C2B — Claude Second Brain

One graph over everything you know: your Obsidian knowledge vault (the durable why — decisions, traps you already paid for, architecture notes) joined to the code files of every project you map — with live Claude Code activity flowing through it, an offline MCP recall layer, and a 3D brain-shaped visualization.

The point is not the picture. The point is that recall happens before code is read: every Claude Code session starts knowing the brain exists, every prompt is matched against it, and every file edit can be preceded by one call that returns both the code neighbourhood and the human knowledge attached to it.

┌─────────────────────────┐        ┌──────────────────────────────┐
│  Obsidian vault          │        │  Your code projects           │
│  (openclaw structure)    │        │  (any language)               │
│  okf/catalog.json        │        │  graphify extract --code-only │
│  okf/graph.json          │        │  → data/raw/<name>/graph.json │
└───────────┬─────────────┘        └───────────────┬──────────────┘
            │        knowledge pages               │  symbol graphs,
            │        + links + tags                │  collapsed to file level
            └──────────────┬───────────────────────┘
                           ▼
                     merge.py  →  data/brain.json  (+ brain.index.json)
                           │
       ┌───────────────────┼──────────────────────────┐
       ▼                   ▼                          ▼
 mcp_server.py       server.py :8930           Claude Code hooks
 (stdio MCP,         (FastAPI + WebSocket      SessionStart  → primer
  reads brain.json    + React frontend:        UserPromptSubmit → recall
  from disk —         3D brain / 2D network    PostToolUse  → live events
  works offline)      / cortical rings)        (buffered when server down)

What you get

Related MCP server: obsidian-emergent-mcp

The recall protocol

The brain is rung 0 of the recall ladder — above code-graph tools and grep — because one call returns the merged picture:

• About to touch a file? brain_context(file_path) → the matching node, its code neighbours, the vault pages a human wrote about that file (the traps, the decisions), and recent Claude access events.

• Starting a task on a topic? brain_search(topic) — finds the knowledge page and the code files in one shot, across all projects.

• Changing something shared? brain_neighbors(node_id, depth) — the blast radius.

• How do two things relate? brain_path(a, b) — the actual chain between them.

brain_context's vault_pages field is the payload: a non-empty result means someone already paid for a lesson about this exact file. Read the page before editing.

Requirements

• Python 3.11+ and uv

• Node.js 24+ (the unit tests import TypeScript directly via type stripping)

• graphifyy — uv tool install graphifyy (local tree-sitter AST extraction, no LLM, respects .gitignore)

• Claude Code — the hooks and MCP registration target its config

• An Obsidian vault for the knowledge layer — the vault is the memory layer of the brain. It must carry an okf/ bundle (a machine-readable catalog of your pages). The expected structure and the exact JSON contract are documented in docs/vault-structure.md. No vault? Set "vault": null in sources.json and you get a code-only brain — but the knowledge layer is the half that makes recall worth it.

• Windows-first: the refresh scripts are PowerShell. Everything else (Python, Node, hooks) is cross-platform; porting refresh.ps1 to bash is a ten-line exercise.

Quick start

1. Configure your sources. Copy sources.example.json → sources.json, point it at your vault and projects, name your layers.

2. Extract + merge:

uv tool install graphifyy
./refresh.ps1        # graphify extract per source → merge.py → deploy to ~/.claude/c2b

Projects without a git repo need a .graphifyignore (like this repo's) so node_modules/build output stay out of the graph.

3. Register the MCP server (user scope, so it works in every project):

claude mcp add --scope user c2b -- uv run --directory <home>/.claude/c2b python mcp_server.py

4. Wire the hooks into ~/.claude/settings.json:

{
  "hooks": {
    "SessionStart": [{ "hooks": [{ "type": "command", "command": "node \"<home>/.claude/hooks/c2b-session-start.js\"" }] }],
    "UserPromptSubmit": [{ "hooks": [{ "type": "command", "command": "node \"<home>/.claude/hooks/c2b-prompt-hook.js\"" }] }],
    "PostToolUse": [{ "hooks": [{ "type": "command", "command": "node \"<home>/.claude/hooks/c2b-hook.js\"", "async": true }] }]
  },
  "permissions": { "allow": ["mcp__c2b__*"] }
}

The mcp__c2b__* allow rule matters: without it every brain call prompts for permission, which silently kills adoption.

5. (Optional) run the visualization:

cd frontend && npm ci && npm run build && cd ..
uv run uvicorn server:app --port 8930    # serves the built UI + live activity

6. Keep it fresh. Re-run refresh.ps1 after structural code changes, or schedule it (nightly task, or a debounced wrapper fired from a Stop hook). The SessionStart primer warns every session when brain.json goes stale — a stale node is worse than no node.

MCP tools

Node ids are namespaced: vault:<page-id> for knowledge, <layer>:<path> for code.

Offline by design

• The MCP reads brain.json from disk at startup; the :8930 server only enriches recent_access, with a fallback to the persisted events.jsonl.

• The PostToolUse hook buffers events to pending.jsonl (2 MB cap, newest-half kept on overflow) whenever the server is down; the server drains the buffer on start, deduplicating replays.

• A corrupt brain.json never blocks the activity pipeline: the server boots with an empty graph and a loud message, because recording activity is the part that cannot be recovered later.

The hardening behind these guarantees (atomic drains, replay dedup, ambiguous-path refusal, and more) is documented in docs/implementation-notes.md.

Demo mode

npm run demo:data regenerates frontend/public/demo/ from your real data/brain.json through a sanitizer that strips every identity: node ids become n0…n, labels become generic, paths and descriptions are dropped — a unit test proves no private string survives. npm run build:preview + vite preview then serves the full visualization with zero API calls, safe to show anyone. This repo ships with a pre-built demo dataset (~1.6k nodes).

Tests

cd frontend
npm run test:unit    # node:test over the layout/sprite/frame/sanitizer logic
npm run test:e2e     # vite preview + Chrome: views, a11y contract, framing, reduced motion

License

MIT