Use the dashboard

The dashboard is a local browser workbench for browsing workflows, configuring agent lineups, previewing dry runs, starting real runs, watching structured per-agent events, gating human review, and handing completed artifacts to follow-up agents. Long multi-agent workflows are easier to inspect as stateful runs than as terminal output alone.

Open it


nax dashboard review


nax dashboard --run <workflow-run-id>


nax dashboard --no-open


nax dashboard --no-open --no-tail

Passing a workflow name opens it on launch; --run <id> jumps straight to a saved run’s details view. The dashboard streams child workflow output to its terminal by default — use --no-tail to keep workflow activity in the browser only, and --no-open to print the URL instead of launching a browser.

When started normally, the dashboard also advertises a private, project-scoped MCP control plane for nax mcp. The advertisement contains the ephemeral loopback address and per-process token with private filesystem permissions. Restarting the dashboard does not require an MCP configuration change because the stdio adapter rediscovers it before every operation. Disable the advertisement with --no-mcp-advertise.

The opened URL carries a per-session ?token=. Mutating endpoints require that token (or an x-nax-token header), so share the exact URL, not just the host.

Surfaces

Surface	What it does
Workflow sidebar	Lists the same workflows as `nax list`, with search and a Run an individual agent action.
Graph canvas	Renders each step as a node and each configured instance as a chip, with inputs, submit mode, and live per-instance status.
Add agent	Opens the two-step lineup wizard on a runnable step (choose how many, then configure).
Remove agents	Clears every instance from a step in one click.
Chip caret	Edits or removes a single provider/model/effort instance.
Dry run	Runs `nax run <flow> --dry --force` through the local API and streams output.
Run	Starts the workflow after an in-browser confirmation that names the Netlify site and credit spend.
Output panel	Shows Dry Run / Run sections with status, command, streamed stdout/stderr, and a raw-event diagnostics view.
Recent runs	Reads `.nax/workflows` with search, a status filter, durable pagination, and resume.
Run details	Shows per-agent results, prompts, artifacts, copy/open actions, review gates, and follow-up controls.
Branch	Suggests current, local, and locally known remote branches while still accepting an unfetched name.
Netlify project	Inspects every local site link and marks the Agent Runner target.

The workflow canvas

Each step is a node showing its number, status badge, submit mode (new agent run, follow-up session, human review), title, and description. Every configured instance is a chip with the provider icon, model label (or Auto), and effort.

During a run, chips are keyed by exact instance id, so two models from the same provider report independent status. Each chip’s control reflects its state:

Chip state	Control
Running	Cancel (✕) — cancels only that runner; others in the step continue.
Running (no cancel)	Spinner.
Completed	Check.
Failed / cancelled	Retry when the step is still active, otherwise an alert icon.

Because instances are tracked individually, a step where some agents succeed and others fail is marked Completed with failures rather than shown as an all-green step. In a saved run (inspect view), click any chip to open that agent’s result.

Configure agent lineups

Each runnable step card carries one chip per instance. Bare-provider flow entries load as Auto chips and stay Auto until you edit them. A step accepts at most four resolved instances, and local execution runs up to four at once.

Add agents

Choose Add agent on a step to open the wizard.

Step 1 — choose how many. The wizard opens on Add new agents with three options:

Choice	Result
One agent	Configure a single model and reasoning effort.
Several agents	Build a lineup across models and efforts.
One of each provider	Adds the flagship config for every provider at once and closes.

Step 2 — configure. After picking One or Several agents you land on Configure one agent / Configure several agents. Pick the Provider, then:

One agent: a Model select (or Auto) and a Reasoning effort select.
Several agents: clearable Models and Reasoning efforts multi-selects. Every selected effort combines with every selected model, and each model creates a separate instance.

Back returns to the choice screen, and Escape closes the wizard at any point. A live summary (“Adds 3 agents.”) and the Add button reflect exactly what will be created; already-selected tuples are skipped.

Quick presets

The wizard’s presets fill the fields in one click. The active preset stays highlighted as long as your selection matches it.

