Inplain
Servidor MCP y CLI que convierten una frase en una presentación, un documento o un panel de control terminados: entrada en texto sin formato y salida como página web independiente. Funciona con tu propia clave de API o completamente en local con Ollama.
Documentación
Plain
Presentations, reports, and dashboards your AI agent can actually make.
Plain turns one sentence into a finished artifact — a deck, a long-form doc, or a data dashboard — rendered as a self-contained web page. It ships as a CLI and an MCP server, so Claude Code, Cursor, or any MCP-aware agent can produce one mid-task and hand back the result.
npm install -g @inplain/cli
export ANTHROPIC_API_KEY=sk-... # or OPENAI_API_KEY, or OLLAMA_BASE_URL
plain install # wires it into Claude Code / Cursor / VS Code
Then, from inside your agent:
"Read src/payments and build me an 8-page deck on what shipped this sprint."
See what comes out: deck ·
doc ·
sheet — each is the
literal output of one plain generate call, published without hand-tuning.
Why an agent needs this
Agents are good at deciding what a document should say. They're bad at
producing the document — most tooling makes them emit a .pptx blob they can't
inspect, or raw HTML they can't reliably edit later.
Plain's artifacts are plain text in, web page out:
- Source is text — Markdown, CSV, a small DSL. An LLM can edit it precisely with a JSON Patch instead of regenerating the whole thing.
- Output is a self-contained page — one HTML file, no build step, no runtime.
- Office formats are an export, not the product.
.pptx/.docx/.xlsxexist for people who need them.
One model, three expressions:
| Form | Is | Source |
|---|---|---|
| deck | A presentation — scroll or present mode | Markdown |
| doc | A long-form report with charts and diagrams | Markdown + frontmatter |
| sheet | A data dashboard (KPIs, charts, tables) | CSV + prose + chart spec |
Use it
As an MCP server — what most people want:
plain install # writes config for Claude Code / Cursor / VS Code
plain mcp # or run it in the foreground to debug
Manual MCP config — the same entry plain install writes. Add it to
~/.cursor/mcp.json or ~/.claude.json:
{
"mcpServers": {
"plain": {
"command": "plain",
"args": ["mcp"]
}
}
}
Tools: generate_artifact, edit_artifact, export, plus per-form variants
and Office exporters (see the caveat below). generate_artifact publishes the
result and returns a shareable URL alongside the source.
As a CLI — scriptable, pipes cleanly:
plain generate "Q3 board update" --as deck -o out.html
plain deck edit out.md --patch '[{"op":"replace","path":"/slides/0/title","value":"Q3"}]'
plain export out.md --to html
Office export (--to pptx|docx|xlsx) currently only handles the older
Marp-style source (marp: true frontmatter), not the v2 format generate
produces. Web-native HTML is the supported path for everything.
As a library — the rendering pipeline is a plain dependency:
import { marpToDeck, deckDocToPptx, findRefs } from "@inplain/core";
Model backends
Any one of these works:
| Provider | Env var |
|---|---|
| Anthropic | ANTHROPIC_API_KEY |
| OpenAI | OPENAI_API_KEY |
GOOGLE_GENERATIVE_AI_API_KEY | |
| Moonshot | MOONSHOT_API_KEY |
| DeepSeek | DEEPSEEK_API_KEY |
| Ollama (local) | OLLAMA_BASE_URL=http://localhost:11434 |
With Ollama, nothing leaves your machine. Each agent role (router, generator,
editor) can point at a different model via ROUTER_PROVIDER / GENERATOR_MODEL
and friends.
What's in here
packages/core/ Rendering pipeline, document model, serializers, Office export
packages/cli/ @inplain/cli — the CLI, the MCP server, and the agent skill
scripts/ Smoke tests (no LLM calls, no network)
packages/core is where an artifact becomes a web page: the V32 document model,
37 templates, the theme engine, and the deck/doc/sheet renderers. It's pure
functions — no IO, no network, no framework.
Develop
pnpm install
pnpm plain --help # run the CLI from source
pnpm build # bundle it
Four checks, all must pass:
pnpm typecheck
pnpm core:smoke # 4 checks — parse, patch, render, export
pnpm build
pnpm cli:smoke # 11 checks — including a live MCP server over JSON-RPC
Neither smoke suite calls an LLM or touches the network, so they run anywhere.
Already have an HTML file?
If you're not generating anything — you already have a finished .html and just
want a link to send someone — that's a different (and much shorter) path:
drop it at inplain.app/html-to-url
and you get an https URL. No account, no build step.
中文说明:把本地 HTML 文件转成链接
From the CLI, same thing in one command:
plain publish ./report.html # → https://inplain.app/s/...
plain publish ./decks/*.html # batch
plain --json publish ./*.html | jq -r .url
Agents can do it too — the MCP server exposes publish_html for files that
already exist on disk (as opposed to generate_artifact, which writes new ones).
Without credentials links expire in 7 days; plain login makes them permanent.
Hosted version
inplain.app runs a hosted workspace on top of this — a web editor, shared links, and a managed model gateway. Everything in this repository works fully against your own API keys; the hosted service is optional.