Neuromcp

neuromcp — Sovereign Memory for AI agents

Any model. Your memory. Stays local.

neuromcp is the first Sovereign Memory layer for AI: an open-source MCP server that gives Claude, GPT, Gemini, and Ollama persistent, searchable memory — stored entirely on your machine. No API keys. No cloud sync. No subscription required to remember who you are.

Sovereign Memory = data that you own outright, lives on hardware you control, and is portable across every model you use. Cloud memory products own your data; Sovereign Memory means you do.

npx neuromcp-init   # one command: detects your MCP clients, writes configs, sets up the wiki

Or run the bare server without any setup: npx neuromcp. Something not working? npx neuromcp-doctor diagnoses the daemon, Ollama, embeddings and the database in one run.

The built-in memory browser: the most-connected entities in a namespace and the relations between them, as a force-directed graph you can drag, zoom and click — served from localhost, never from a cloud. See Memory browser & Obsidian.

Why neuromcp

The LLM is a commodity. Your memory is the moat. GPT-5, Claude 4, Gemini — they all converge. The model you use next year will differ. The memory of every conversation, decision, and preference you build is yours. neuromcp keeps that layer on your machine and makes it portable across any MCP-compatible client.

Local-first is a design choice, not a limitation. No telemetry. No data leaves your laptop. No vendor has a copy of your conversations. Audit every line of code that touches your memory. SQLite + local embeddings; everything fits on one disk.

One install. Every client. Claude Desktop, Cursor, Windsurf, Codex CLI, Continue, LibreChat, Open WebUI — neuromcp speaks MCP, so it works wherever MCP is supported. Switch models tomorrow; your memory follows.

Real recall, not keyword matching. Hybrid retrieval combines vector search (nomic-embed-text, 768-dim), BM25 full-text, graph links, and a learned usefulness prior. At 500 distractors on LongMemEval, R@5 holds at 93.3%. Your context window gets the right memory, not just the most recent.

LongMemEval-S accuracy

Repro: OMB_ANSWER_LLM=claude OMB_ANSWER_MODEL=opus OMB_JUDGE_LLM=claude OMB_JUDGE_MODEL=opus uv run omb run --dataset longmemeval -s s -m neuromcp -c "single-session-user,single-session-assistant,multi-session,temporal-reasoning,knowledge-update,single-session-preference" --query-limit 17

Sample size honesty. n=102 (17 per category × 6 categories). Wilson 95% CI for 98/102 ≈ 90.5–98.7%. Full 500q run with the same config is the next milestone before any "top-tier" claim.

Benchmarks (v0.18.0)

Oracle split (clean — easy mode)

Oracle-split LongMemEval isolates the correct memory in a small corpus. Every local MCP memory system claims ~99% here. It measures "does the ranker work on clean inputs" — nothing more.

Distractor split (v0.18.0, honest)

Same 30 questions + 1000 random distractor memories drawn from other questions' haystacks. The correct memory now competes against real noise.

Reproduce: npx tsx eval/longmemeval-distractor-runner.ts --limit 5 --distractors 1000

Sample sizes. The 500-distractor row is n=30 (Wilson 95% CI for 28/30 ≈ 78-99% R@5). The 1000-distractor row is n=5 — preliminary, Wilson 95% CI [57%, 100%]. The 1000-distractor n=30 run takes ~36 min on a single Ollama instance; cached-distractor batching is v0.19.0 work. Treat 500-distractor numbers as defensible, 1000-distractor as directionally positive but underpowered.

Head-to-head comparison is explicit v0.19.0 work. Hindsight (local OSS MCP, ~94.6% LongMemEval claimed) and Mem0/Zep publish their own numbers on their own harnesses. Until we run all of them against the same corpus + embedder, calling any local MCP server "state of the art" is marketing, not measurement. neuromcp publishes its numbers with sample-size caveats so you can judge direction; don't read absolute superiority into them yet.

