CodeBrain

An MCP server that lets Claude Code offload bulk work to a local LLM running on your own hardware.

What this is (and isn't)

Is: A Model Context Protocol (MCP) server that Claude Code registers as a sub-agent backend. When a session includes the kind of task a 14B local coder model handles well — generating 50 event templates, polishing 20 React components, drafting boilerplate — Claude Code calls into CodeBrain instead of spending its own output tokens. The local model does the bulk draft, Claude reviews and applies.

Is not: A Claude replacement. The reasoning, architecture decisions, debugging, and anything where "close enough" isn't good enough stays with Claude. CodeBrain is a Claude-offloader, not a Claude-competitor.

Why: Large-volume content and polish work burns through Claude's context and rate limits fast. A local model you can run unlimited costs nothing extra per call and keeps the high-value context free for the hard parts of the session.

Status

Phases 1–4 complete, Phase 5 deferred. Nine tools exposed, .brain/context.md passthrough live, per-file brain summaries scanner, verifier loop, consensus decoding. MCP integration verified in a real Claude Code session. Phase 5 (RAG) was explicitly scoped as "only if needed" and current use doesn't show cross-file search as a bottleneck, so it stays deferred.

How it works

Claude Code session                     CodeBrain MCP server              Local machine
─────────────────────      stdio       ───────────────────                ─────────────
Claude delegates a         ────────►   codebrain_generate()     ────►    Ollama HTTP
bulk / polish task                     codebrain_explain()                (localhost:11434)
                                       codebrain_status()                      │
                                                                                ▼
                                                                        Qwen2.5-Coder 14B
                                                                              (GPU)
Claude reviews,            ◄────────   tool result string        ◄────    streamed response
applies, or pushes back

Nine tools are exposed today:

The use_brain flag on generation tools automatically prepends .brain/context.md from the current working directory to the system prompt, so project-specific context travels with every call without Claude having to pass it manually.

Requirements

• Python 3.11+

• Ollama — download for your OS. Tested with Ollama on Windows native, talking over localhost:11434.

• A coder model pulled locally:

  ollama pull qwen2.5-coder:14b

  ~9 GB download. Fits in 12 GB VRAM at Q5. Other models work too (DeepSeek-Coder, Qwen3 if available) — set via CODEBRAIN_MODEL env var.

• Claude Code CLI on the machine that will call the server (obviously).

Install

git clone <this repo> CodeBrain
cd CodeBrain
python -m venv .venv
.venv\Scripts\activate                         # on Windows
# source .venv/bin/activate                    # on macOS / Linux
pip install -e .

Configure Claude Code

Add CodeBrain to your Claude Code MCP config. On Windows, that's usually ~/.claude.json (adjust path to where you cloned):

{
  "mcpServers": {
    "codebrain": {
      "command": "C:\\Users\\YOU\\Desktop\\CodeBrain\\.venv\\Scripts\\python.exe",
      "args": ["-m", "codebrain"]
    }
  }
}

Restart any Claude Code session — the five codebrain_* tools should now appear in the available-tools list.

Keep brain files in sync automatically

Once you've run codebrain_init on a repo and scanned it with codebrain_scan_repo, you probably want brain files to refresh automatically whenever Claude edits source. Two pieces wire that up:

1. Project CLAUDE.md snippet — tell Claude to read brain files before opening source:

## Brain files

This repo has per-file `.brain` summaries next to each source file.
Before reading a full source file, read its `<path>.brain` sibling first.
Only open the source when the brain file is insufficient for the task.

2. PostToolUse hook — regenerate the brain after every Edit/Write.

Add to .claude/settings.json in the repo root:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "python -c \"import asyncio, json, sys; from codebrain.brain_scanner import scan_file; d = json.load(sys.stdin); p = d.get('tool_input', {}).get('file_path'); p and p.endswith(('.py', '.ts', '.tsx', '.js', '.jsx', '.java', '.go', '.rs')) and print(asyncio.run(scan_file(p)))\""
          }
        ]
      }
    ]
  }
}

The hook inspects the edited path, skips non-source files via the extension filter, and kicks off a scan. Hash-gated: unchanged files don't hit Qwen.

Sanity check

Inside a Claude Code session, ask Claude:

Call codebrain_status and tell me what's installed.

If Ollama is running and the model is pulled, you'll get back qwen2.5-coder:14b in the list.

Configuration

Environment variables read by the backend:

Project structure

CodeBrain/
├── codebrain/
│   ├── __init__.py
│   ├── __main__.py            # `python -m codebrain` entry
│   ├── backend.py             # Ollama HTTP client
│   ├── server.py              # FastMCP server + tool definitions
│   ├── brain_scanner.py       # scan_file / scan_repo + hash gate
│   ├── brain_init.py          # one-shot .brain/context.md seeding
│   ├── verifier.py            # deterministic output checks
│   └── prompts/
│       └── brain_few_shot.md  # few-shot for brain-file generation
├── tests/                     # 96 unit + integration tests
├── .spec/
│   ├── CURRENT.md             # phase state
│   └── brain-file-format.md   # brain-file format v1
├── pyproject.toml
├── LICENSE
└── README.md

Roadmap

Phase 1 — scaffold ✓

• Ollama HTTP client with error handling

• FastMCP server with stdio transport

• Three core tools: generate, explain, status

• Documented setup + Claude Code config

• Verified in a real Claude Code session

Phase 2 — batch & context ✓

• codebrain_batch_generate for mass content with one shared system prompt, index-stable errors

• codebrain_polish for targeted transforms (shorten / rephrase / translate) instead of regeneration

• .brain/context.md passthrough — cwd project context auto-prepended to every generation call

• Dogfood: coding tasks solid, text-transform tasks revealed real limits (informs Phase 3)

Phase 2.5 — brain system ✓

Per-file <source>.brain summaries sit next to each source file. Claude reads the brain first and only opens the source when the brain is insufficient.

• codebrain_scan_file(path, force) — generate or refresh one brain file

• codebrain_scan_repo(root, force, extensions, exclude_dirs) — bulk walk + scan

• codebrain_init(root, force) — seed .brain/context.md with stack detection

• Hash-gated regeneration (SHA256) — idempotent reruns

• Programmatic frontmatter — deterministic source, source_hash, model; Qwen only writes the five sections

• Defense-in-depth validation: fence-strip, skip-empty-sources (<10 chars), section-presence/order, retry-on-invalid

• CLAUDE.md convention + PostToolUse hook snippet in this README

Phase 3 — VERIFIER loop ✓

Dogfood showed the local model drifts on text transforms. The verifier catches no-ops, length violations, and schema misses deterministically before they reach Claude.

• detect_noop — whitespace-normalised equality check (auto-retries inside codebrain_polish)

• check_word_count(min_words, max_words) — bounded-window gate

• check_regex_schema(pattern) — structured-output check

• codebrain_generate_verified(prompt, min_words, max_words, must_match, max_retries) — loop with tightened retry instructions, returns [codebrain warning] ... if verification fails after retries

Phase 4 — consensus decoding ✓

• codebrain_consensus_generate(prompt, n) — generate N candidates (clamped to [2,5]), Qwen picks the best verbatim. N+1 inference calls, tightens quality on high-variance tasks.

• Multi-pass skeleton→logic→edges→polish: deferred (low measured value; individual tools already compose).

Phase 5 — RAG (deferred — not a bottleneck)

Brain files already act as an index; cross-file RAG only makes sense if future use actually shows that indexing is the blocker. No current signal for it, so not built.

License

MIT — see LICENSE.