Rune

The Software Intelligence Runtime.

Cursor, Codex, Claude Code — they all read your codebase cold, every session, and forget everything when it ends. Rune is the thing that runs underneath them, continuously, so every session starts already knowing your codebase instead of re-deriving it.

Without Rune:

flowchart LR
    A1["New AI session"] --> A2["Reads files cold"]
    A2 --> A3["Reasons, answers"]
    A3 --> A4["Session ends"]
    A4 -.->|"understanding lost"| A1

Loading

With Rune:

flowchart LR
    B1["Code changes"] --> B2["rune watch rebuilds automatically"]
    B2 --> B3["Understanding always current"]
    B3 --> B4["New AI session"]
    B4 --> B5["Queries Rune over MCP"]
    B5 --> B6["Evidence-backed answer, cited to file + line"]
    B6 -.->|"stays current"| B3

Loading

What it does · How it works · Install · Quickstart · See real output · CLI · MCP tools · Limitations · Security

---

What Rune does

• Always current — rune watch keeps the understanding graph up to date as you edit, not just when someone remembers to re-run a scan. It's a runtime, not a one-shot CLI.
• Persistent — understanding survives past any single session. Every AI you connect starts already knowing your software, instead of re-reading it cold.
• Shared — one understanding, many AI clients. Claude, GPT, Gemini, whatever you're running next month — they all ask the same Rune instance instead of maintaining separate, inconsistent mental models of your code.
• Explainable — nothing Rune tells an AI is a black box. Ask it to justify any claim and it shows the exact source location behind it.
• Read-only, suggest-only — Rune never writes to your source. That's the actual moat: everything else in this space is racing toward autonomous code-writing; Rune stays strictly on the reading, understanding, and (later) suggesting side of that line. It never becomes the thing you have to double-check for silently breaking your code.

How it works

Two layers, and every conclusion traces back to evidence — not asserted, always inspectable:

flowchart LR
    F1["Fact: import react, UserCard.jsx line 1"] --> D["Derived: React component UserCard"]
    F2["Fact: function UserCard, UserCard.jsx line 3"] --> D
    F3["Fact: useState hook, UserCard.jsx line 4"] --> D
    D --> C["AI answer: this project has a UserCard component"]
    C -.->|"rune_explain"| F1
    C -.->|"rune_explain"| F2
    C -.->|"rune_explain"| F3

Loading

Call rune explain <id> (or the rune_explain MCP tool) on anything and get that chain back. An AI using Rune isn't guessing about your architecture — and neither is Rune.

Install

npm install @pypl100/rune

A Python distribution exists at python/ with full feature parity (same fact/derived model, same detectors, same MCP tool surface) for teams who'd rather not require Node.js — published as pip install north-rune (the CLI command is still rune; see python/README.md).

If you install both (@pypl100/rune via npm globally, and north-rune via pip) on the same machine, only one rune command will actually be on your PATH — whichever your shell finds first, not whichever you installed most recently. Check with which -a rune; if it lists more than one path, that's why. There's no version-detection magic here — it's plain OS PATH resolution, same as any two unrelated tools that happen to install a same-named binary. If you need a specific one, invoke it by its full path rather than relying on bare rune.

Quickstart

cd your-project
npm install -g @pypl100/rune   # or: npx -p @pypl100/rune rune <command>
rune init      # sets up Rune in your project
rune watch &    # keeps the understanding current in the background as you work
rune serve      # starts an MCP server exposing it to any AI client

rune watch is the recommended default — it's what makes Rune a runtime instead of a tool you have to remember to re-run. (For CI or a one-shot check, use rune scan instead.)

sequenceDiagram
    participant Dev as You
    participant Watch as rune watch
    participant Graph as graph file
    participant AI as AI client (Claude Code, Cursor, etc.)

    Dev->>Watch: save a file
    Watch->>Watch: detect change (debounced)
    Watch->>Graph: rebuild + write
    AI->>Graph: rune_search / rune_explain (via rune serve)
    Graph-->>AI: current, evidence-backed answer

Loading

Point any MCP-compatible client (Claude Desktop, Claude Code, custom agents, etc.) at the rune serve process. Every connected AI shares the same, continuously current understanding — no restart, no manual rescan.