Hybrid ranker (BM25 + vector + attention + graph + usefulness prior) keeps R@5 = 100% at 1000:1 distractor:target ratio on the observed sample. MRR drops to 74% because the correct memory is sometimes not rank-1 but always rank ≤ 5 in what we saw. Earlier v0.18.0 numbers (R@5 23%) were from a test FakeEmbedder — fixed in v0.18.1.

What this benchmark does NOT prove: end-to-end answer correctness, long-horizon multi-session reasoning, or superiority over commercial cloud systems (Mem0, Zep) on their own benchmarks. Those comparisons need their numbers on the same distractor split, which hasn't been published.

Why

AI agents forget everything between sessions. Existing solutions either store flat key-value pairs (useless for real knowledge) or require cloud infrastructure and API keys.

neuromcp gives you two layers of memory:

1. MCP Server — hybrid search (vector + full-text + graph), verbatim recall, memory governance, automatic consolidation, all in a single SQLite file
2. Wiki Knowledge Base — compiled Markdown knowledge that survives crashes, compounds over sessions, and gives your agent project-aware context at every startup

Inspired by Karpathy's LLM Wiki, Mastra's Observational Memory, and Zep's temporal knowledge graphs — but simpler than all of them. No vector DB, no embeddings pipeline, no cloud. Just Markdown files + Git + hooks.

Architecture

~/.neuromcp/
├── memory.db               ← SQLite: hybrid search, MCP tools
├── wiki/                   ← Compiled knowledge (git-tracked)
│   ├── index.md            ← Routekaart — LLM reads this FIRST
│   ├── schema.md           ← Operating rules for the LLM
│   ├── log.md              ← Append-only changelog
│   ├── people/             ← User profiles, preferences
│   ├── projects/           ← Project knowledge (stack, auth, URLs)
│   ├── systems/            ← Infrastructure (tools, MCP servers)
│   ├── patterns/           ← Reusable patterns (error fixes, routing)
│   ├── decisions/          ← Architecture decisions with context
│   └── skills/             ← Repeatable procedures
└── raw/sessions/           ← Raw session logs (auto-generated)

How the wiki works

Self-healing consolidation pipeline (v0.15.0+)

Every ~4h the launchd agent runs run-consolidation.sh, which orchestrates four steps end-to-end:

1. consolidate-sessions.py — batches raw sessions per project, asks Claude for a factual summary, and fact-checks it against the raw sources. When the auditor flags specific unsupported claims the consolidator now auto-strips those lines and re-audits once — so one speculative sentence no longer kills a whole batch.
2. rescue-rejected.py — any batch that still fails is parsed, the unsupported claims are removed, and the cleaned summary is appended to its wiki page. Pure text surgery, no LLM calls.
3. entity-linker.py — cross-links every page: a bare-word mention of another registered entity (people/, projects/, systems/) is added to the page's related: frontmatter. Makes the wiki act like a graph without a separate graph database.
4. rebuild-index.py — regenerates index.md and per-category -index.md files. Categories over 10 pages are auto-split so the session-start router stays compact as the wiki scales.

The pipeline is idempotent — safe to re-run at any time.

What the LLM knows at session start

Schema (operating rules) → How to maintain the wiki
Index (knowledge map)    → What knowledge exists
User profile             → Who you are, how you work
Project page             → Current project details (auto-detected from cwd)
Last session             → What happened last time

Quick Start

One command (recommended)

npx neuromcp-init

Detects your installed MCP clients (Claude Desktop, Claude Code, Cursor, Windsurf), writes the neuromcp entry into each config (with a backup of the original), initializes the wiki + hooks, and checks whether Ollama is available. --dry-run previews everything without writing.

For GUI clients (Claude Desktop, Codex Desktop) install permanently first: npm install -g neuromcp && neuromcp-init. Only a permanent install lets init write the absolute node + script paths those clients need — from the bare-npx cache (garbage-collected) it deliberately falls back to a PATH-dependent npx entry, which GUI clients usually cannot resolve.

