for 5 minutes

🤖 Agent:
  ⚡ Analyzing request...
  ▸ Safety check ✓ — cms-demo is not blacklisted, no overlapping experiments
  ▸ Plan generated ✓ — pod-cpu fullload, cpu-percent=80, timeout=300
  ▸ Waiting for human confirmation...  → user types yes
  ▸ Executing ✓ — ChaosBlade experiment created (uid: 4d2e...)
  ▸ Verifying effect ✓ — Layer 1: blade_status=Running; Layer 2: kubectl top pod CPU=82%
  ✅ Injection complete! Task ID: task-20260507-a1b2c3

No need to memorize blade create k8s pod-cpu fullload --cpu-percent 80 --namespace cms-demo … — just say what you want done.

Three usage modes

# 1) Conversational TUI (recommended for interactive use)
blade-ai

chaosblade

Most Popular