novyxlabs/novyx-mcp

Governed actions, shared memory, audit, and rollback for MCP clients such as
**Claude Desktop**, **Cursor**, and **Claude Code**. Works locally with zero
config (SQLite) for memory and policy-evaluation workflows, or connects to
Novyx Cloud for policy-as-code, approval workflows, governance dashboards,
Runtime v2, replay, and eval baselines. The threat-intelligence and
defense-tracking tools were removed in 2.9.0 and never enforced anything;
upgrade to 2.9.0 or later.
What To Try First
Use Novyx MCP when coding agents need reviewable shared memory instead of
untracked context sprawl:
That workflow keeps agent memory reviewable before it becomes permanent. The
same server also exposes governed-action tools (action_submit,
list_pending, approve_action, explain_action) when you connect it to
Novyx Cloud.
Install
Configuration
Claude Desktop
Add to your claude_desktop_config.json:
Cursor
Add to your Cursor MCP settings:
Claude Code
Set the NOVYX_API_KEY environment variable before starting Claude Code. Omit it to use local mode (SQLite, zero config).
CLAUDE.md Integration
After installing and configuring the MCP server above, add this to your project's CLAUDE.md so Claude Code uses Novyx automatically:
This turns isolated Claude Code sessions into a coordinated team — each session stores what it learned and checks what other sessions have done before starting work.
Canonical Workflow: Draft, Review, Merge
This is the highest-signal Novyx workflow for coding agents:
1. The agent learns something important, but uses draft_memory(..., branch_id="feature-x") instead of writing directly.
2. Review the whole branch with memory_branch("feature-x").
3. Use draft_diff when one draft needs a closer look.
4. Merge the whole branch with merge_branch("feature-x"), or reject it with reject_branch("feature-x").
Example:
This keeps agent memory reviewable instead of letting every session write directly into permanent state.
Available Tools
Memory-only tools work in local SQLite mode (zero config, no API key).
Cloud-only governance and runtime tools require a Novyx API key.
Core Memory (20 tools)
Store, recall, supersede, and audit individual memories.
| Tool | Description |
|------|-------------|
| remember | Store a memory observation with tags, importance, context, TTL |
| recall | Semantic search using natural language |
| list_memories | List stored memories with optional tag filtering |
| memory_stats | Total count, average importance, conflict count |
| memory_health | Health score, stale memory count, contradiction count |
| forget | Delete a memory by UUID |
| supersede | Replace a memory with a new version, preserving history |
| link_memories | Create a directed link between two memories |
| unlink | Remove a link between memories |
| get_links | Retrieve all links for a memory |
| add_triple | Add a knowledge graph triple (subject → predicate → object) |
| query_triples | Query knowledge graph triples with filters |
| delete_triple | Remove a knowledge graph triple |
| get_entity / list_entities / delete_entity | Knowledge graph entity CRUD |
| graph_edges | List edges between memories or entities |
| audit | Get the cryptographic audit trail |
| audit_export / audit_verify | Export and verify the audit chain |
Memory Drafts & Branches (8 tools)
Stage memory changes for review before committing.
| Tool | Description |
|------|-------------|
| draft_memory | Create a reviewable draft before writing to canonical memory |
| memory_drafts | List open, merged, or rejected drafts |
| draft_diff | Show field-level changes before merging a draft |
| merge_draft / reject_draft | Merge or reject an individual draft |
| memory_branch | Review a whole branch/session of drafts at once |