Mode	Preset	Effect
One agent	Best {provider} model available	Strongest model at its highest effort.
One agent	Auto	Let Netlify pick the model and effort.
Several agents	Every effort level	One model at each supported effort.
Several agents	Every {provider} model	One instance per available model.

Edit or remove instances

Edit one instance: open a chip’s caret to change its provider, model, or effort, then Save. Remove agent deletes just that tuple without affecting other instances of the same provider.
Clear the step: Remove agents (next to Add agent) removes every instance at once — handy before rebuilding a lineup from scratch.

Follow-up steps inherit instances

Follow-up step chips are read-only. They inherit the surviving instances from the first input step and continue each instance’s own runner session; in the definition view, inherited chips mirror changes to that source step. An input configured with results: peers gives each continued instance the other instances’ outputs for cross-grading.

Transport notes

Pinned or repeated-provider lineups resolve to the Netlify API under transport Auto. GitHub transport disables model and effort configuration because the pinned GitHub Action supports provider selection only.

Run and watch a workflow

Dry run previews the workflow without spending credits and streams to the Output panel.
Run opens a confirmation naming the workflow, branch, exact Agent Runner site (and selection reason), and the agents — with a spend warning, since it creates remote work in your Netlify account. You can append optional instructions/context before starting.
While a real run is active, a red Cancel appears in the controls, and each running chip can be cancelled or (once failed) retried individually.
The Output panel exposes Copy output and View event diagnostics (the Workflow event diagnostics modal lists parser errors and recent structured events).

Human review gates

When a step reaches awaiting_review, open the run details and use Continue to approve or Cancel flow to stop before later steps run.

Browse and resume runs

The Runs panel reads .nax/workflows. Search with Search runs, filter by All statuses / Running / Completed / Failed / Cancelled, and use Load older to page through durable history. Each run shows its status, a Stalled badge (with a last-event tooltip) when a run has gone quiet, and usage. Resumable runs expose Resume run.

Inspect a run

Open View run details for the timeline. Each entry shows its result and, when a prompt was captured, a Results / Prompt switch with Copy prompt markdown / Copy results markdown. Retryable entries show Retry result. The metadata panel lists workflow, status, usage, transport, branch, target (with SHA and caveats), runner and session ids, and — when local file access is allowed — actions to open the run directory or copy its path. Artifacts expose Copy file path, Open file, and Open in Netlify.

Follow up from a run

In run details, choose Send to next agent. The menu offers Copy file path of results output, Copy results as markdown, and Run a followup. The follow-up composer requires fresh instructions, defaults to the last meaningful artifact, and lets you choose the mode (follow up on a previous Agent Run, or start a fresh runner), the agents and their configuration, and which workflow/step/runner/session/result artifacts to include. Run history keeps the requested model/effort separate from the backend-observed agent_config.

Run a single agent

For work that does not need a workflow, choose Run an individual agent in the sidebar. In Run one agent, select a provider, Configure its model and effort from the same server-owned catalog, enter instructions, and Run agent. Standalone runs need the Netlify API transport; they are persisted and appear in Recent runs with the resolved provider, model, and effort.

Verify the Agent Runner site

When the dashboard starts, the terminal lists every .netlify/state.json link found in the repository and marks the site selected for Agent Runner execution. The dashboard header exposes that target through the quiet Netlify project · site-slug menu, where you can inspect every local link and open the corresponding Netlify Agent Runs page.

In a monorepo, the repository-root link is not automatically the run target. If exactly one discovered netlify.toml contains a package-manager --filter, nax selects that app and its adjacent .netlify/state.json. The startup output and dashboard menu include this selection reason, and the run confirmation names the exact Netlify site before remote work begins.

Develop the dashboard UI


node src/cli/nax.js dashboard --no-open --port 53734
NAX_DASHBOARD_API_URL=http://127.0.0.1:53734 npm run dashboard:dev

Open the Vite URL with the printed token:


http://127.0.0.1:5173/?token=<token>&workflow=do-next

⚠️

Mutating dashboard endpoints require the token embedded in the opened URL or supplied as x-nax-token.