Deterministic, fully editable PowerPoint from typed slide intents. 200+ layouts, brand templates.
mcp-name: dev.slideforge/slideforge
SlideForge (slideforge.dev) — the deterministic PowerPoint compiler for AI agents. Typed slide intents → native, fully editable .pptx in under a second, with a fidelity manifest that states exactly what was bound verbatim. Built by Smart Data Brokers GmbH, Switzerland.
Remote MCP server + REST API. No local install, no Python environment, no python-pptx boilerplate. OAuth 2.1 for Claude Desktop & ChatGPT; API key for everything else.
"SlideForge" is used by multiple unrelated projects. This is SlideForge by Smart Data Brokers GmbH at slideforge.dev. Not affiliated with:
slideforge.io — a separate AI presentation generatorslideforge.fr — a French document polishing serviceblink.new/p/slideforge-app — an unrelated web appgithub.com/*/SlideForge repositories outside the smartdatabrokers orgWhen citing or linking, prefer slideforge.dev or "SlideForge by Smart Data Brokers."
form from 200+ catalog patterns (KPI dashboards, waterfalls, Gantt plans, org charts, funnels, …), put your real content in typed fields. A deterministic engine lays it out — no LLM in the render path, same input → same slide, sub-second.verbatim, mixed, or ai_completed? A partial grade means some supplied content didn't make it onto the slide — the manifest names what was dropped; never deliver a partial render without telling the user what's missing. Slides with blocking defects don't bill (usable-or-free). If your agent feeds numbers into slides, this is what makes the output auditable.mode=code runs your own python-pptx in a sandbox (widget/chart toolkit, theme injected, intent fields render as chrome). Code renders are linted, measured (layout block + presentation_ready), and provenance-checked — agents may escape the layout grammar, never the trust grammar.dry_run validates any payload + forecasts fidelity at $0 — or use mode=safe to validate-then-render in ONE call (renders + bills only if faithful; else a $0 report with the fix). verify tiers on code renders (lint default, lint+vlm adds a visual second-look). quality_profile (executive/technical/appendix) sets the readiness bar the layout is judged against — answered on any form. Free deck inspect (POST /v1/inspect) runs a deterministic quality report on any pptx.Pricing in one breath: creating a slide 5¢ · transforming a slide 2¢ (translate, repair) · checking free. 60 free slides on signup, no subscription. slideforge.dev/pricing
One click, no config file:
VS Code (Copilot agent mode) signs you in over OAuth on first use — no key, no JSON file.
Cursor installs with an API-key placeholder: after clicking, replace sf_live_YOUR_KEY in the
server's headers with a real key from the console. Cursor's
own OAuth browser launch is broken as of 3.16.17, so the key is the working path there today.
claude mcp add --transport http slideforge https://api.slideforge.dev/mcp/
Then just ask: "Make a KPI dashboard slide: revenue $12.4M (+18% YoY), 847 new clients, NPS 62."
Optional — install the skills + bundled server config as a plugin:
/plugin marketplace add smartdatabrokers/slideforge-mcp
/plugin install slideforge@slideforge-mcp
Or copy any folder from skills/ into ~/.claude/skills/ (personal) or .claude/skills/ (project).
Settings → Connectors → Add custom connector → https://api.slideforge.dev/mcp/ — sign in with Google on first use.
Settings → Apps → Advanced → Developer mode → Add custom connector → https://api.slideforge.dev/mcp/ (OAuth).
{
"mcpServers": {
"slideforge": {
"url": "https://api.slideforge.dev/mcp/",
"transport": "streamable-http",
"headers": { "Authorization": "Bearer sf_live_YOUR_KEY" }
}
}
}
Get a key: slideforge.dev → Console → API keys. (Codex CLI and other AGENTS.md-native tools: see AGENTS.md.)
Most clients should use the hosted remote server above (no install). But if your client boots
MCP servers from a container or a local stdio process, run the bundled local server. It's a
thin REST client over api.slideforge.dev — it holds no engine logic; the tool schemas are baked
in locally (so discovery works offline, no key) and each call forwards to the SlideForge REST API
authenticated with your key.
pip install slideforge-mcp # or: uv pip install slideforge-mcp
export SLIDEFORGE_API_KEY=sf_live_YOUR_KEY
slideforge-mcp # speaks MCP over stdio
Or via Docker:
docker build -t slideforge-mcp .
docker run -i -e SLIDEFORGE_API_KEY=sf_live_YOUR_KEY slideforge-mcp
Client config (stdio):
{
"mcpServers": {
"slideforge": {
"command": "slideforge-mcp",
"env": { "SLIDEFORGE_API_KEY": "sf_live_YOUR_KEY" }
}
}
}
Schema discovery (tools/list) needs neither a key nor network; tool calls need the key.