Manual steps (what init does under the hood)

1. Start the MCP server

npx neuromcp

2. Initialize the wiki + hooks (required for closed-loop attribution)

npx neuromcp-init-wiki

This creates the wiki structure, installs hooks (Claude Code) and rules (other editors), and configures everything automatically. Without this step, npx neuromcp still runs as a plain MCP server with 46 tools, but the critic hook that closes the attribution loop is not installed — retrieval works but usefulness scores never accumulate. Safe to run multiple times — won't overwrite existing config.

Editor Compatibility

neuromcp works with any MCP-compatible editor. Two tiers of integration:

Claude Code gets the full experience via native hooks — context injection and persistence happen automatically, even if the LLM forgets.

Other editors get rules files that instruct the LLM to call neuromcp tools at session start/end. This depends on LLM compliance — it works well in practice but is not guaranteed like hooks.

# Auto-detect installed editors
npx neuromcp-init-wiki

# Target a specific editor
npx neuromcp-init-wiki --editor cursor

# Install rules for all supported editors
npx neuromcp-init-wiki --editor all

Supported editors: cursor, windsurf, cline, copilot (VS Code), jetbrains, zed

Recommended: Add Ollama for real semantic search

ollama pull nomic-embed-text

neuromcp auto-detects it. No config needed.

Non-English memories: use a multilingual embedder

nomic-embed-text is English-centric. On a real Dutch memory store (2,402 memories, 40 paraphrased Dutch queries, 15 off-topic queries) it found the right memory in the top 5 only 2.5% of the time, and off-topic queries scored higher than true matches. bge-m3 (multilingual, 1024-dim) on the same data:

To switch an existing install:

ollama pull bge-m3
npx neuromcp-reembed                 # dry run on a copy
npx neuromcp-reembed --apply         # rebuild the index (keeps a backup)

Run the reembed with NEUROMCP_EMBEDDING_MODEL=bge-m3 set. After that every client on the default auto setting follows the model the index was built with, so nothing else needs configuring. With bge-m3, search_memory also applies a calibrated relevance floor (0.45) so off-topic queries return nothing instead of noise; keyword (FTS) and knowledge-graph hits always pass.

Keep NEUROMCP_RERANKER=none for non-English memories: the bundled cross-encoder (ms-marco-MiniLM-L-6-v2) is English-only and dropped MRR on the Dutch benchmark above from 0.46 to 0.18.

Upgrading

npm install -g neuromcp@latest
npx neuromcp-init-wiki        # updates installed hooks; safe to re-run
npx neuromcp-doctor check     # confirms hooks, daemon and embeddings

neuromcp-init-wiki copies the Claude Code hooks once; until 0.30.0 it never touched them again, so hook fixes did not reach existing installs. Re-running it now updates every hook that is an unmodified older neuromcp version (a backup is kept next to each one), migrates old .js hooks to .cjs (leaving a forwarding stub at the old path) and leaves hooks you edited alone — it lists them, and --update-hooks replaces those too. neuromcp-doctor check reports outdated hooks.

Installation

The reliable way is npm install -g neuromcp && neuromcp-init — it writes the entry below into every detected client config (with a backup first). Run init from the permanent install, not via bare npx: from the npx cache it deliberately falls back to a PATH-dependent entry that GUI clients usually cannot resolve. The manual form is:

Claude Code

// ~/.claude.json → mcpServers
{
  "neuromcp": {
    "type": "stdio",
    "command": "npx",
    "args": ["-y", "neuromcp"]
  }
}

npx is fine here because Claude Code runs from your terminal and inherits your PATH.

