prunus

A shared knowledge store for AI coding sessions. Insights, decisions, and patterns accumulate in plain Markdown trees and are surfaced via MCP during future sessions across all your machines and tools.

How it works

Capturing knowledge — explicit, curated:

/prunus update "focus on the architectural decisions"
  └─ Session AI composes a summary document from the current session context
       └─ Sends document to prunus via update_notes MCP tool
            └─ Server LLM extracts distinct knowledge chunks from the document
                 └─ Dedup check against existing tree → save as Markdown → index

Using knowledge — automatic, per-prompt:

User sends a prompt
  └─ Hook/plugin queries prunus: GET /tree/{tree}/context?query=<prompt>
       └─ Relevant note summaries injected as context
            └─ Session AI calls read_note or search_notes via MCP for full content

Architecture

Component	Role
Deno HTTP server	MCP at /mcp, context search at /tree/{tree}/context, health endpoint
SQLite (default)	Metadata + FTS5 + JS-side cosine similarity — zero external dependencies
PostgreSQL + pgvector	Optional — HNSW index + TSVECTOR for larger deployments
OpenAI-compatible API	Embed model for search; chat model for note extraction (e.g. Ollama, LM Studio)
Tree directories	Source of truth — one subdirectory per named tree under grove.path

The database is derived from the grove and fully rebuildable by rescanning.

MCP Tools

Tool	Description
create_note	Create a new note at a given path
read_note	Read a note's full Markdown content by path or ID
update_note	Update an existing note's body and summary
delete_note	Delete a note; watcher handles link cleanup and index removal
list_notes	List all note paths in a tree
search_notes	Hybrid vector + FTS search across all notes in the tree
update_notes	Submit a prepared session summary document for LLM extraction
list_profiles	List all profiles + which are enabled for a tree
enable_profile	Enable a profile for a tree
disable_profile	Disable a profile for a tree

Setup

Server

cp settings-dev.toml settings.toml
# edit settings.toml: set grove.path, llm.hostname, llm.chat.model, llm.embed.model
deno task start

For PostgreSQL, set db.type = "postgres" and fill in the [db.postgres] section.

Clients

Supports Claude Code, Gemini CLI, OpenCode, and Qwen-Code. The installer detects which are installed:

deno run --allow-all http://prunus-host:9100/cli/install

Installs per tool: MCP server registration, per-prompt context hook, /prunus slash command.

Then in any project:

/prunus init   # creates .prunus/settings.json, sets tree and project

See CLIENTS.md for per-tool details, hook formats, and settings.

Server Settings (settings.toml)

[srv]
hostname = "0.0.0.0"
port = 9100

[srv.auth]
token = ""

[log]
level = "info"   # debug | info | warn | error

[db]
type = "sqlite"   # sqlite | postgres

[db.sqlite]
path = "/path/to/data"

[llm]
hostname = "localhost"
port = 1234

[llm.chat]
model = "..."   # chat model for note extraction (e.g. qwen2.5, llama3.2)

[llm.embed]
model = "..."          # embedding model (e.g. nomic-embed-text, bge-m3)
dimension = 1024       # must match model

[grove]
path = "/path/to/trees"   # required; each subdirectory is a named tree

Client Settings

• ~/.prunus/settings.json — user-level: url, token, default tree
• .prunus/settings.json — project-level (walk-up from cwd): tree, enabled, project

Note Format

Notes are plain Markdown with YAML frontmatter managed by the server:

---
id: <uuid>
summary: '2-3 sentence summary used for search and embeddings'
created: 2026-03-23T10:00:00Z
updated: 2026-03-23T10:00:00Z
projects:
  - my-app
tags:
  - typescript
---

Note content here...

Trees are Obsidian-compatible — [[wikilinks]] between notes are stored as link relationships in the index.

Development

deno task check   # type check
deno task format  # format
deno task lint    # lint
deno task start   # run with hot reload
deno task test    # tests