Graph inputs
No graph inputs. Define named values here (e.g. question, sources) and
bind them into any node with {{slots}} — one input can feed many nodes.
Settings
Context
Window (tokens)
manual
Last peak—
Warn at %
Alert at %
per-model window sizes
Chat
Expanded width px
Channels
Default max rounds
Inter-agent delay ms
Max @-reply hops per turn
Data
state also mirrors to localStorage on every change; a loaded graph file that carries its own settings / agents / channels wins over it.
Agents
No agents. An agent is a named persona with its own endpoint, model and system prompt —
the roster that channels draw on. Agents are not graph nodes. Blank base / model inherit the topbar values.
Channels
Benchmark
Test
Results
Pick a test and hit Run. Artifacts render here — use the header to widen the viewport.
Nothing run yet. The message enters at INPUT and flows along the links.
Nodes whose dependencies are met run in parallel; a chain of links runs in
series. Nodes with no outgoing links are replies.
Each traced node shows its output, and under "payload sent" the exact system + user text that went to the API after slot substitution — byte-for-byte, no guessing about handoffs.
Each traced node shows its output, and under "payload sent" the exact system + user text that went to the API after slot substitution — byte-for-byte, no guessing about handoffs.
Empty channel. Post a message and every enabled member replies —
in parallel with turns off, or round-robin with turns on. Write @name to address one agent only.
Test blob (session / character — message comes from the chat box)
idle
— / 8k