Example MCP client config entry:

{
  "mcpServers": {
    "rune": {
      "command": "npx",
      "args": ["-p", "@pypl100/rune", "rune", "serve", "/absolute/path/to/your-project"]
    }
  }
}

What this actually looks like

Real output, from the fixture project in this repo — not staged:

$ rune scan .
[rune] scanned 4 file(s) in 34ms
[rune] facts: 10, derived conclusions: 4
[rune] graph written to /project/.rune/graph.json

$ rune explain component_2
{
  "kind": "function",
  "id": "component_2",
  "type": "react_component",
  "file": "components/UserCard.jsx",
  "line": 3,
  "name": "UserCard",
  "evidence": "export function UserCard({ user }) {"
}

That's the whole trust model in one example: Rune doesn't just say "there's a UserCard component" — it points at the exact file, the exact line, and the exact text it matched. Ask it to explain anything, and you get the receipts, not a claim.

CLI

Command	What it does
rune init [dir]	Sets up .rune/ in the current (or given) project
rune scan [dir]	Builds (or rebuilds) the understanding graph, once
rune watch [dir]	Keeps the understanding graph current as files change (Ctrl+C to stop)
rune serve [dir]	Starts the MCP server
rune explain <id>	Prints the evidence trail behind any fact or conclusion
rune --version	Prints the installed Rune version

Configuration

rune init writes .rune/config.json. The only setting today is ignore — an array of directory/file names to exclude from scanning, on top of the built-in defaults (node_modules, .git, dist, build, .next, coverage, etc., and all dotfiles unconditionally):

{
  "ignore": ["legacy", "vendor"],
  "version": 1
}

MCP tools exposed

Tool	Purpose
rune_get_overview	Architecture summary — start here
rune_list_components	All detected React components
rune_list_routes	Unified Express + Next.js route list
rune_search	Find facts/derived nodes by name, file, or route substring
rune_explain	Full evidence trail for any id
rune_get_file_dependencies	Internal import graph for a file
rune_rescan	Re-scan on demand after code changes

Current scope and honest limitations

Rune is intentionally narrow right now:

• Framework support: React, Next.js (pages + app router), Express. Everything else gets generic file/import scanning only.
• Extraction method: heuristic, regex-based pattern matching — not a full AST parser. This keeps the scanner dependency-free and fast, and every fact still carries file/line/matched-text evidence, but it will miss unusual code shapes (e.g. components returned via React.createElement with no JSX, dynamically constructed route strings, deeply re-exported components). A real AST-based extractor is the natural next upgrade — the fact schema is designed so extraction method can be swapped without touching anything downstream.
• No cross-file route-prefix resolution: an Express route mounted via app.use('/api', router) in one file and defined in another isn't stitched into a single path yet.
• Read-only: Rune never writes to your source files. It only ever writes its own graph to .rune/graph.json.
• Single-process, stdio MCP transport — no multi-client daemon yet.

Security notes

• Rune never scans dotfiles or dot-directories (.env, .env.*.js, .git, .ssh, editor configs, etc.), with no exceptions. This is enforced in the scanner and covered by a regression test — a config file with a matching code extension (e.g. .env.js) will not have its contents read or embedded as evidence.
• Symlinks are not traversed, so a symlink pointing outside the project root can't pull external files into the scan.
• .rune/graph.json is excluded from git by rune init (it creates .gitignore if one doesn't already exist).
• The graph itself is the only thing Rune writes. If you share .rune/graph.json with an AI client, you're sharing everything in it — treat it like any other file that quotes snippets of your source.

Verifying the MCP server

The unit test suite (npm test) covers scanning and the understanding graph, but doesn't spin up a real MCP client. scripts/verify-mcp-server.mjs does: it spawns rune serve as a real child process and drives it through the actual JSON-RPC handshake (initialize → notifications/initialized → tools/list → tools/call), checks that all expected tools are exposed, and confirms a genuinely invalid call is rejected cleanly rather than crashing the server.

npm install
npm run verify:mcp
# or against a real project instead of the bundled fixture:
npm run verify:mcp -- /path/to/your-project

Roadmap

• AST-based extraction (swap-in replacement for the regex scanner)
• Cross-file route resolution
• Data-flow tracing between frontend calls and backend routes
• True incremental re-scan — rune watch exists now, but it still does a full rebuild on every change, not a diff of just what changed. Fine for small-to-medium projects; will matter on very large ones.
• Suggestion tools (e.g. flagging a known-vulnerable dependency pattern with evidence, proposing a reviewable fix) — strictly additive to the read-only model, never auto-applied

License

MIT

---

If you want to support the circus: ETH 0xbc0979dde621c353737d21f6d7b4eb361f7bc11f