Install and run

How to install Avalanche AI from the public one-liner, run it locally, and take the first steps in the web UI. Written for a fresh machine. If a step fails, the exact error text is quoted so you can match it.

Requirements

Requirement	Minimum	Why
Node.js	24.9	the runtime is built and run on Node 24; the build refuses to start below this
pnpm	11.x	workspace tooling (npx pnpm@11 works if pnpm is not installed)
git	2.26	build metadata and the local install preparation
OS	Linux, macOS	Windows is untested for the source install
Disk	~3 GB	dependencies plus build output
Memory	8 GB free	the client compiler is the peak; below ~6 GB free the build can be OOM-killed

Install

curl -fsSL https://tool.aimb-x.xyz/install.sh | bash

What the script does, in order:

1. Checks Node, pnpm (falls back to npx pnpm@11), git, tar.
2. Downloads avalanche-ai-dsh-0.2.2-rc.1.tar.gz from https://tool.aimb-x.xyz/tarballs/ and verifies its SHA-256.
3. Extracts it to ~/.avalanche-ai (override with AVALANCHE_DIR).
4. pnpm install --frozen-lockfile, then pnpm run build.
5. Smoke-checks the CLI and prints how to start.

The release tarball is served from the project's own domain, so the source repository can stay private. Pinning AVALANCHE_VERSION selects a different published release.

If it fails

Message	Cause	Fix
Node.js vX is too old; 24.9 or newer is required.	old Node	install Node 24 (nvm recommended) and rerun
download too small	the URL returned an error page	check curl -I https://tool.aimb-x.xyz/tarballs/avalanche-ai-dsh-0.2.2-rc.1.tar.gz
checksum mismatch	truncated or modified download	rerun; if it persists, do not install and report it
build stops with no error line, process killed	OOM during the client compiler	free memory (close other builds) and rerun; 8 GB free is safer
could not prepare git metadata (git rev-parse HEAD)	git missing	install git and rerun

Run

cd ~/.avalanche-ai
pnpm dsh web          # or: npx pnpm@11 dsh web

The web UI starts at http://127.0.0.1:3080 and opens in your browser. It only binds to localhost. Use --no-open to skip the browser, and --host 0.0.0.0 only inside a network you control.

Other entry points:

pnpm dsh --help       # CLI reference
pnpm dsh --version    # installed version

First steps in the UI

1. Pick a domain. The runtime hosts seven intelligence domains on one shared core: developer, security, devops, data, protocol, testing, and defi. Routing is deterministic by keyword; choosing a domain sets the tool set and the policy profile for the session.
2. Read the tool catalog. Only five read tools are wired in phase 0–1, all read-only against Fuji/Mock networks. Nothing signs a transaction and nothing writes to a mainnet.
3. Run a read task. Ask a question the read tools can answer (for example a balance or a contract read on Fuji). The nine-stage lifecycle runs to completion and produces an evidence receipt.
4. Inspect the receipt. Every tool result is hash-chained into the evidence ledger. The receipt names the domain, the tools called, their inputs and outputs, and the policy decision that allowed each call. Treat it as the record of what happened, not a summary.

What phase 0–1 does not do

Stated plainly, because the site and the receipt format are easy to over-read:

• No mainnet writes. Read-only tools only, Fuji/Mock networks.
• No signing, no key material. No wallet is connected and no private key is ever loaded.
• Default-deny policy. Any tool not on the read allowlist is refused; approvals are fingerprint-bound and single-use.
• The policy simulator on the site is illustrative. The shipped policy engine is the authority.

Updating

Rerun the installer. It moves the previous install to ~/.avalanche-ai.bak-<timestamp> and installs the new release beside it, so a bad release can be rolled back by moving the backup back.

Uninstall

rm -rf ~/.avalanche-ai ~/.avalanche-ai.bak-*

License and attribution

MIT. This is an independent fork of DeepSeek Harness, not affiliated with Ava Labs or the Avalanche Foundation. Upstream attribution and the license are preserved in the repository.