MarsDawn
Review without spending your agent's tokens.
Two separate things stay cheap in this loop: what the agent gets back from calling the tool, and what it takes to confirm the result looks right.
The tool call returns a small JSON object, not the rendered page, and the rendered page itself is reviewed by a person in MarsDawn — never read back into the agent's context.
The tool call itself is cheap
Call marsdawn export — from the CLI, the skill, or the MCP server — and what comes back is a compact JSON object: ok, output, pages, theme, paper and diagramErrors. The full schema is export.v1.json. None of that is the rendered document. A 50-page PDF with a dozen Mermaid diagrams returns the same handful of fields as a one-page note.
Review happens off to the side
Once the PDF exists, a person opens it — in MarsDawn, or any PDF viewer — and reads the diagrams, the math and the layout rendered. The agent never needs that rendered output read back into its own context window to confirm it looks right: the review happens in a separate window, on a separate screen, not as another round-trip of tokens spent describing what a diagram looks like.
What this avoids
- Pasting rendered Markdown, a screenshot, or a description of one back into the conversation just so the agent can confirm the export worked.
- An agent that has to reconstruct what a Mermaid diagram or a KaTeX formula renders as, rather than a person just looking at it.
- A second tool call to fetch the PDF's contents after the first one already reported success.
Next
- The three ways to call marsdawn — CLI, skill file, MCP server: three ways in.
- Every field in the JSON result: marsdawn for agents.
- Why a person still needs to read what an agent wrote: the case for review.