> ## Documentation Index
> Fetch the complete documentation index at: https://evalgate.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Quick start

> Start with one local gate. Add hosted evidence after proving a regression blocks.

# Start with one local gate

The local CLI gate is **Experimental**, as recorded in [feature status](https://evalgate.com/docs/platform/feature-status.md).
In your first hour, use **`gate`**. Write a reviewed case, accept a passing
baseline, break the case, and read the blocking result. Start with the
[TypeScript / JavaScript five-minute loop](https://evalgate.com/docs/quickstart-typescript.md).

```bash theme={null}
npm install --save-dev @evalgate/sdk
npx @evalgate/sdk init --local --format json
npx @evalgate/sdk init --local --apply --format json
# Replace the placeholder with your own reviewed behavior check.
npx @evalgate/sdk discover --manifest
# Commit the reviewed scaffold and specs before accepting a baseline.
npx @evalgate/sdk baseline update
npx @evalgate/sdk gate --format json
```

Preview writes nothing; `--apply` authorizes the scaffold. Package installation
requires network access, but deterministic initialization and evaluation need
no account, browser, organization, GitHub remote, EvalGate API key or model
call. Always use the scoped `npx @evalgate/sdk`; the unscoped npm package is
unrelated. In a Python project use the installed `evalgate` executable and
the [Python guide](https://evalgate.com/docs/sdk/python.md), checking its released capability boundary.
Each runtime executes its own specs.

The loop is complete when a non-empty reviewed baseline passes unchanged,
a deliberate behavioral regression exits `2`, and restoring the behavior
passes against that same baseline. `status --json` exit zero only means the
status document was produced. A gate pass supports its configured evidence;
it is not universal safety or deployment permission.

| Local file | Purpose |
| - | - |
| `evalgate.project.json` | Local execution target and baseline lineage |
| `evalgate.config.json` | Command and evidence required by the gate |
| `evals/baseline.json` | Initially pending; accepted only after a reviewed passing run |
| `evalgate.md` | Repository instructions |
| `.github/workflows/evalgate-gate.yml` | The same local gate for pull requests |

`evalgate.product.json` is **not** an init artifact: Experimental
`understand --apply` writes product intent separately. `.evalgate/project.json`
is **not** local-init output either: `link` writes hosted linkage. Remember
`evalgate.project.json` as **local execution**, and `.evalgate/project.json`
as **hosted linkage**.

## After the first gate

Use [CI/CD gates](https://evalgate.com/docs/guides/cicd-integration.md) for the command-mode comparison:
what `gate`, later `ci`, and later hosted `check` prove, their costs, and what
a red build means. [Quality score](https://evalgate.com/docs/concepts/quality-score.md) is the shared
descriptive scoring contract; a blended score does not replace release checks.

Use [migration guidance](https://evalgate.com/docs/guides/migrate-to-evalgate.md) for existing cases.
The [public playground](https://www.evalgate.com/playground) offers a synthetic
review fixture; it makes no production or provider-quality claim.

## Optional hosted setup

The hosted platform is **Beta**. Repository Intelligence, Agent Skills, MCP and
product-understanding availability follow [feature status](https://evalgate.com/docs/platform/feature-status.md).
Hosted collaboration is a separate, browser-approved step:

```bash theme={null}
npx @evalgate/sdk login
npx @evalgate/sdk link --preview --format json
npx @evalgate/sdk link --apply --format json
```

`login` saves a revocable human session outside the repository. `link` adds
non-secret organization/repository/root identity without replacing local
specs, baselines, configuration or user edits. Cloud intake is pinned to an
exact commit; branch movement changes checkout status, not repository identity.
A behind checkout may remain linked and locally ready. Unpushed or diverged
work cannot silently borrow evidence from another commit.

Repository Intelligence supports **GitHub** repositories. GitLab CI can run
the local gate recipe; it does not provide GitLab repository connection or
cloud repository discovery. [Creating a connected repository scan](https://evalgate.com/docs/help/repository-intelligence/scan-repository.md)
requires `eval:write` and an organization member role because it persists evidence;
reading existing evidence requires `eval:read`. Model-backed judges require separate BYOK
credentials and provider spend. Check [production adoption](https://evalgate.com/docs/platform/production-adoption.md)
before expanding a hosted pilot.

## Optional coding-agent instructions

The official [Skills repository](https://github.com/evalgate/skills) is
Experimental. Inspect with `npx skills add evalgate/skills --list`, then
install deliberately with `npx skills add evalgate/skills`. Instructions do
not issue credentials, approve spending or enforce release policy. See
[write evals with your agent](https://evalgate.com/docs/guides/write-evals-with-your-agent.md) after the
deterministic first loop.

## Inspect installed capabilities

`npx @evalgate/sdk version` reports the installed CLI; `npm view @evalgate/sdk version`
observes the current registry version. `npx @evalgate/sdk capabilities --format json`
reports supported commands. Do not treat a source-tree version as published.
Package publication does not establish that a matching hosted API is deployed.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.