Claude Desktop / Codex Desktop / any GUI client

GUI apps are started by the window manager, not by a login shell, so they inherit a minimal PATH that usually has neither node nor npx (nvm, fnm, Volta and Homebrew all install outside /usr/bin). Use absolute paths — this is exactly what neuromcp-init writes since 0.29.3:

// ~/Library/Application Support/Claude/claude_desktop_config.json
{
  "mcpServers": {
    "neuromcp": {
      "command": "/opt/homebrew/opt/node@22/bin/node",
      "args": ["/opt/homebrew/lib/node_modules/neuromcp/bin/neuromcp.mjs"]
    }
  }
}

Find your own two paths with:

which node                      # → "command"
npm root -g                     # → "args": ["<that>/neuromcp/bin/neuromcp.mjs"]

(bin/neuromcp.mjs keeps its #!/usr/bin/env node shebang, so running it straight from a terminal still works.)

Cursor / Windsurf / Cline

Same format — add to your editor's MCP settings. Copy-paste configs for every client live in examples/.

Shared daemon (recommended when you run multiple clients)

By default each client spawns its own neuromcp process. One shared background daemon serves them all instead — one database connection, one embedding pipeline, no cold start per client:

npx neuromcp-enable-daemon --port 3200   # macOS launchd agent; verify with: curl -s http://127.0.0.1:3200/health

Then point clients at the daemon:

// Claude Code (~/.claude.json) — native HTTP transport
{ "neuromcp": { "type": "http", "url": "http://127.0.0.1:3200/mcp" } }

// Claude Desktop — stdio-only, bridge via neuromcp-connect.
// The bridge waits for the daemon on cold boot (plain mcp-remote exits
// fatally when the client starts before the daemon has bound its port,
// leaving a permanent "Server disconnected").
{ "neuromcp": { "command": "npx", "args": ["-y", "--package=neuromcp", "neuromcp-connect", "http://127.0.0.1:3200/mcp"] } }

The daemon binds loopback only, rejects non-allowlisted Host and Origin headers (DNS-rebinding defense), and is unauthenticated by design inside that boundary. Uninstall: npx neuromcp-enable-daemon --uninstall.

Platform support

Per-project isolation

// .mcp.json in project root
{
  "mcpServers": {
    "neuromcp": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "neuromcp"],
      "env": {
        "NEUROMCP_DB_PATH": ".neuromcp/memory.db",
        "NEUROMCP_NAMESPACE": "my-project"
      }
    }
  }
}

Memory browser & Obsidian

Memory you cannot see is memory you cannot trust. neuromcp ships two ways to look at what your agents actually know — both local, both read-only.

Built-in web view

When the shared daemon is running, open http://127.0.0.1:<port>/ui (default port 3200):

• Graph — the top entities in a namespace and the relations between them, as a force-directed canvas. Drag nodes, scroll to zoom, click a node to list the memories linked to it (current memories only; superseded ones are hidden unless you ask for them).
• Timeline — type a topic and see how the facts about it evolved over time, including which memory superseded which.

The UI is a single page with no external assets (strict CSP, no CDN), and it is only served on loopback: bind the daemon to a non-loopback host and the /ui and /api/* routes are disabled entirely.

Obsidian graph of the wiki

The wiki (~/.neuromcp/wiki) is plain Markdown, so it opens directly as an Obsidian vault. Obsidian draws its graph from [[wikilinks]] while neuromcp keeps relationships in frontmatter, so run the bridge once (idempotent, --dry-run supported):

npx neuromcp-obsidian-bridge

neuromcp-init-wiki also seeds a minimal wiki/.obsidian/ (colour groups per page type, no plugins) — only if none exists; an existing vault setup is never touched. Details: docs/OBSIDIAN.md.

MCP Surface

46 tools across 8 families — the full auto-generated reference with every parameter lives in docs/TOOLS.md (regenerated from the actual registrations on every change; CI fails when it drifts).

