northcua — Claude Code skill: what it does, how to install
northcua is a Claude Code skill for running NorthCUA Harness across Desktop/Web tasks, local browser batches, and evidence checks. It guides setup and task verification.
Apache-2.0 ★ 128 updated 2026-11-09
npx skills add beipo-ai/NorthCUA --skill northcua -a claude-codeWhat northcua does
northcua operates NorthCUA Harness from a clean checkout. It helps select a Web or native Desktop target, configure the model, run a bounded task, and inspect the outcome, saved trace, and evaluator verdict separately. Local batches use the real controller, browser workers, and the existing scheduler.
The skill covers config init, config validate, Python and serving modes, remote-device execution, runtime skill loading, and troubleshooting. Source or wheel installation needs uv 0.13.6+; a real model needs a key and, when applicable, an endpoint. A scripted demo can check the installation without credentials. Desktop work requires an exclusive physical display, so concurrent Desktop tasks are not allowed. The local provider shares the host OS and network rather than running in a VM.
What it helps with
- set up NorthCUA Harness from a clean checkout
- run a Desktop or Web task with a selected model
- process several local tasks through browser workers
- separate actor completion from outcome verification
- diagnose a model, browser, or stuck-task failure
Example prompts for the agent
- Install and configure NorthCUA Harness from a clean checkout, then run one bounded Web task.
- Run several local tasks through
local-process, real browser workers, and the batch commands with concurrency 1–2. - After a Desktop task, inspect its saved trace, actual outcome, and evaluator verdict separately.
How to install northcua
1Skills CLI (Vercel)
npx skills add beipo-ai/NorthCUA --skill northcua -a claude-code2Manually: put the files in place
~/.claude/skills/northcua/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/northcua/
1Skills CLI (Vercel)
npx skills add beipo-ai/NorthCUA --skill northcua -a codex2Manually: put the files in place
.agents/skills/northcua/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add beipo-ai/NorthCUA --skill northcua -a cursor2Manually: put the files in place
.cursor/skills/northcua/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install beipo-ai/NorthCUA skills/northcua2Manually: put the files in place
.github/skills/northcua/SKILL.md1Skills CLI (Vercel)
npx skills add beipo-ai/NorthCUA --skill northcua -a gemini-cli2Platform CLI
gemini skills install https://github.com/beipo-ai/NorthCUA --path skills/northcuaManually: put the files in place
.devin/skills/northcua/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about northcua
How do I install northcua in Claude Code?
Run npx skills add beipo-ai/NorthCUA --skill northcua -a claude-code, then use ~/.claude/skills/northcua/SKILL.md.
What does northcua do?
northcua operates NorthCUA Harness: it guides project setup and configuration, runs Desktop/Web tasks and local browser batches, and checks evidence against the evaluator result.
What does northcua need to run?
Source or wheel installation needs uv 0.13.6+. A real model needs a model, key, and an endpoint when applicable; the credential-free scripted demo runs without a key. Desktop tasks require a physical display.
