runanywhereai/wally/runanywhere

RunAnywhere

wally is one CLI for two things: models running on this machine, and models served from RunAnywhere Cloud. The same commands cover both — if a model is on the machine it is served locally, otherwise the request goes to the console the user is signed in to and is metered against their balance.

Install wally if it is missing

If wally is not on PATH, install it first — one line, no configuration and no key to copy:

curl -fsSL https://raw.githubusercontent.com/RunanywhereAI/wally/main/install.sh | sh

macOS (Apple Silicon) and Linux x86-64. On Windows: irm https://raw.githubusercontent.com/RunanywhereAI/wally/main/install.ps1 | iex. The installer only puts wally on PATH and writes this skill; it signs nothing in on its own.

First check what is already true

wally account whoami     # signed in? which console?
wally backends   # which engines this build linked

wally account whoami failing with "not signed in" is the only thing that needs fixing before anything else works against the cloud. On-device models need no account.

Signing in

wally account login

Opens the console in a browser. The person signs in with Google or GitHub, approves the terminal, and the CLI stores a key in ~/.config/wally. There is no password and no organization step. If a browser cannot open, wally account login --no-browser prints the URL to visit.

Coding harnesses

A harness is an existing coding tool that wally wires to a model. Local harnesses use the certified Qwen3 4B Instruct 2507 artifact.

wally opencode -m qwen3-4b-instruct-2507

If opencode is not installed, wally says so and prints the install command (npm i -g opencode-ai) rather than failing. Install it, then run the same line again.

Offer to explain a harness before running it. Someone who has just signed up does not yet know what opencode is, and "would you like me to explain how the opencode harness works?" is a better second message than a launched TUI.

Shannon and other tools

wally shannon -- -u "$APP_URL" -r "$REPO_PATH" runs Shannon, an open-source AI pentester, on a model from the person's account. Only suggest it for a system they own or are authorised to test. It needs Docker running, plus Node 18+ unless Shannon 3.4.0 is installed on PATH, and wally sets up the model config and key itself. For a scan longer than a sign-in lasts, the person exports a console Cloud key as WALLY_CLOUD_KEY. Setup details, and how to point any other OpenAI-compatible tool at the endpoint by hand, are in docs/EDITORS.md.

Security tools

wally deepsec and wally vvah run hosted security scanners on a model from the person's account. Only suggest them for a repo or system they own or are authorised to test. DeepSec (2.3.10) needs Node 18+ for npx unless a deepsec on PATH reports exactly 2.3.10; VVAH (1.5.1) needs Python 3.11–3.14. Wally sets up routing per tool (command-line flags for DeepSec, a config file in the Wally profile for VVAH); keys go in the child environment only, never on the command line, and competing variables are removed for that child only. For a run longer than a sign-in lasts, they export a console Cloud key as WALLY_CLOUD_KEY. Details are in docs/EDITORS.md.

wally deepsec -- init . "$REPO_PATH"
wally vvah -- scan --repo "$REPO_PATH"

Running a model directly

wally models pull qwen3-4b-instruct-2507  # download it
wally models list                         # what is downloaded
wally run qwen3-4b-instruct-2507          # talk to it

Models land in ~/.local/share/runanywhere. Nothing is downloaded until asked.

Spend

wally account usage               # credit left, then input/output/cache tokens and spend
wally account usage --json
wally account usage --requests      # one page of settled requests, last day; --follow walks up to 100 pages

Read-only, and scoped to the signed-in account.

Decisions with Eve

eve, the default for wally decisions, scores named labels without generating prose:

wally decisions --input "Checkout is blank" --ask "Is this a bug?"
wally decide --input-file ticket.txt --choice "Owner=frontend,payments,account"

--image FILE (PNG, JPEG or WebP, up to 8, hosted only; HEIC is converted to JPEG on macOS) asks the questions about images too; a model that does not take them refuses with 400.

Use --json when another tool needs the raw response. Do not send eve (or its retired names pplx-decider-v1 and qwev) to wally run or a coding harness; those surfaces refuse it and point back here.

When something is wrong

• "not signed in" — wally account login.
• "not available on this account" — that model may need approval. Check

the model id first; if it is right and the model is gated, request access at https://runwally.com/security/request. Do not log in again.

• DeepSec or VVAH prerequisites missing — install Node 18+ (DeepSec) or Python 3.11–3.14 (VVAH); see docs/EDITORS.md.
• "that key is not valid" — the key was revoked or expired; wally account login again.
• opencode not installed — npm i -g opencode-ai.
• a model is slow or unavailable — wally backends shows which engines this

build actually linked; a model needing an engine that is not there will not run.

What not to do

Do not print, log or echo the contents of ~/.config/wally/credentials.json. It holds a key with the person's credit behind it.