Quickstart

This gets a developer preview running from source on a disposable copy of a repository. It is not an installation guide for a supported product.

Read this first. Bridlewise is pre-1.0 and pre-admission. Do not use it on proprietary, regulated, client-confidential, or production code. Expect breaking changes and data loss. See Security and limits.

Prerequisites

• Node.js >=24.18.0 <25.0.0 and npm.
• Linux for the Landlock path. The server and cockpit can run elsewhere, but OS-level confinement is Linux-only; WSL2 depends on the host kernel and Landlock ABI.
• Git, a C compiler and make if you want to build the guardian yourself.
• A disposable copy of a repository, and a model provider endpoint you can call.
• Enough free disk for node_modules in both packages and local SQLite state.

Check the runtime before installing:

node --version   # must be v24.18.x or later in the 24 line
npm --version

Prefer a prepared workspace?

The dev container installs both lockfiles, builds the guardian’s native dependencies, forwards ports 5173 and 5174, and still runs the cockpit on loopback. It does not include a model credential or the development-floor corpus. See Open it in a Codespace for what it configures and what it deliberately leaves to you. A Codespace consumes your GitHub plan’s included quota; it is not a free hosted deployment.

Install

From the repository root:

git clone https://github.com/LLM-4-People/bridlewise-testing.git
cd bridlewise

npm ci
(cd engine && npm ci)

chmod 600 engine/bridlewise.config.json

Three details matter here:

1. Both packages install separately. The engine has its own package.json and lockfile. A root-only install leaves the engine incomplete.
2. The config must be 0600. Git does not preserve exact non-executable Unix modes, so a fresh checkout can materialize engine/bridlewise.config.json as 0644 or 0664. Self-protection is enabled in the checked-in config, and the runtime fails closed with config_mode_insecure until the mode is fixed.

If you prefer to inspect before running, the config is a strict, typed document. It contains no secrets; secrets live behind handles outside the repository.

Start

node scripts/sd.mjs start

The lifecycle script starts the API server and the Vite UI, waits for health, and reports what it started. Open:

http://127.0.0.1:5173

The API listens on 127.0.0.1:5174 by default. The UI is Vite’s; the API server alone does not serve the app.

Useful commands:

node scripts/sd.mjs status    # PIDs + listeners on 5173 and 5174
node scripts/sd.mjs restart   # stop what this lifecycle started, start fresh
node scripts/sd.mjs stop      # stop only what this lifecycle started

sd.mjs deliberately does not kill a process it did not start. If another process owns a port, it names that process and stops. Do not reach for pkill or kill -9; you will lose the record of what you started.

Ports

If you change the UI origin, set HARNESS_WEB_ALLOWED_ORIGIN to match before starting. If you change the API port, sd.mjs reads HARNESS_WEB_PORT for its health check.

First run

1. Create the first account

The first account owns the harness and is the administrator. Open registration is allowed only while no users exist; after that, registration is closed unless an administrator creates the account. There is no hosted account recovery service.

Choose a username and a password you are willing to lose with the disposable environment. The session cookie is HttpOnly and SameSite=Lax; the database stores a hash of the session token, not the token itself.

2. Add a model provider

Open Settings → Model providers and choose the add action. You need:

• a name for the provider (what the picker shows);
• an API address — an HTTP or HTTPS base URL such as https://your-provider.example/v1, with no embedded credentials, query string, or fragment;
• an API key;
• at least one model id the endpoint serves.

You can add optional per-model limits (context window, max reply tokens, temperature) and extra non-secret headers. The key is stored in the encrypted provider record and is never echoed back by the read path. If a provider is saved without a key, it is listed as incomplete and a run that selects it refuses by name. There is no fallback provider.

Set a default model for new conversations, or choose a model per conversation. Providers must speak the OpenAI Responses or OpenAI Chat Completions protocols. Anthropic-native and Gemini-native APIs are not verified adapters here.

3. Add a project

Choose a disposable copy of a repository. The harness reads and writes inside the selected project and rejects paths outside it. Deleting a project removes it from the harness database only; it never deletes files from disk. Keep your own backup and a clean git state.

4. Send a run

Send a prompt. The cockpit opens a run and streams its beats. Expect to see:

• the user turn and the model turn;
• tool calls with their arguments and results;
• file effects with diffs;
• check deltas when a check changes state;
• findings authored by the harness;
• typed refusals where a capability is blocked.

A run executes once. If you open the same run in another view, it attaches to the recorded stream rather than starting a second execution. Cancelling or steering is an explicit action on the run.

Provider-free demo mode

To exercise the interface without live model calls:

HARNESS_WEB_LLM_FIXTURES=1 node scripts/sd.mjs start

Fixture mode replaces the model stream with deterministic events. It is not a free pass around setup: provider and model resolution still run first, so you must have a provider and model selected. The configured endpoint is not called in fixture mode, and fixture turns are marked as fixtures in the record. Do not use fixture output to evaluate model quality.

To return to live calls, stop the servers and start them without the variable:

node scripts/sd.mjs stop
node scripts/sd.mjs start

What to expect

• A boot column, not a splash screen. Registration reports real probes with measured timings. On a warm signed-in boot it compresses to a few lines.
• Guided refusal instead of an empty chat. With no provider, a run refuses with provider_not_configured and names the remedy.
• Local state. The web database defaults to .bridlewise/web.db; the timeline defaults to .bridlewise/timeline.db. Runtime state is ignored by git.
• Keys outside the checkout. Database key material lives in the key home: HARNESS_KEY_HOME if set, otherwise $XDG_CONFIG_HOME/bridlewise/keys, otherwise ~/.config/bridlewise/keys. Back it up. A database copy without its key is unreadable.
• No hosted service. Nothing is uploaded to a Bridlewise account. Prompts and context do go to the provider you configure, subject to the redaction policy.
• Rough edges. Layout, providers, migrations, and error copy are still changing. If something looks wrong, capture the run id and the typed error before restarting.

Troubleshooting

If a failure is reproducible, include the exact command, the typed error code, the run id, and whether the provider was live or fixture mode. Redact keys, tokens, and private code before posting.

Verify the install

npm run typecheck
(cd engine && npm run typecheck)

The product tree exposes type checks, not the full prevention suite. The complete test inventory, benchmarks, and verification tooling live on main. The engine’s preflight is an integrity receipt:

(cd engine && node src/cli.ts preflight)

It reports harnessReady: false while the governance gate is open. That is the current, honest state of the preview — not a bug in your install.

Next

• Architecture — how the pieces fit and where the trust boundary is.
• Security and limits — what is protected, what is not, and how to report a vulnerability.
• Design — the cockpit’s surfaces, tokens, and accessibility rules.
• FAQ — readiness, providers, licensing, and the main and product branches.