Source text
SKILL.md · Apache-2.0
--- name: northcua description: Install, configure and operate the NorthCUA Harness, including Desktop/Web tasks, local browser batches and evidence checks. Use for running this project, not its training or data-generation research. --- # Operate NorthCUA from a clean checkout The project is **NorthCUA**, the shipped component is **NorthCUA Harness**, and the package/CLI is `northcua`. This skill is distributed with the repository. Resolve the real location of this file (including symlinks); the repository root is two directories above its folder. Run commands there, or use an installed release and the corresponding version of these guides. Do not copy this file alone: its linked guides are part of the skill's operating contract. For a new installation, use the `main` branch of [NorthCUA](https://github.com/beipo-ai/NorthCUA), following the first-run guide's clone/install commands. Check [resource availability](../../docs/resources.md) for the current report and model links; do not assume a referenced model checkpoint has been published or that this project is already available on PyPI. ## Choose the user's workflow - **Nothing installed / first task:** follow [from-zero setup](../../docs/first-run.md). It includes Bash/zsh and PowerShell setup, model templates, native permissions, a credential-free browser/native check and success criteria. - **Model or parameter configuration:** [configuration](../../docs/configuration.md). `config init --model ... --base-url ...` populates both agents. The user normally supplies only model, key and optional endpoint via [the env template](../../.env.example). Read [models](../../docs/models.md) for endpoint requirements and weight availability. Use the matching shell instructions: PowerShell does not source a Bash `.env` file. Keep credentials as environment references; do not ask for keys in chat. Before a Desktop run, match the model's coordinate convention: for 0–1000 relative output, set `desktop.coordinate_policy.input_space: normalized_1000` in the app YAML used by that run; `clamp_to_bounds: true` keeps edge positions inside the capture. The default is `image_pixels`. Read [coordinate configuration](../../docs/configuration.md#desktop-coordinates) for the mapping. Web refs and viewport CSS-pixel actions follow their own contract; this Desktop setting does not change them. - **Several local tasks / recovery:** follow [local batch](../../docs/local-batch.md) end to end. It starts a real controller, real browser workers and the existing scheduler without cloud services. Then consult [batch operations](../../docs/batch.md). - **Python, serving, remote device:** [execution modes](../../docs/running.md). Local application WebSocket and controller submission are different protocols. - **Custom runtime skill:** [skill loading](../../docs/skills.md). - **Failure, missing browser, model error or stuck task:** [troubleshooting](../../docs/troubleshooting.md). - **Result inspection / evaluation:** [evaluation](../../docs/evaluation.md). Read the chosen tutorial before executing; it contains dependencies and expected outputs. Do not infer command flags from another CUA project or old experiment scripts. Use `northcua <group> --help` for optional controls and `config schema` for current app/batch fields. `config validate` does not call the model. ## Execution contract 1. Determine the requested target (Web or native Desktop), execution mode, model endpoint and success criterion. Use the local scripted demo when checking the installation without credentials; label it as such. Use the user's real model for their actual task. Do not substitute demo success for task success. 2. Use uv 0.13.6+ for source or wheel installation; older uv releases can omit Darwin PyObjC dependencies. First read the [known upstream dependency risk](../../docs/security/dependencies.md): the retained NexAU constraint requires affected Transformers 4.x; do not load untrusted tokenizer/model configuration or treat the policy exception as a fix. Install from this repository's lock or matching release wheels. Activate the environment in every terminal. `uv run --no-sync` preserves installed extras. Create configuration in a new writable directory; preserve existing outputs. 3. Run one bounded task first. Desktop GUI tools return screenshots directly; there is no alternative observation mode. A physical display is exclusive: do not run concurrent Desktop tasks or use another agent's GUI session. 4. Inspect terminal status, the actual outcome, saved trace and evaluator verdict separately. A passing configuration check or actor completion is not outcome verification. Retain failed evidence and report unverified modes honestly. 5. For local batching use `local-process`, real browser workers and the existing batch commands. The `fake` provider is only a scheduling simulation. Start with concurrency 1–2; keep task IDs/output paths stable for resume. Drain and wait for the current runner to exit before resuming. Do not hand-edit state DBs. 6. Stop the processes you started and confirm cleanup. Preserve results for the user. The local provider shares the host OS and network; it is not a VM. The self-contained path needs no NorthJetty, cloud key, private dataset or historical deployment checkout. E2B deployments and prepared Desktop HTTP endpoints are optional environments managed separately. Use the [recommended Desktop settings](../../examples/desktop-settings/README.md) through the ordinary Desktop entry points when those model/prompt recipes are useful. Report the exact commands/mode, source revision, model type (scripted or real), actual outcomes, evidence paths, recovery/cleanup result and any remaining blocker. For an independent acceptance run, use a fresh output directory and follow these instructions without relying on author-provided success claims.
© beipo-ai · Apache-2.0 · beipo-ai/NorthCUA