Skip to main content

Graph Copilot agent workbench

Graph Copilot is an agent workbench for CodefyUI. It can build and edit a graph, run bounded experiments on isolated candidates, rank numeric results, optionally promote a parameter-only winner, and turn observations into testable research follow-ups.

Installing the plugin adds a floating assistant button at the bottom-right of the editor. The workbench presents the active model and graph context, chat/tool progress, Experiment Lab, history, and settings. Edit is the default; Ask is executor-enforced read-only and cannot start execution. Each applied GraphOp batch is one host undo step (Ctrl+Z), not a whole agent turn. Stop does not undo earlier changes or external node effects.

What “agent” means today

Planning and study coordination run in your browser. CodefyUI's existing backend validates and executes submitted graphs; this plugin installs no durable agent server or background job queue. Keep the page open during a study. A server-owned host run may survive reload and be followed again, but reattachment does not resume the browser agent or an entire study. See Architecture and CodefyUI compatibility.

Features​

  • Build graphs by conversation — use the installed-node catalog, schema-redacted live graph, exact schema lookups, and batched operations (add_node, connect, set_params, remove_node, auto_layout, and more). Indexed receipts support bounded correction; automatic validation/repair follows actual current-turn edits, not unrelated pre-existing faults.
  • Run the live graph — structural validation and separate approval precede host execution. Track node/training progress, loss, compact outputs, supported media, and exact recorded metrics. Structure passing alone is not execution proof.
  • Follow host runs — supported hosts expose editor-started and plugin runs through list_runs/get_run, reload reattachment, and optional completion notifications. Dismissing/hiding an existing-run follower stops observation only; explicit Stop run requests cancellation.
  • Run isolated experiments — validate and execute in-memory graph clones through CodefyUI's WebSocket engine, then measure/rank them without temporary canvas graphs. Isolation does not sandbox file/network side effects.
  • Search bounded parameter spaces — complete grids and versioned seeded-random plans turn explicit safe parameter domains into deterministic, reviewable candidates under the same approval and run budget.
  • Promote safely — apply a requested parameter-only winner only if graph-revision and baseline-fingerprint checks pass. Structural winners remain review proposals.
  • Inspect research evidence — per-run scalars, descriptive intervals, baseline effect sizes, failures, CSV, and evidence-labeled Markdown preserve observations without presenting a pilot as proof.
  • Carry portable studies — import checks canonical semantic content and raw/derived evidence, shows a preview, and requires an explicit click; it never executes or applies data. Derived candidate graph JSON reapplies redacted GraphOps without changing the canvas, but is not exact execution replay.
  • Attach files — add images, PDFs, text, or code. Text/PDFs are extracted into the prompt; images use multimodal content with vision-capable models. See Attachments.
  • Keep conversation context — browser-local history holds up to 50 conversations and complete tool-call/result groups. Settings/History/Lab navigation or close stops the current turn while preserving its mounted draft/attachments; drafts do not survive reload. See Conversation history.
  • Choose a provider — OpenAI API, OpenAI Codex (ChatGPT account), OpenRouter, Claude API, or a custom OpenAI-compatible endpoint, with capability-negotiated catalogs/reasoning controls. See Provider setup.
  • Follow staged progress and recover safely — numbered rounds, Markdown, live tool stages/durations, and a status bar distinguish agent, structure, and execution evidence. Stop preserves partial text/receipts and blocks new dispatch without rollback. Retry requires no canvas mutation and no approved execution-runner dispatch, keeping the original request once; after attempted or unknown effects use Continue from current graph, not replay. See Graph editing.

Next steps​