Why Agentic Mermaid

Use it when you want to describe a diagram in plain language and get back something you can publish without a design cleanup pass.

The agent workflow is the guardrail behind the polish: agents should not guess from pixels, concatenate strings, or regenerate whole diagrams when a structured edit is available.

Highlights

• Descriptor-registered diagram families — built-ins and namespaced extensions share one discovery and capability contract.
• SVG, PNG, ASCII, Unicode, JSON — one resolved request with explicit graphical, terminal, and positioned-layout projections.
• Synchronous, zero-DOM SVG renderer — no Puppeteer, no browser flash.
• Composable styles — { style: ['hand-drawn', 'dracula'] } stacks a look over a palette; discoverable full looks cover sketch, watercolor, blueprint, accessibility, print, operational, physical-media, architecture, and editorial/report use cases. Custom styles are plain JSON records any agent can author (docs/style-authoring.md). seed re-rolls the ink, never the layout.
• Discoverable palettes + Shiki compatibility — a theme is a palette-only style: discover the canonical catalog at runtime, theme from two colors, or adapt a VS Code theme.
• Agent-native editing — typed mutation for every registered renderable family; source-level round-trip only for opaque fallbacks containing unmodeled syntax.
• CLI + MCP + library — am, agentic-mermaid-mcp, agentic-mermaid, agentic-mermaid/agent, and the browser/workerd-safe agentic-mermaid/agent/core. Audit reports and trusted host-resource helpers remain repository tooling instead of published runtime entry points.

Installation

npm install agentic-mermaid       # or: bun add agentic-mermaid / pnpm add agentic-mermaid
npx --no-install agentic-mermaid --help
npx --no-install agentic-mermaid mcp

For repository development, install from source and run the Bun entrypoints:

git clone https://github.com/adewale/agentic-mermaid
cd agentic-mermaid
bun install
bun run build
bun run bin/am.ts --help
bun run bin/agentic-mermaid-mcp.ts   # MCP stdio server

ESM-only. agentic-mermaid ships ES modules (there is no CommonJS build); require() consumers should use dynamic import() instead. Requires Node ≥ 22.

The am … examples below name the published bin. After a project-local npm install, invoke it from a shell as npx --no-install agentic-mermaid … (or from an npm script as am …). From a source checkout, use bun run bin/am.ts … instead.

Output quick starts

Use agentic-mermaid/agent when you want one import path for styled renders, output formats, and the structured edit API.

SVG

import { renderMermaidSVG } from 'agentic-mermaid/agent'

const svg = renderMermaidSVG(`flowchart TD
  Start --> Done`, { security: 'strict' })

PNG

import { writeFileSync } from 'node:fs'
import { renderMermaidPNG } from 'agentic-mermaid/agent'

const png = renderMermaidPNG(`flowchart TD
  Start --> Done`, {
  fitTo: { width: 1200 },
  background: '#fff',
})

writeFileSync('diagram.png', png)

CLI equivalent:

am render diagram.mmd --format png --output diagram.png

ASCII / Unicode

import { renderMermaidASCII } from 'agentic-mermaid/agent'

const unicode = renderMermaidASCII(`flowchart LR
  A --> B`)
const ascii = renderMermaidASCII(`flowchart LR
  A --> B`, { useAscii: true })

Agent quick start

If your coding agent can read repo files, point it at:

• skills/agentic-mermaid-diagram-workflow/SKILL.md for diagram authoring/editing.
• skills/agentic-mermaid-live-editor/SKILL.md for editor changes.

If it only has shell access:

am --agent-instructions
am capabilities --json
am preview diagram.mmd --security strict --open
am mutate diagram.mmd --op '{"kind":"add_node","id":"Cache","label":"Cache"}' --json

Zero-install prompt for a coding agent: read https://agentic-mermaid.dev/llms.txt and follow the parse → narrow → mutate → verify → serialize workflow. To wire Agentic Mermaid into another repo, run npx agentic-mermaid init-agent (or bun run bin/am.ts init-agent from a source checkout); it writes a non-clobbering AGENTS.md section, root skills/ bundle, and .mcp.json sample.

Use strict preview for human inspection and mutate --op/--ops for verified one-shot or batched edits.

For multi-step MCP edits, connect agentic-mermaid-mcp and use Code Mode execute(code) with the same mermaid.* SDK names. Stdio is the default transport; agentic-mermaid-mcp --transport http starts HTTP/SSE and managed PNG file/URL artifacts. See the agent API cookbook for copy-pasteable library, CLI, and MCP recipes.

MCP server

Agentic Mermaid ships a Model Context Protocol server so MCP-capable agents can render and safely edit diagrams without shelling out.

• Self-hosted (default). agentic-mermaid-mcp runs a stdio server exposing execute (Code Mode sandbox), describe_sdk (one family's mutation schema on demand), render_png, and describe. Package runners can use npx -y agentic-mermaid mcp; the mcp argument routes the package-name binary to the same stdio server. Add --transport http for HTTP/SSE with managed PNG file/URL artifacts. See docs/mcp-http-transport.md and docs/mcp-code-mode-rationale.md.
• Hosted. A stateless Streamable HTTP endpoint is available at https://agentic-mermaid.dev/mcp (tools: execute, describe_sdk, render_svg, render_ascii, render_png, verify, describe, mutate, build; 64 KB input caps). Call describe_sdk for compact signatures or exact fields before authoring unfamiliar ops. It is MCP JSON-RPC only, not a REST render API. Hosted execute runs the same mermaid.* facade in a Cloudflare Dynamic Worker isolate with no network; hosted PNG returns base64 only.

Local-first is the default posture: prefer the library, CLI, or a self-hosted MCP for anything sensitive, offline, larger than the hosted caps, or needing local PNG file/URL artifacts. The hosted endpoint is a public, unauthenticated convenience for zero-install render/verify/describe and bounded structured edits.

Directory maintainers can use the canonical MCP listing record. Hosted data handling is described in the MCP privacy notice.

Structured edit example

import { parseRegisteredMermaid, asFlowchart, mutate, verifyMermaid, serializeMermaid } from 'agentic-mermaid/agent'

const parsed = parseRegisteredMermaid('flowchart TD\n  API --> DB')
if (!parsed.ok) throw new Error('parse failed')

const flow = asFlowchart(parsed.value)
if (!flow) throw new Error(`not a structured flowchart: ${parsed.value.kind}`)

const next = mutate(flow, { kind: 'add_node', id: 'Cache', label: 'Cache' })
if (!next.ok) throw new Error(next.error.message)

const verify = verifyMermaid(next.value)
if (!verify.ok) throw new Error(JSON.stringify(verify.warnings, null, 2))

const source = serializeMermaid(next.value)

Rules:

• Use the matching exported as<Family> narrower before mutating an existing structured diagram.
• Mutation ops use kind, not type.
• Run verifyMermaid before every commit point.
• Do not call mutate on opaque fallback bodies; the narrower returns null for unmodeled syntax.

Supported diagram families

Family support and its executable evidence are projected from the FamilyDescriptor registry into the generated Section A capability report. See diagram families for syntax examples and compatibility notes.

More documentation