Skip to main content
Glama
Kilonova MCP
by MilnaOS

Kilonova MCP

Persistent knowledge base tools for Claude Code.
Claude remembers your projects, decisions, and patterns across sessions — stored on your machine, no cloud required.

Built by AIM Studio · Free · MIT License


The Problem

Every Claude Code session starts cold. You re-explain your project structure, re-describe decisions you made last week, re-state what's in flight. Context burns fast.

Related MCP server: xtage

The Solution: DOT + KB

Kilonova gives Claude a persistent knowledge base on your local machine. At session start, Claude loads your DOT (Document of Truth) — a compressed, structured reference doc with your project state, active tasks, decisions, and patterns. During the session, Claude writes new discoveries back to the KB. Next session, it's all there.

Session 1: Claude learns your architecture → kb_write saves the decision
Session 2: dot_load → Claude already knows. No recap needed.

Install

pip install kilonova-mcp

Or from source:

git clone https://github.com/MilnaOS/kilonova-mcp
cd kilonova-mcp
pip install -e .

Wire Up Claude Code

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "kilonova": {
      "command": "python",
      "args": ["-m", "kilonova_mcp"],
      "env": {
        "KILONOVA_KB_ROOT": "/path/to/your/kb"
      }
    }
  }
}

Copy CLAUDE.md.template to ~/.claude/CLAUDE.md (or append to your existing one).

Quick Start

1. Create your first KB topic:

In Claude Code, just start writing:

mcp__kilonova__kb_write(
  topic="claude_context",
  entity_type="project_state",
  name="my-project",
  data={
    "name": "my-project",
    "status": "active",
    "location": "/path/to/project",
    "summary": "What this project is",
    "next_action": "What to do next"
  }
)

2. Load it next session:

mcp__kilonova__dot_load(topic="claude_context")

3. Search it:

mcp__kilonova__kb_search(topic="claude_context", query="authentication decision")

The DOT Format

A DOT is a plain text file with two sections:

---TOC---
1:Projects|1.1:My_Project
2:Active_Tasks
3:Decisions

---CARDS---

## [1] PROJECTS
### [1.1] My Project
STATUS: active
NEXT: wire up the auth flow

The TOC lets Claude fetch only the section it needs. Cards hold the actual content.

See example_dot/ for a starter template.


Starter Schema: claude_context

Copy schemas/claude_context/ into your KB directory under <kb_root>/claude_context/schemas/:

Entity type

Use for

project_state

Current status, location, next action per project

decision

Architectural choices with rationale

pattern

Code conventions, gotchas, file locations

active_task

In-flight work across sessions

session_note

End-of-session summaries


Tools

Tool

Description

dot_load(topic)

Load full DOT document into context

kb_search(topic, query)

Search records by natural language query

kb_write(topic, entity_type, name, data)

Write/update a record (merges with existing)

kb_load(topic, entity_type, name)

Load one specific record

kb_topics()

List all KB topics and record counts

kb_schema(topic, entity_type?)

Show field schema for an entity type

corpus_status(topic)

Show KB size and record counts

kb_backup(dry?)

Mirror KB to OneDrive


Bring Your Own KB

Kilonova doesn't care what you store. Define your own schemas:

// kb/my_topic/schemas/component.json
{
  "name": {"type": "string", "description": "Component name"},
  "file": {"type": "string", "description": "Path to file"},
  "purpose": {"type": "string", "description": "What it does"},
  "dependencies": {"type": "array", "description": "What it depends on"}
}

Then write records to it and search them naturally.


Part of the Kilonova Ecosystem

Kilonova MCP is the free, standalone KB layer extracted from Milna OS — a full BYOK multi-model AI terminal. If you want the whole thing (parallel model legs, distillation engine, web ingestion, local+cloud hybrid inference), check out Milna OS.


MIT License · © AIM Studio

Available Tools

8 tools
corpus_statusA

Show ingestion progress for a topic — record counts, file size, last modified.

Inputs: topic (required).

dot_loadA

Load the full DOT (Document of Truth) for a KB topic. A DOT is a compressed, structured reference document with a TOC and cards. Use dot_load("claude_context") at session start to load your persistent working memory.

Inputs: topic (required).

kb_backupB

Mirror the KB folder to OneDrive. Use dry=true to preview without copying.

Inputs: dry, kb_only.

kb_loadC

Load a specific entity record from a KB by name.

Inputs: name (required), topic (required), entity_type (required).

kb_schemaA

Show field schema for an entity type in a KB topic. Call with no entity_type to list all schema types for the topic.

Inputs: topic (required), entity_type.

kb_topicsA

List all available KB topics and their entity type counts.

Inputs: no named parameters.

kb_writeA

Write or update an entity record in a KB. Use this mid-session to persist decisions, task state, patterns, or any structured note. Merges with existing record — only non-empty fields overwrite.

Inputs: data (required), name (required), topic (required), entity_type (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.6/5.0

    Scored across 8 tools

    Disambiguation4/5

    The core CRUD/search set (kb_search, kb_load, kb_write, kb_schema) is clearly differentiated by verb. The main soft overlap is between kb_topics (lists topics with entity type counts) and corpus_status (record counts, size, last modified), which could be confused by an agent wanting topic-level metrics; kb_load vs dot_load is also close but distinguished by entity record vs whole-document.

    Naming Consistency4/5

    Six of eight tools use a clean prefix_resource pattern (kb_topics, kb_search, kb_load, kb_write, kb_schema, kb_backup), and dot_load is readable in context. corpus_status breaks the naming family by dropping the kb_ prefix, which is a minor but noticeable deviation.

    Tool Count5/5

    Eight tools is well within the sweet spot and each maps to a distinct need: discovery, search, load, write, schema, full-document load, status, and backup. Nothing feels padded or redundant.

    Completeness4/5

    The surface covers discovery, search, read, write/update, schema introspection, bulk-document load, and backup, which is solid lifecycle coverage for a knowledge base. The visible gap is deletion/removal of entity records — kb_write can create and overwrite, but there is no way to delete an obsolete record.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      C
      maintenance
      Persistent project memory + architectural decisions + pre-execution safety hooks for Claude Code. Local-only storage, multi-repo workspace, automatic knowledge extraction via background auditor.
      8 npm
      14
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Provides Claude with persistent, semantic knowledge of a codebase across sessions by maintaining a local knowledge base of structural maps, compressed file summaries, and insights.
      7 npm
      1
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Provides persistent, searchable memory for Claude Code using local SQLite, semantic embeddings, and full-text search, enabling Claude to recall and retrieve context across sessions and projects without external services.
      23 npm
      4
      MIT
    • F
      license
      Not graded
      quality
      D
      maintenance
      Provides persistent semantic memory for Claude Code via local embeddings and six MCP tools, enabling context storage and retrieval across sessions without cloud dependencies.
      -