Claude Code Setup

Add SlideForge

claude mcp add --transport http slideforge https://api.slideforge.dev/mcp/

Verify: claude mcp list → slideforge with 7 tools.

Optional — skills + server config as a plugin:

/plugin marketplace add smartdatabrokers/slideforge-mcp
/plugin install slideforge@slideforge-mcp

Usage

> Make a KPI dashboard slide: revenue \$12.4M (+18% YoY), 847 new clients, NPS 62
> Create a 5-slide QBR deck from the numbers in q3.csv
> Translate deck.pptx to German
> Check board_deck.pptx for quality issues before I send it

The headless self-review loop

There are no inline widgets in a terminal, but the tool result embeds the preview PNG inline — the agent can look at its own render directly out of the response and fix it:

1. create_slide(form=..., data=..., dry_run=true)   # free validation + fidelity forecast
2. create_slide(form=..., data=...)                  # \$0.05 → inline preview PNG in the result
3. (view the inline preview)                         # spot the issue on slide
4. fix the intent → re-render                        # identical input would be free;
                                                     # a changed intent is a fresh \$0.05
5. curl -H "Authorization: Bearer sf_live_YOUR_KEY" \
     -o slide.pptx https://api.slideforge.dev/v1/jobs/<job_id>/pptx   # deliverable

Read the fidelity manifest in every response: verbatim means every number came from your input; ai_completed/mixed names the fields a model filled — tell the user which is which.

Unattended runs

• status: completed_with_errors never bills — read each error's remedy, fix, re-render. Never retry an identical failed payload.
• Identical successful input re-renders free (repeat_of) — idempotent retries are safe.
• Deck renders are parallel and per-slide isolated: a failed slide never blocks (or bills inside) the rest of the deck.