Skip to content

Playground

The playground is a built-in conversational client wired to the same /api/v1/* endpoints your apps use. It's not a sandbox — it talks to the same providers, the same MCP servers, and is gated by the same authentication.

Chat (/playground/chat)

  • Pick one or several models and stream simultaneous responses for comparison
  • Tool calls, when triggered by the model, render inline with full request / response inspection
  • Per-conversation settings: model selection, temperature, max tokens, default-params overrides
  • Conversations are persisted; reopen any past one and continue
  • Branching: edit any past message to fork the conversation

Live metrics

Every assistant message carries:

  • Token usage (prompt / completion / total)
  • Time-to-first-token (TTFT)
  • End-to-end latency
  • Provider + model the request actually hit

Stop, retry, fork

  • Stop cancels an in-flight stream and persists what arrived
  • Retry re-sends the last user turn against the same model
  • Editing any prior user message creates a new branch from that point

Working with tool calls

When MCP servers are registered and enabled, every chat request advertises their tools. Tool invocations are rendered as collapsible cards with the full input / output. Large tool transcripts collapse by default — open the ones you care about.

Embedding (/playground/embedding)

  • Side-by-side embedding comparison across multiple models / providers
  • Pairwise cosine similarity matrix for the input list
  • Toggle to view raw vectors or only their distance metrics

Settings persistence

The playground splits state in two places:

  • Per-device UI: send-on-enter, show-timestamps, compact-mode → localStorage
  • Cross-device: default model, default temperature, default tools enabled → server-side user_preferences table

So your provider / model defaults follow you between browsers, but window-level UI tweaks stay local.