OpenFusion

Use

The dashboard ​

The dashboard is OpenFusion's local web UI. It opens automatically at http://localhost:9077 the first time your agent starts the server — or start it anytime:

bash
openfusion-ui        # or: npx openfusion-ui

It binds to 127.0.0.1 only — nothing outside your machine can reach it. Four tabs: Playground · Dashboard · Settings · Errors.

Playground ​

The Playground is a fusion workbench in your browser: write a prompt (with optional context), pick a persona, press Run, and watch the candidates answer and the judge synthesize live. It runs the exact same engine your agent uses, so a Playground run shows up in your history and stats like any other.

The Playground — write a prompt, pick a persona, run the panel

The Playground — a fusion workbench in the browser.

Useful when you're tuning the panel: try different candidate mixes here before committing them as the active setup. The history rail on the left keeps every past run — click one to re-read its full answer and the candidate-by-candidate breakdown, including runs your agent kicked off.

Dashboard ​

The overview tab: activity and cost at a glance, from the same local database every fusion writes to.

OpenFusion dashboard — KPIs, charts, and expandable activity log

The usage dashboard — KPIs, per-model charts, and the activity log.

  • KPIs — fusions run, total spend, token usage.
  • Fusions per day and cost by model and tokens by model charts.
  • Expandable activity log — every fusion opens into its sub-calls: each candidate and judge step, with model, cost, tokens, and latency.

Because each fusion stores one row per sub-call (N candidates + two judge steps), the charts can slice spending by model — you'll see exactly which panel member earns its keep.

Settings ​

Models, judge, keys, execution mode, and personas — all covered in Configuration. Changes apply immediately; the ● Configured badge turns green when the fusion tool is ready.

Errors ​

Fusions that failed or were interrupted land here with the reason — unconfigured setup, failing keys, candidates that timed out. If a client reported a fusion as failed but you suspect it finished, check here first: results often complete and log even when the client's own timeout gave up (more on that).

The error log — failed and partial fusions with their reasons

The error log — failed and partial fusions, with the reason attached.

Privacy ​

The dashboard, the engine, the database, and your encrypted keys all run on your machine. The only outbound traffic is the model calls themselves — from your machine directly to the providers you configured.

Next up