Skip to content

— agent · mcp

claude drafts
the workflow.
your designer
ships it.

onfield speaks the model-context-protocol. claude desktop, cursor, or any mcp client can list step types, draft a workflow, validate it against the schema, and save a draft against your org. the designer remains the publisher. agents propose. humans approve.

mcp/1v0.4
live today get started

— 01 — list

[mcp/1]

the catalog, before the draft.

the agent calls list_step_types first. it gets the catalog of valid types, the config shape each one expects, and example bodies for branch, repeater, and form. it has the same ground truth your designer does.

  1. agent list_step_types
  2. server 35 types · example bodies · config schemas
list_step_types ok · 38ms

— 02 — draft

[mcp/1]

from transcript
to runnable workflow.

paste the transcript of how your senior tech describes the job. ask claude to draft the workflow. it picks the right step types, writes the conditional logic, names the inputs — and validates against the schema before it saves. workflows with branches or repeaters typically cost 2–3 save→fix iterations without that validate step. with it, the save is clean.

  1. agent create_workflow
  2. agent validate_workflow_body
  3. server valid: true · 0 errors · 0 warnings
  4. agent save_workflow_body
  5. server saved · version 2 · parent v1

— 03 — publish

[mcp/1]

only humans publish.

the agent saves a draft. it cannot publish. a human in the designer reviews the body, reads the validation report, and hits publish. version 2 lights up on every device on next login.

— clients

[mcp/1]

  • claude desktop stable
  • cursor stable
  • any mcp client stable

we publish the server spec. anything that speaks mcp can talk to onfield.

— governance

[mcp/1]

  • · org-scoped api key, rotated from settings
  • · agent acts as a draft author, never a publisher
  • · every agent call is logged