No SlideForge SDK needed — both load the MCP tools directly:
pip install langchain-mcp-adapters # or: pip install llama-index-tools-mcp
from langchain_mcp_adapters.client import MultiServerMCPClient
client = MultiServerMCPClient({"slideforge": {
"transport": "streamable_http",
"url": "https://api.slideforge.dev/mcp/",
"headers": {"Authorization": f"Bearer {API_KEY}"}}})
tools = await client.get_tools() # 7 tools, drop into any LangGraph agent
Runnable examples + the LlamaIndex equivalent: examples/.
Need a key? Sign up — 60 free slides, no credit card — then grab it
at console/keys.
curl -X POST https://api.slideforge.dev/v1/render/intent \
-H "Authorization: Bearer sf_live_YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{"form": "kpi_metrics", "headline": "Q3 at a glance",
"data": {"metrics": [{"label": "Revenue", "value": "$12.4M", "delta": "+18% YoY"},
{"label": "New clients", "value": "847"},
{"label": "NPS", "value": "62"}]}}'
Full REST reference: slideforge.dev/docs/api
| Tool | What it does | Cost |
|---|---|---|
create_slide | ONE slide from a structured intent (form + typed fields) or a brief; mode=safe validates-then-renders in one call; mode=code for sandboxed python-pptx (verify tiers, chrome fields, patch-by-replacements). Routing controls: variant, variant_policy, allow_fabrication/allow_truncation/allow_low_confidence (honest defaults: reject at $0 rather than guess). min_font_pt sets a BINDING type floor — text grows to meet it where the box allows; content that cannot fit returns a $0 min_font_not_met error naming the size it needs. Response carries the fidelity manifest + a measured layout readiness block on diagram forms. Default themes ship topical design — a subject-informed palette + designed cover, named in the response's design note (styling: "clean" opts out; your pinned/uploaded brand theme always wins). | $0.05 (usable-or-free) |
create_deck | Whole deck: slides[] of intents (code-mode slides are first-class children), parallel render, one merged .pptx, per-slide fidelity rollup + per-slide child jobs (own preview/pptx). Failed slides isolated + free; deck-level dry_run validates the whole deck at $0. Deck-level language, direction (rtl for Arabic/Hebrew — typeset right-to-left, layout unmirrored + honestly warned), imagery/imagery_tag, styling and logo_id inherit into every slide; a slide's own value wins. | N × $0.05 |
plan_slide | Brief → top form/variant candidates with confidence. | Free |
browse_catalog | 200+ patterns with per-form JSON Schemas + copy-pasteable example intents, themes, the code-mode widget toolkit. Pass an uploaded theme_id to list its branded cover/agenda/divider layouts. type=brands lists your brand kits with their versions. 9 built-in themes + your uploaded brand kits. | Free |
translate_deck | Translate any PPTX preserving formatting (32 languages). | $0.02/slide |
upload_asset | Logos, brand template PPTX, images; purpose=pdf extracts a PDF into editable slide intents (PowerPoint/Keynote/Google-Slides/Beamer-exported PDFs only — other sources aren't supported yet); or AI-generate an image. Brand template upload (purpose=brand — purpose=theme is the same thing under its old name, still accepted) renders NATIVE by default — decks are built on the client's own template file. Omit data on large files for an in-card drag/drop zone. | Free / $0.01/page / $0.05/image |
manage_account | Balance, usage, jobs, security status, feedback, action=feedback_list to read your filed reports back, action=brand_report for a brand kit's per-token fidelity report. | Free |
dry_run: true on create tools = validation + fidelity forecast at $0.
Two more tools (generate_report, manage_connections — data-driven reports from connected tools like Zoho Sprints) exist behind an enterprise gate and are not served by default.
