FlowZap Diagram Skill
Generate valid FlowZap Code (.fz) diagrams from natural-language requests,
validate them, and return shareable FlowZap playground URLs.
This skill is designed to be paired with the flowzap-mcp server. The skill
gives the agent FlowZap-specific diagram knowledge and output rules; the MCP
server provides the actual tools.
Trust and privacy summary
- This skill bundle is documentation only:
SKILL.md plus reference files. It
does not execute code by itself.
- The
flowzap-mcp server runs locally on the user's machine over stdio.
- Outbound requests are restricted to
https://flowzap.xyz.
- Raw HTTP logs, OpenAPI specs, and code snippets are parsed locally inside the
MCP package. Only generated FlowZap Code is sent to FlowZap public endpoints
when validation or playground creation is requested.
When to use this skill
- User asks for a workflow, flowchart, process diagram, sequence diagram, or architecture diagram.
- User pastes HTTP logs, OpenAPI specs, or code and wants them visualised.
- User wants to compare two diagram versions (diff) or patch an existing diagram.
Installation model
This setup has two parts:
- Install the skill bundle so your agent knows when to use FlowZap and how
to produce correct FlowZap Code.
- Install the MCP server so your agent can validate code, create
playground URLs, diff diagrams, and apply structured changes.
Install the skill bundle
Recommended:
npx skills add flowzap-xyz/flowzap-mcp --skill flowzap-diagrams
If your agent does not support skills.sh, install this bundle through your
agent's skill manager or by copying the folder into the agent's local skills
directory.
Manual locations:
- Claude Code:
.claude/skills/flowzap-diagrams/SKILL.md
- Windsurf:
.windsurf/skills/flowzap-diagrams/SKILL.md
- Cursor:
.cursor/skills/flowzap-diagrams/SKILL.md
Install the MCP server (required for tools)
If the FlowZap MCP server is not already configured, install it:
# Claude Code
claude mcp add --transport stdio flowzap -- npx -y flowzap-mcp@1.3.6
# Or add to .mcp.json / claude_desktop_config.json / cursor / windsurf config:
{
"mcpServers": {
"flowzap": {
"command": "npx",
"args": ["-y", "flowzap-mcp@1.3.6"]
}
}
}
Package verification
The pinned version 1.3.6 can be verified against the npm registry:
To verify locally before use:
npm view flowzap-mcp@1.3.6 dist.integrity dist.shasum
Compatible tools: Claude Desktop, Claude Code, Cursor, Windsurf, OpenAI Codex,
Warp, Zed, Cline, Roo Code, Continue.dev, Sourcegraph Cody.
Not compatible: Replit, Lovable.dev.
Without the MCP server, this skill can still help an agent draft FlowZap Code,
but it cannot validate diagrams, create playground URLs, or use the FlowZap
tooling workflow described below.
Available MCP tools
| Tool | Purpose |
|---|
flowzap_validate | Check .fz syntax before sharing |
flowzap_create_playground | Get a shareable playground URL |
flowzap_get_syntax | Retrieve full DSL docs at runtime |
flowzap_export_graph | Export diagram as structured JSON (lanes, nodes, edges) |
flowzap_artifact_to_diagram | Parse HTTP logs / OpenAPI / code → diagram + playground URL |
flowzap_diff | Structured diff between two .fz versions |
flowzap_apply_change | Patch a diagram (insert/remove/update nodes/edges) |
FlowZap Code DSL — quick reference
FlowZap Code is not Mermaid, not PlantUML. It is a unique DSL offering a simple syntax for a triple-view option to workflow, sequence and architecture diagrams.
Shapes (only 4)
| Shape | Use for |
|---|
circle | Start / End events |
rectangle | Process steps / actions |
diamond | Decisions (Yes/No branching) |
taskbox | Assigned tasks (owner, description, system) |
Syntax rules
- Node IDs are globally unique, sequential, no gaps:
n1, n2, n3 …
- Node attributes use colon:
label:"Text"
- Edge labels use equals inside brackets:
[label="Text"]
- Handles are required on every edge:
n1.handle(right) -> n2.handle(left)
- Directions:
left, right, top, bottom
- Cross-lane edges: prefix target with lane name:
sales.n5.handle(top)
- Lane display label: one
# Label comment on the same line as the opening brace
- Loops:
loop [condition] n1 n2 n3 — flat, inside a lane block
- Layout: prefer horizontal left→right; use top/bottom only for cross-lane hops
Multi-lane sequence design
- Ping-pong rule: For multi-participant processes, every cross-lane interaction must alternate back-and-forth between lanes. A request from Lane A → Lane B must be followed by a response from Lane B → Lane A before any new major cross-lane request begins. This is now a strict validation requirement, not just a readability suggestion.
- Chronological order: The sequence view follows cross-lane edge definition order. Define request, response, then next request in the exact order they happen.
Gotchas — never do these
- Do NOT use
label="Text" on nodes (must be label:"Text").
- Do NOT use
label:"Text" on edges (must be [label="Text"]).
- Do NOT skip node numbers (n1, n3 → invalid; must be n1, n2).
- Do NOT omit lane prefix on cross-lane edges.
