endiagram
# @endiagram/mcp
MCP server for [EN Diagram](https://endiagram.com) — deterministic structural analysis powered by graph theory. Every result is backed by a named mathematical theorem. No AI inside the computation.
## Installation
Run directly:
```bash
npx @endiagram/mcp
```
Or install globally:
```bash
npm install -g @endiagram/mcp
```
## Connect
### Claude Code
```bash
claude mcp add endiagram npx @endiagram/mcp
```
### Claude Desktop
Add to your `claude_desktop_config.json`:
```json
{
"mcpServers": {
"endiagram": {
"command": "npx",
"args": ["@endiagram/mcp"]
}
}
}
```
### Cursor
Add to `.cursor/mcp.json` in your project root:
```json
{
"mcpServers": {
"endiagram": {
"command": "npx",
"args": ["@endiagram/mcp"]
}
}
}
```
### HTTP (zero install)
Any MCP client that supports HTTP transport:
```
https://api.endiagram.com/mcp
```
### Smithery
```bash
smithery mcp add dushyant30suthar/endiagram
```
## Environment Variables
| Variable | Default | Description |
|----------|---------|-------------|
| `EN_API_URL` | `https://api.endiagram.com` | API endpoint for the EN Diagram service |
## Tools
Six questions about any system, plus a render tool. Every tool takes `source` (EN code or `.en`/`.txt` file path). Tool names are shorthand, not specs — read each description before calling; `compose` and `equivalent` are mode-based, not general analyzers.
| Tool | What it answers | Levers |
|------|-----------------|--------|
| `structure` | What is this system? Shape, stages, bridges, cycles, critical path, dominator tree, min-cuts, subsystems, actors, locations. | `detect_findings=true` flags risks (unguarded-sink, single-cut-path, multi-cut-path); `node=X` returns per-node centrality (betweenness, closeness, eigenvector). |
| `invariant` | What's always true? Conservation laws, T-invariants (sustainable cycles), depletable sets, deficiency, reversibility. | `rules` (one per line) checks custom claims. Four supported shapes: `no bridge that is also hub` · `every path from X to Y passes through at least one of [A,B,C]` (precedence) · `no node with centrality above N` · `removing any single node disconnects at most N others`. |
| `live` | Can it deadlock? Can entities overflow? Siphons, traps, unbounded cycles, structural liveness and boundedness. | — |
| `reachable` | Can X reach Y? Path, distance, boundary crossings. `from`/`to` accept entity or action names. | `defense_nodes=a,b,c` checks whether guards cover every path. |
| `equivalent` | Are two systems the same, or what changes if I change this one? | Compare mode (`source_a`+`source_b`): edit distance + spectral cospectrality. Evolve mode (`source`+`patch`): plain EN adds; `- name` removes; same-name replaces. |
| `compose` | How do parts combine (merge) or how does a part stand alone (extract)? | Merge: `source_a`+`source_b`+`links` (`a.entity=b.entity` per line). Extract: `source`+`subsystem` (names come from `structure.subsystems`). |
| `render` | SVG or PNG diagram. Only call when the user asks to visualize. | Themes: `Editorial`, `Primer`, `Carbon` (each ± `isDark`) or seed-derived from `color=#RRGGBB`. `structure_layers` bitmask (1=subsystems, 2=pipelines, 4=cycles, 8=forks, 16=joins, 32=hubs, 64=deadlock, 128=overflow). |
## EN Syntax
One statement per line:
```
actor do: action needs: input1, input2 yields: output1, output2
```
Shared names between yields and needs create connections automatically:
```
customer do: place order needs: menu yields: order
kitchen do: prepare food needs: order yields: meal
waiter do: deliver needs: meal yields: served customer
```
## Modeling
Same name = same thing. Put all required inputs in one `needs:` list (AND). Give two actions the same yield-name to offer alternatives (OR). Sequence = one action consuming another's yield. Re-yield stateful resources to keep them persistent; one-shot data stays consumed. Model at the real atomic granularity — split only when the pieces could be accessed independently.
Learn more at [endiagram.com](https://endiagram.com).
## Telemetry
`@endiagram/mcp` generates a random install ID on first run, stored at
`~/.endiagram/install-id` (mode `0600`). It is sent with every request as
the `X-Endiagram-Install-Id` HTTP header so we can correlate requests
from the same install for debugging issues that the per-IP signal alone
cannot track (mobile networks, VPNs, CGNAT all collapse or churn IPs).
**No source code, no file paths, no environment variables, and no PII
are sent.** The install ID is a random opaque UUIDv4 generated locally.
A first-run notice prints to **stderr** (never stdout — stdout is the
MCP JSON-RPC channel) with the disclosure and the opt-out instructions.
The notice fires once per install and never again.
### Opting out
Any of these three methods disables the install ID:
1. Set `ENDIAGRAM_TELEMETRY=off` as an environment variable (also
accepts `0`, `false`, `no`).
2. Create a file at `~/.endiagram/telemetry` containing the word `off`.
3. Delete `~/.endiagram/install-id`. (A new one is generated on next
run unless option 1 or 2 is also set.)
When any of these is active, the `X-Endiagram-Install-Id` header is not
sent at all — the server falls back to its per-IP HMAC `cid` for
correlation, which works fine for short-term per-session tracing.
Full privacy policy: [endiagram.com/privacy](https://endiagram.com/privacy)
## License
MIT
TDQS
Scored across 12 tools
Most tools target distinct graph operations, but trace, distance, and between all revolve around paths and centrality, while analyze_system, detail, and categorize overlap in structural analysis. Descriptions help but an agent could reasonably pick the wrong tool for some analytical tasks.
Tool names are mostly single lowercase words, but they mix verbs, nouns, and prepositions (between, distance, impact, detail), and analyze_system breaks the pattern with snake_case. There is no consistent verb_noun or predictable naming convention.
Twelve tools is well-scoped for a dependency-graph analysis server. Each tool has a recognizable responsibility within analysis, rendering, transformation, and composition, without feeling padded or thin.
The toolkit covers traversal, centrality, structural diffing, subsystem discovery, deep analysis, rendering, merging, and dry-run evolution. Minor gaps like explicit node creation/deletion or cycle detection exist, but the server's analytical focus is well covered.