Picking the right retrieval tool: search_memory returns ranked raw memories; recall_answer synthesizes a cited extractive answer (or honestly says not_in_memory); recall_memory is a plain ID/filter lookup with no semantics; search_all adds the verbatim store to the sweep.

Resources (13)

Prompts (3)

Wiki Knowledge Base

The wiki is the compiled, human-readable knowledge layer. It replaces the chaos of session logs with structured, interlinked Markdown pages.

Why a wiki instead of more vector search?

Wiki page format

---
title: My Project
type: project
created: 2026-04-06
updated: 2026-04-06
confidence: high
related: [other-project, oauth-setup]
---

# My Project

Description, stack, auth, deployment details...

How to use

The wiki works automatically once hooks are installed. The LLM:

1. Reads index.md at session start to know what knowledge exists
2. Reads specific pages when relevant to the current task
3. Updates pages when learning something new
4. Gets reminded every 8 tool calls if the wiki needs updating

You can also browse and edit the wiki manually — it's just Markdown files.

Auto-consolidation (optional)

Once you accumulate raw session logs, the wiki can be kept fresh automatically. A scheduled job reads unprocessed sessions, groups them per project (by detecting $HOME/projects/<name> paths in the session content), and uses the claude CLI to synthesise a ## [date] entry into the right wiki page.

npx neuromcp-enable-consolidation

What it installs:

• ~/.neuromcp/scripts/consolidate-sessions.py — the worker
• ~/.neuromcp/scripts/run-consolidation.sh — threshold-guarded runner
• macOS: a launchd agent that fires every 4 hours (com.neuromcp.consolidate)
• Linux: prints a cron snippet to add manually

Requirements:

• python3 ≥ 3.8 on PATH
• the claude CLI on PATH

Guards built in:

• Threshold: skip if fewer than 5 unprocessed sessions
• Output is extracted from a fenced markdown block; apology/narration text is rejected
• Ledger (~/.neuromcp/consolidation-ledger.json) makes re-runs idempotent
• Large project backlogs are auto-batched (default 15 sessions per claude call; override with --max-sessions)

Uninstall: npx neuromcp-enable-consolidation --uninstall

Change interval: npx neuromcp-enable-consolidation --interval 7200 (every 2 hours)

Hallucination guard (eval-loop). Every consolidator output goes through a second Haiku audit before the wiki is touched. If any factual claim in the generated summary is not traceable to the raw sessions, the chunk goes to ~/.neuromcp/review-queue/ instead of the wiki. No hallucinated claims leak through.

Atomic facts with temporal supersession. After a summary is approved, it is also distilled into short standalone facts and stored as category='fact' rows with valid_from=today. When a new fact is Jaccard-similar to an existing one in the same project, Haiku decides whether NEW supersedes OLD — if yes, the old row gets superseded_by_id and valid_to set. Retrieval defaults to current facts only (superseded_by_id IS NULL), so outdated conclusions never resurface.

Auto-retrieve + hybrid indexing

Once the wiki has content, make it searchable so the UserPromptSubmit hook can surface relevant pages automatically (no more "LLM must remember to call search"):

npx neuromcp-index-wiki              # index wiki pages into memories_fts + memories_vec
npx neuromcp-index-wiki --rebuild    # wipe wiki entries first, then reindex
npx neuromcp-index-wiki --dry-run    # preview what would change
npx neuromcp-index-wiki --no-embed   # FTS-only mode (no embedding provider needed)

npx neuromcp-backfill-embeddings     # embed any memory still missing a vector

The indexer splits each page on ## section headers and stores every section as a deduplicated memory (source='wiki', category='wiki'). Each section is both written to the FTS5 index and embedded via the configured provider (Ollama → OpenAI → ONNX) so vector search works too.

At prompt time the neuromcp-auto-retrieve.js hook calls neuromcp-query, which runs FTS5 BM25 and sqlite-vec cosine search in parallel and fuses the rankings via Reciprocal Rank Fusion (k=60). The top-3 merged results are injected as <neuromcp-recall> context.

