eidetic-works/nucleus-mcp

> The portable decision log your AI tools all read. One MCP server. Any AI tool. Plain files.
Every AI coding session starts by re-explaining context the last session already knew. .brain is a folder in your repo that Claude Code, Cursor, and Codex all read via one MCP server. Decisions, policies, plans — written once, remembered across every session and every tool.
MIT licensed. File-based (plain JSON + markdown). No embeddings. No vendor lock-in.
---
Also included: nucleus-rabbithole
nucleus-mcp ships a second, fully independent tool: **nucleus-rabbithole**,
a rabbit-hole depth tracker for focus-prone developers.
It gives your AI a push/pop depth stack, a context-switch thrash detector,
an open-loop externaliser, and a weekly review — all backed by local SQLite,
no network, no daemon.
Claude Code .mcp.json snippet:
Full documentation: docs/RABBITHOLE.md — also installed with the package, at /mcp_server_nucleus/docs/RABBITHOLE.md.
---
Three Frontiers
The core loop that makes AI reliability compound over time:
**GROUND** — 5-tier execution verification. Syntax, imports, tests, runtime. Goes outside the formal system to check the AI's work.
**ALIGN** — One-call corrections. nucleus review records a verdict against a prior decision, which becomes a training pair and an event.
**COMPOUND** — Deltas measure the gap between intent and reality. Recurring patterns become strategy. Negative deltas become training signal.
Every tool response shows frontier health:
---
Quick Start
**Hosted access is not currently available.** The remote MCP endpoint is
withdrawn pending an access-control review. Use the local install below.
**Local install (Cursor, Windsurf, Claude Desktop):**
Two commands. Nucleus is running. nucleus init seeds .brain/, writes a project-local .mcp.json, and offers to configure every AI client you have installed. Say yes, then restart the client.
nucleus build additionally needs a coding-agent CLI on PATH — any one of claude (npm i -g @anthropic-ai/claude-code), devin, or agy, already authenticated (claude/devin/agy login once, separately). Without one, build stops at the plan stage. If build reports status='error' without a clear cause, the vendor CLI's own error (e.g. an expired login) lands in .brain/relay/cross_vendor/*.json — check the newest file there.
---
What It Does
**30 MCP tools** across 16 facades, exposing **210 actions**, plus 19 brain://
resources and 8 prompts. Regenerate these with python3 tools_audit/count_surface.py
rather than editing them by hand — three hand-maintained copies of this number had
drifted apart before that script existed.
- **GROUND** — Execution verification (5 tiers: diff, syntax, imports, tests, runtime)
- **ALIGN** — Human corrections (verdict + delta + DPO + event in one call)
- **Memory** — Engrams that persist across sessions. Write once, recall forever.
- **Sessions** — Save context, resume later. Session arc shows your last 3 sessions.
- **Tasks** — Priority queue with escalation, HITL gates, and heartbeat monitoring.
- **Governance** — Kill switch, compliance configs (EU DORA, MAS TRM, SOC2), audit trails.
- **Orchestration** — Agent slots, multi-brain sync, task dispatch.
- **Archive** — Training pipeline (SFT + DPO), delta tracking, frontier health dashboard.
**Benchmark:** decision-retention-evals — does your AI agent remember why the code is the way it is? (ask on Discord for the current mirror)
---
Nucleus Pro
Everything above is free (MIT). Nucleus Pro adds verifiable governance:
**$19/month** or **$149/year** — nucleusos.dev/pricing
| | Free | Pro |
|---|---|---|
| 30 tools, 19 resources, 8 prompts | Yes | Yes |
| Persistent memory | Yes | Yes |
| Governance & HITL | Yes | Yes |
| Audit trails (DSoR) | Yes | Yes |