Novyx MCP - Desktop Extension

Desktop Extension (.mcpb) for Claude Desktop. This repo is the one-click installer shell for the Python novyx-mcp engine.

Current desktop extension: 1.0.6 Current MCP engine target: novyx-mcp 2.8.0

Features

• Reviewable Memory - Draft, review, diff, merge, or reject agent memory before it becomes permanent.

• Governed Actions - Submit, approve, deny, and explain agent actions when connected to Novyx Cloud.

• Time-Travel Rollback - Preview and restore memory to a known-good point.

• Audit Trails - Inspect cryptographic proof of memory operations.

• Knowledge Graph - Link memories and query subject-predicate-object triples.

• Context Spaces - Keep project and team memory separated.

• Local-First - Works instantly with SQLite and no API key.

• Cloud Upgrade - Optional hosted governance, approvals, replay, Runtime v2, threat intelligence, and team workflows.

Related MCP server: Engram

Installation

From the Anthropic Directory (recommended):

Install directly from Claude Desktop -> Settings -> Extensions.

Manual install:

1. Download the latest .mcpb file from Releases

2. Double-click the file, or drag it into Claude Desktop

Prerequisites: Python 3.10+ must be installed. The extension first tries uvx --upgrade novyx-mcp, then falls back to python3 -m novyx_mcp, python -m novyx_mcp, and finally the novyx-mcp executable.

Configuration

No configuration required for local mode. The extension works out of the box with a local SQLite database.

Optional - Cloud mode:

When prompted during installation, enter your Novyx API key. Get a free key at novyxlabs.com (5,000 memories, no credit card).

Cloud mode enables:

• Cross-device memory sync

• Hosted governance and approval workflows

• Runtime v2 agents, missions, capabilities, checkpoints, and interventions

• Replay, cortex, and threat-intelligence workflows

• Team spaces when hosted sharing is available

share_space is currently visible but disabled in novyx-mcp 2.8.0 until hosted invitation redemption is available. It returns a clear local-only response instead of issuing unusable tokens.

What To Try First

Use the desktop extension when Claude needs reviewable memory instead of untracked context:

draft_memory(
  observation="Deploys fail if REDIS_URL is unset in staging",
  tags=["ops", "staging"],
  importance=8,
  branch_id="staging-fixes"
)

memory_branch("staging-fixes")
draft_diff("drf_abc123")
merge_branch("staging-fixes")

That flow keeps memory changes inspectable before they become permanent.

Usage Examples

Example 1: Store and recall memories

User prompt:

Remember that the project deadline is March 15th and we're using React with TypeScript.

What happens: Claude calls the remember tool to store tagged memories. Later:

What tech stack are we using for this project?

What happens: Claude calls recall with a semantic search, finds the stored memory about React + TypeScript, and answers accurately.

Example 2: Roll back a mistake

User prompt:

I accidentally told you the deadline was March 15th - it's actually April 1st. Roll back the wrong memory and fix it.

What happens: Claude calls rollback_preview and rollback to undo the incorrect memory, then remember to store the corrected date. The audit trail shows the full history: original store -> rollback -> corrected store.

Example 3: Build a knowledge graph

User prompt:

Track these relationships: Alice manages the frontend team, Bob manages the backend team, and both teams report to Carol.

What happens: Claude calls add_triple to create knowledge graph entries:

• Alice -> manages -> frontend team

• Bob -> manages -> backend team

• frontend team, backend team -> reports_to -> Carol

Later, asking "Who does the frontend team report to?" triggers a query_triples call that returns the answer.

Example 4: Isolated project contexts

User prompt:

Create a separate memory space for my side project so it doesn't mix with work memories.

What happens: Claude calls create_space to create an isolated context. Memories stored in that space are scoped and do not appear in general searches.

Privacy Policy

Novyx MCP operates in two modes:

Local mode (default): All data is stored locally in a SQLite database at ~/.novyx/local.db. No data is sent to any external server. No analytics or telemetry.

Cloud mode (opt-in): When you provide an API key, memories are sent to the Novyx API (novyx-ram-api.fly.dev) for storage and sync. Data is encrypted in transit (TLS) and at rest. We do not share your data with third parties. See our full privacy policy at novyxlabs.com/privacy.

You can switch between modes at any time by adding or removing your API key.

Data retention: Local data persists until you delete it. Cloud data is retained until you delete it or close your account. Audit trails are immutable by design.

For privacy questions, contact blake@novyxlabs.com.

Support

• Issues: github.com/novyxlabs/novyx-mcp-desktop/issues

• Documentation: docs.novyxlabs.com

• Email: blake@novyxlabs.com

How It Works

This Desktop Extension is a thin Node.js wrapper that spawns the Python novyx-mcp server as a child process. The Node.js layer handles process lifecycle; the Python server handles all MCP logic.

Launch order:

1. uvx --upgrade novyx-mcp (fastest; no install needed)

2. python3 -m novyx_mcp (if pip installed)

3. python -m novyx_mcp

4. novyx-mcp (if pipx installed)

Release

Build the desktop extension bundle:

npm run validate
npm run package

The package script writes dist/novyx-mcp-desktop-<version>.mcpb. The bundle contains manifest.json, server/index.js, icons, license, README, and package metadata with manifest.json at the archive root.

License

MIT