The hook is installed automatically by neuromcp-init-wiki and registered under UserPromptSubmit in Claude Code's settings.json. Re-run the indexer after large wiki updates (or schedule it — it's idempotent).

Tuning:

Known upstream issues

memories_vec does not reclaim space after DELETE — sqlite-vec #54 / #265. When you re-index after editing wiki sections, the old vector rows are marked deleted but their storage stays. The database file grows monotonically until you run npx neuromcp-index-wiki --rebuild, which drops and re-creates the vector rows. Run a rebuild every few weeks if you edit the wiki heavily.

claude CLI streaming hangs from non-TTY subprocesses on macOS — if you script interactions with claude -p from another process (e.g. scheduled jobs), pipe it through script -q /dev/null to allocate a pseudo-TTY. Without that the stdout buffer never flushes. We work around this inside the consolidator where needed.

Memory Governance

Namespaces isolate memories by project, agent, or domain.

Trust levels (high, medium, low, unverified) rank search results and control decay resistance.

Soft delete tombstones memories — recoverable for 30 days.

Content hashing (SHA-256) deduplicates at write time.

Lineage tracking records source, project ID, and agent ID per memory.

Configuration

All via environment variables. Defaults work for most setups.

What's new

Full history in CHANGELOG.md. Recent highlights:

• v0.29 — current-validity everywhere: superseded memories are hidden from every read path by default (include_superseded / valid_at to opt in); built-in web memory browser with force-directed graph + timeline; Obsidian bridge and non-destructive vault seed; cold-boot race fix (daemon binds its port at process start); neuromcp-doctor and neuromcp-connect find the daemon on its configured port; switching embedding provider no longer bricks an existing database (neuromcp-reembed, degraded mode); npm releases via Trusted Publishing (no tokens); graph overview no longer polluted by the system's own contradiction plumbing.
• v0.28 — neuromcp-init one-command setup (detects clients, writes configs with backups); neuromcp-doctor rebuilt as real triage; auto-generated docs/TOOLS.md + docs/BENCHMARK.md; examples/ client configs; store-time namespace pushdown; entity prefix-merge now needs evidence.
• v0.27 — security release: CWE-22 path-traversal fix in wiki_ingest, MCP-spec Origin validation on the daemon, neuromcp-connect boot-race-safe Claude Desktop bridge, runtime health-check hook.
• v0.26 — recall_answer (deterministic extractive answers with citations + gap-analysis, no LLM on the read path), optional local cross-encoder reranker (ships default-off after an honest A/B), recall-quality correctness sweep.
• v0.20–0.25 — shared HTTP daemon, session isolation, critic hook, entity dedup canonicalization, distractor benchmark hardening.

Explain mode

Every search_memory result includes an explain field so you can audit what the system remembers and why it surfaced:

{
  "explain": {
    "source_trust": { "level": "high", "reason": "Directly provided by user" },
    "temporal_validity": { "currently_valid": true, "superseded_by": null },
    "contradictions": [{ "memory_id": "abc", "content_preview": "...", "resolution": "coexist" }],
    "claims": [{ "subject": "neuromcp", "predicate": "version", "object": "0.26.0" }],
    "confidence": { "retrieval_score": 0.016, "source_trust_score": 1.0, "overall": 0.85 }
  }
}

Contradiction resolution is three-tier, and the two tiers that touch the graph require claim-level evidence — same subject, mutually-exclusive predicate (is, uses, runs, … see src/config/predicate-classes.json), different object: supersede (evidence + heuristic score > 0.5: old memory invalidated), coexist (evidence + score in (0.3, 0.5]: both kept, linked via a contradicts graph edge), flag (heuristic signal only — numeric difference, negation words — reported in the store_memory result for review, never materialised in the graph). Keyword heuristics alone can never invalidate a memory or assert a contradiction: a false one is a hallucination vector for every LLM downstream.

