Legend

Legend (/legend) chains agents and tools in a DAG (directed acyclic graph) for multi-step automation. This guide covers building, tools and approval, execution and observability.

flowchart LR
    A[Trigger] --> B[Agent]
    B --> C{Condition}
    C -->|true| D[HTTP / Lore]
    C -->|false| E[Agent]
    D --> F[End]
    E --> F

Overview

Build a DAG

1. Use New, or Templates (Starter workflow, Support triage, Content pipeline).
2. Click or drag palette nodes onto the canvas.
3. Drag an output port to another node's input port.
4. Select a node and configure it in NODE INSPECTOR.
5. Save; every save creates a revision.

Condition — IF/ELSE branching

Conditions provide deterministic decisions rather than relying on an LLM prompt.

flowchart LR
    A[Research agent] --> C{Condition}
    C -->|true| B[Write report]
    C -->|false| D[Report error]
    B --> E[End]
    D --> E

• Operators: contains, not_contains, equals, not_equals, not_empty, is_empty, case-insensitive.
• Multiple rules use AND; no rules means false.
• Select outgoing edges and set LABEL to true or false.
• The unselected branch is skipped.
• Check uses the same rules but stops the run instead of branching.

Variable templates

Agent/prompt nodes reference earlier output with {{node_id.output}} or {{node_id}}:

Summarize this research: {{researcher.output}}

node_id is the referenced node ID. This syntax also works in HTTP URLs/bodies, Lore queries and condition fields. JSON-body values are escaped. Text without {{ remains unchanged.

Canvas edge interactions

Draw: drag output to input; nearby ports snap. Esc cancels.

Select: click near an edge to open CONNECTION, set LABEL, Add condition, or Delete connection.

Delete: Delete or Backspace removes the selected node or edge.

View: mouse wheel, pinch, +/−, Fit to canvas. The toolbar includes grid visibility, snapping and layers. Undo / Redo restores or reapplies recent changes.

Tools, permissions and approval

MCP TOOLSETS: attach one or more OpenAPI, native MCP or Lore toolsets to a Claude Agent node. The model calls tools and reads results, up to five rounds per node; then it produces a final answer without more tools.

flowchart LR
    M[Model] -->|call| T[Toolset tool]
    T -->|result| M
    M -->|enough information| O[Output]

Require approval (HITL) pauses at the node with Awaiting. In Observability, Approve & continue resumes or Reject fails the run. Waiting does not deduct credits. Approval workflows execute in-process.

Lore context: relevant knowledge from a bound agent's Lore is added before its node runs; see Lore.

Execute

1. Choose Execute; the workflow is saved first.
2. Enter initial input in the dialog.
3. Observability opens; the run also appears in Tale.

The toolbar displays an estimated credit cost.

Import, publish and versions

Import: paste n8n, Flowise or Langflow JSON, or upload a .json file up to 2 MB. Analyze previews mappings. Unmapped nodes become Unsupported placeholders that pass input through and can be manually connected. Dify import is available through the API.

Publish: provide a name/description and select Public to list in Discover → Workflows; Private removes it.

Versions: Version history opens snapshots; Compare versions shows structural differences.

Observability

/legend/observability/:id shows each node's execution. The header includes workflow, run ID and status; counters show Duration, Credits spent, Steps done. TIMELINE lists DONE/FAILED, duration and output/error per node. Approval nodes show Approve & continue / Reject.

flowchart LR
    A[Node 1: done] --> B[Node 2: done]
    B --> C[Node 3: failed]
    C -. resume .-> C2[Retry node 3]

Next step

Connect external tools with Toolsets; manage service registrations and tokens in Connectors.