.workflow.md and .canvas.md are plain-text specs — hand the spec to any AI, or let an MCP-connected agent edit the file directly. Either way, Arg0 renders the result.
Two doors in
Door 1
Paste the spec
workflow-format.md and canvas-format.md are portable specs written to be pasted into any AI — ChatGPT, Claude, Gemini. It replies with a valid file; save it into a workspace and Arg0 opens it.
Door 2
MCP-connected agents
Arg0 hosts MCP servers — stdio, streamable-http, and SSE transports, OAuth flows, secrets resolved from env or the OS keychain. Give an agent a filesystem MCP server (or any file-writing tool) and it edits .workflow.md / .canvas.md in place — the app re-renders the same file.
Which door?
Both end at the same file — pick by how much loop you need around the write.
| 1Paste the spec | 2MCP-connected agent | |
|---|---|---|
| Setup | None — copy the spec, paste, save the file | Configure an MCP server once in Settings |
| Iteration | One file per round-trip through chat | Live loop — every edit re-renders in Arg0 |
| Context | The spec — nothing else | Your whole workspace — files, integrations, past runs |
| Review | The file lands fully formed — read, then enable | Watch each edit land; checkpoints gate risky steps |
| Best for | One-shot authoring, sharing templates, web AIs | Iterative builds, refactors, in-app agents |
Most people start at Door 1 for a first draft — then move to Door 2 when the file starts changing every day.
Accuracy & cost
The honest trade-off: Door 1 spends less per file. Door 2 spends more — and self-corrects.
Token cost — Door 1 wins
One spec in, one file out — no tool calls, no re-reads. The agent loop pays for every read, edit, and parse check.
Accuracy — Door 2 wins
A paste can’t verify itself — a dangling edge ships until Arg0 rejects it. The loop catches errors before runtime.
The frugal pattern: draft in Door 1, then let a Door-2 agent parse-check and repair — you spend loop tokens only where errors would have shipped.
The formats
Both specs fit in one Markdown page — sized so an AI can hold the whole grammar in context and emit a valid file in one shot.
.workflow.md
— automation graph.canvas.md
— infinite whiteboardWhy plain text
If a model can write Markdown, it can author for Arg0. No client library, no auth dance — the file is the API.
An agent's edit is a git diff like any other — review it, revert it, branch it. Nothing hides in a binary blob.
The same file drives the canvas UI, the executor, and the docs — an agent's write and your click edit land in the same place.