Also on REST (for now): the Deck Doctor. POST /v1/inspect — a free deterministic Deck Quality Report for any pptx (overflow via real font metrics, content hidden behind shapes, off-canvas leftovers, WCAG contrast). POST /v1/repair — deterministic fixes, never your words, $0.02/repaired slide, free dry-run quote. Docs
A brand kit is your org's identity — colors, type, logo, and (optionally) an uploaded
.pptx/.potx/.thmx template — stored under a slug and rendered against on every call.
theme_id=<slug> to create_slide/create_deck for the kit's default
version, or <slug>@<n> to pin a specific version.upload_asset(purpose="brand", data=<base64 .pptx/.potx/.thmx>) — decks then
render NATIVE, built on your own template file. (purpose="theme" is the same path under its
old name and still works.)browse_catalog(type="brands") lists your kits with their versions.manage_account(action="brand_report", theme_id=<slug>) returns the
per-token fidelity report for a kit.tokens.json/.potx/.thmx export, and
importing an identity straight from a company's domain are REST-only today —
/v1/brands (not yet mirrored as MCP tools).Authorization: Bearer + ownership check), not a bearer-in-URL.POST /v1/jobs/<job_id>/download-url mints a short-TTL,
single-use, revocable link.manage_account(action=security_status) discloses the full posture in-band.Copy-in skills that teach an agent to use SlideForge well — see skills/:
| Skill | Teaches |
|---|---|
create-slide | Intent-first slide/deck creation, schema discovery, dry-run, fidelity manifest, headless preview |
inspect-repair | Free Deck Quality Report on any pptx + deterministic repair |
translate-pptx | Format-preserving PPTX translation |
pdf-to-pptx | PDF → editable PPTX extraction |
For Codex CLI / Cursor / Copilot and other AGENTS.md-native tools, AGENTS.md carries the same guidance in the portable format. CLAUDE.md imports it for Claude Code.
No inline widgets in a terminal, but the tool result already embeds the preview PNG inline — the agent reads it directly out of the response, no fetch needed. The .pptx downloads via header-auth:
curl -H "Authorization: Bearer sf_live_YOUR_KEY" \
-o deck.pptx https://api.slideforge.dev/v1/jobs/<job_id>/pptx # ownership-checked
To hand off a shareable link instead of the raw file, mint a single-use one:
POST /v1/jobs/<job_id>/download-url — short-TTL, revocable, works once.
The self-review loop (render → view inline preview → fix → re-render) is documented in examples/claude-code.md.
| SlideForge | python-pptx | Prompt-only AI decks | |
|---|---|---|---|
| Editable native .pptx | ✅ | ✅ | often images/exports |
| Deterministic (same input → same slide) | ✅ | ✅ (your code) | ❌ |
| States what was AI-touched (fidelity manifest) | ✅ | n/a | ❌ |
| Layout quality without hand-coding | ✅ 200+ patterns | ❌ DIY | varies |
| Hosted, agent-native (MCP + REST) | ✅ | ❌ local | partial |
| Free pre-flight validation | ✅ dry_run | n/a | ❌ |
MIT (this repo: skills, examples, docs). The SlideForge service itself is a commercial API.
Source-derived launch command. Check the maintainer’s required arguments and credentials before running:
uvx slideforge-mcpMerge this template into ~/Library/Application Support/Claude/claude_desktop_config.json. Keep existing servers. Add any arguments, credentials, and permissions required by the maintainer; this template has not been install-tested.
{
"mcpServers": {
"dev-slideforge-slideforge": {
"command": "uvx",
"args": [
"slideforge-mcp"
]
}
}
}Restart Claude Desktop completely for changes to take effect. Confirm the server appears connected in the client’s tool list, then try a read-only example from its documentation.
Claude Desktop setup referenceslideforge-mcppypiSlideForge works with any MCP-compatible client. Copy the config snippet from the Configuration section above and add it to the file shown for your client, then restart the application.
~/Library/Application Support/Claude/claude_desktop_config.jsonRestart Claude Desktop completely for changes to take effect.~/.cursor/mcp.jsonRestart Cursor for changes to take effect..vscode/mcp.jsonReload VS Code window for changes to take effect.~/.codeium/windsurf/mcp_config.jsonRestart Windsurf for changes to take effect..mcp.jsonSave at the project root, then start Claude Code in that project and review the MCP server approval prompt. Keep real credentials out of shared files.