- Do NOT output Mermaid, PlantUML, or any other syntax.
- Do NOT add comments except the single
# Display Label per lane.
- Do NOT place
loop outside a lane's braces.
- Do NOT use a
taskbox shape unless the user explicitly requests it.
Minimal templates
Single lane:
process { # Process
n1: circle label:"Start"
n2: rectangle label:"Step"
n3: circle label:"End"
n1.handle(right) -> n2.handle(left)
n2.handle(right) -> n3.handle(left)
}
Two lanes with cross-lane edge:
user { # User
n1: circle label:"Start"
n2: rectangle label:"Submit"
n5: rectangle label:"Receive result"
n1.handle(right) -> n2.handle(left)
n2.handle(bottom) -> app.n3.handle(top) [label="Send"]
}
app { # App
n3: rectangle label:"Process"
n4: rectangle label:"Respond"
n3.handle(right) -> n4.handle(left)
n4.handle(top) -> user.n5.handle(bottom) [label="Result"]
}
Decision branch:
flow { # Flow
n1: rectangle label:"Check"
n2: diamond label:"OK?"
n3: rectangle label:"Fix"
n4: rectangle label:"Proceed"
n1.handle(right) -> n2.handle(left)
n2.handle(bottom) -> n3.handle(top) [label="No"]
n2.handle(right) -> n4.handle(left) [label="Yes"]
}
For the full DSL specification and advanced multi-lane examples: See references/syntax.md
Workflow: how to generate a diagram
- Identify the actors/systems (→ lanes) and steps (→ nodes) from the user's description.
- Write FlowZap Code following all rules above.
- Call
flowzap_validate to verify syntax.
- If valid, call
flowzap_create_playground to get a shareable URL.
- Return the FlowZap Code and the playground URL to the user.
- Always output only raw FlowZap Code when showing the diagram — no Markdown fences wrapping .fz content, no extra commentary mixed in.
Full MCP documentation: flowzap.xyz/docs/mcp
Security and data transparency
The trust boundary is intentionally narrow:
- The skill bundle is static Markdown and reference text.
- The MCP server runs locally and only calls public FlowZap APIs.
- Outbound traffic is restricted to
https://flowzap.xyz.
- Validation is stateless; playground sessions are time-limited.
The flowzap-mcp server runs locally on the user's machine (stdio transport) and enforces the following safeguards:
| Control | Detail |
|---|
| TLS only | All outbound requests require https:// and are restricted to flowzap.xyz (SSRF protection) |
| No authentication | Uses only public FlowZap APIs; no API keys, tokens, or user credentials are stored or transmitted |
| No user-data access | Cannot read diagrams, accounts, or any data beyond what the agent explicitly passes in |
| Input validation | Code capped at 50 KB, total input at 100 KB; null bytes and control characters stripped |
| Rate limiting | Client-side 30 requests/minute sliding window |
| Request timeout | 30-second hard timeout with AbortController |
| Response sanitization | Only expected fields are returned; playground URLs validated against allowlist |
| Audit logging | All tool calls and API requests logged to stderr (not exposed to the MCP client) |
Data flow scope
The MCP server processes raw inputs locally and sends only generated or agent-provided FlowZap Code to FlowZap public endpoints:
POST https://flowzap.xyz/api/validate — returns syntax validation resultPOST https://flowzap.xyz/api/playground/create — returns an ephemeral playground URL (60-minute TTL, non-guessable token)
If the agent uses flowzap_artifact_to_diagram, the raw HTTP logs, OpenAPI spec,
or code snippet are parsed locally inside the MCP package first. Only the
resulting FlowZap Code is sent when a playground URL is created.
No local file paths, environment variables, user identity, repository contents,
or credentials are transmitted by the MCP package.
Playground URL access controls
Playground URLs are ephemeral, time-limited (60-minute TTL), and use non-guessable cryptographic tokens. They are read-only views of the diagram code submitted at creation time. No account or login is required to view them; no data persists beyond the TTL.
Data lifecycle
| Endpoint | Data stored | Retention |
|---|
POST /api/validate | None — stateless; code is parsed in memory and discarded after the response | 0 (not persisted) |
POST /api/playground/create | FlowZap Code only (in PostgreSQL) | 60 minutes (database row + playground URL both expire) |
The playground session is stored server-side with a cryptographic token (UUID v4). After the 60-minute TTL, the session is deleted — either on the next access attempt or during a periodic sweep. No user identity, file paths, environment variables, or host metadata are attached to the session.
What the MCP server does NOT do
- No filesystem access — cannot read or write files on the host machine
- No environment variable access — does not read or transmit
process.env or shell variables
- No code execution — does not evaluate, compile, or run any user code; it only transmits FlowZap DSL text
- No network scanning — outbound connections are restricted to
flowzap.xyz over TLS (SSRF-protected allowlist)
- No long-term data persistence — playground sessions expire after 60 minutes; the validate endpoint stores nothing
- No telemetry or tracking — no analytics, device fingerprinting, or usage data is collected by the MCP server; server-side API logs record only IP, user-agent, and code length (not code content)
Further resources