Troubleshooting

npx neuromcp-doctor check          # human-readable report
npx neuromcp-doctor check --json   # same checks, machine-readable

One run checks: Node version, native modules actually loadable (better-sqlite3, sqlite-vec), database openable, shared daemon /health, Ollama reachable + nomic-embed-text pulled, ONNX fallback model present, and whether the vector index in your database still matches a provider this machine can reach. Exit codes: 0 healthy, 1 degraded (e.g. no Ollama — ONNX fallback active), 2 broken. Start every bug report with its output.

"Embedding dimension mismatch" / vector search is off

The vector index is a vec0 table with a fixed width: 384 for the ONNX fallback, 768 for Ollama's nomic-embed-text. If the provider changes (you installed Ollama after running on the fallback, or Ollama was down and the auto cascade dropped to ONNX), the new provider cannot write into the old index.

Since 0.29.3 this no longer stops the server. It:

1. prefers the provider that matches the existing index,
2. retries a briefly-unavailable provider with backoff,
3. and otherwise starts in degraded mode — full-text search keeps working, vector search is disabled, and new memories are stored without embeddings and queued for backfill_embeddings. Nothing is deleted and the index is never rebuilt behind your back.

neuromcp-doctor check names the state. To get vector search back, pick one:

# a) restore the provider that matches the index
ollama serve && ollama pull nomic-embed-text      # 768-dim index
# or force the fallback for a 384-dim index:
export NEUROMCP_EMBEDDING_PROVIDER=onnx

# b) rebuild the index for the provider you want (works on a COPY first)
npx neuromcp-reembed                 # dry run: rebuild a copy, report, discard
npx neuromcp-reembed --keep-copy     # same, but keep the copy to inspect
npx neuromcp-reembed --apply         # swap it in; original kept as *.pre-reembed-<ts>

Stop every client before --apply. Rollback is a file move back.

Installing without Xcode / with install scripts disabled

better-sqlite3 fetches its prebuilt binary from its install script. With npm install --ignore-scripts (or a hardened .npmrc, or pnpm's build allowlist) the package installs but has no native binding and nothing works. npm 11's own default is ignore-scripts=false, so a plain install is fine:

npm install -g neuromcp                      # normal path, scripts run
npm install -g neuromcp --ignore-scripts=false   # if your .npmrc disables them

If you already installed with scripts off, reinstall as above (a bare npm rebuild in a global prefix does not reliably produce the binding).

The ~33 MB ONNX fallback model no longer depends on the postinstall at all: it is downloaded lazily on first use into ~/.neuromcp/models (user-writable even when the package lives in a root-owned global prefix). Fetch it ahead of time with npx neuromcp-download-model, or forbid the download entirely with NEUROMCP_DISABLE_MODEL_DOWNLOAD=1.

Comparison

License

AGPL-3.0 for the engine in src/. MIT for bin/, templates/, scripts/, docs/, and examples/ (carve-out — see LICENSE-EXAMPLES).

License FAQ

Can I use neuromcp commercially? Yes. Running neuromcp as part of your own application, on your own infrastructure, is unrestricted. AGPL only imposes obligations if you modify the engine code AND distribute or host it as a network service.

Can I install neuromcp from npm in my closed-source product? Yes. Using the published binary as a dependency does not trigger AGPL contagion.

What if I host neuromcp as a SaaS? Then AGPL §13 applies: you must make the source code (including your modifications) available to your users. This is the explicit anti-fork clause we chose for the engine — it stops well-funded competitors from taking the code, putting it behind a login, and shipping it as their own product.

Can I copy a CLI script or template? Yes. Everything in bin/, templates/, scripts/, docs/, and examples/ is dual-licensed AGPL-3.0 OR MIT. Pick MIT in your downstream project.

Need different terms for the engine? Commercial dual-license is available — contact the maintainer.