Human-first notes (Notion / Obsidian)
Page-centric and great for people; AI access depends on connectors or plugins, and structure is a human convention.
LOCAL-FIRST · MCP-NATIVE · OPEN SOURCE
A queryable, traceable knowledge base for your AI agents
hyseed turns team decisions, lessons, APIs, and runbooks into structured facts that MCP clients like Claude Code, Cursor, and Cline can use directly — with data staying on your machine.
Windows x64 · VS Code 1.101+ · Beta · MIT
$ hyseed search "sqlite fts5 decision"
dec/use-sqlite-fts5decision · accepted
→ kb_get(..., section="Architecture")target section only
context70–90% less on long docs
Section-level retrieval keeps whole documents out of the context window.
Notes, vector stores, and AI memory each miss a piece
They solve how humans write, how to find similar text, and how to remember a conversation — but not how AI reliably uses facts your team has reviewed.
Vector stores / RAG (Chroma / LanceDB, etc.)
Strong at similarity search, but no schema, versioning, or team collaboration; you build the retrieval pipeline.
AI memory (Mem0 / Zep, etc.)
Per-agent conversational memory, mostly auto-extracted; not designed for team review or long-term governance.
hyseed fills the gap: structured, team-reviewed facts that AI can use through MCP.
Designed to be used by AI
Not a chat box bolted onto a wiki — retrieval, structure, versioning, and collaboration live in the protocol and the data model.
MCP-native
CRUD, search, graph, versioning, and scheduling are exposed through MCP. Claude Code, Cursor, Cline, and other clients can connect directly. When the tool surface is too large, thin mode exposes only three read-only tools.
Typed schema, not free text
Nine built-in document types: project, decision, lesson, glossary, person, FAQ, API, runbook, and release. Typed metadata and stable IDs let AI extract exact facts.
Section-level retrieval
kb_get can return a single heading section; when the section is missing, it returns the table of contents. Long-document context drops by 70–90%.
Versioning, audit, rollback
Every change has a version number, with field-level diffs, restore to any version, soft-delete recovery, and an audit log. AI writes become safe to trust.
Typed knowledge graph
Ten standard relation types, automatic backlinks, multi-hop traversal, impact analysis, and decision chains. Change a decision and follow the graph to affected docs.
Git-native collaboration
Knowledge lives as Markdown in Git; tags, aliases, links, and metadata are three-way merged. PR review, branches, and Code Owners keep working.
Side by side
Compared by typical product form; check each product's current capabilities for details.
| Dimension | Human-first notes (Notion / Obsidian) | Vector stores + RAG | AI memory | hyseed |
|---|---|---|---|---|
| Design center | People and pages | Models and vectors | One agent's conversation | AI agents + people, one shared body of knowledge |
| AI access | Web / editor; connectors or plugins | SDK + custom retrieval pipeline | SDK / API tied to a framework | MCP-native + CLI + admin console + VS Code extension |
| Structure | Free-form, by convention | Vectors only, no schema | Auto-extracted memory entries | 9 document types + typed metadata + stable IDs |
| Retrieval | Keywords / plugins | Mostly vector similarity | Semantic recall | FTS5 BM25 / fuzzy / semantic / RRF hybrid |
| Context cost | Whole page | Chunks, still need trimming | Injected memory, uncontrolled granularity | Section-level; 70–90% less on long docs |
| Versioning & audit | Page/file history; audit depends on plan or plugin | Usually none | Not the focus | Versions, field-level diffs, restore, soft delete, audit |
| Team collaboration | Notion: real-time; Obsidian: Git/sync | None | Weak, platform-bound | Git-native + three-way merge + PR review |
| Data & deployment | Notion: cloud; Obsidian: local files | Local or cloud, self-managed | Self-hosted open source or cloud | Single-file SQLite, fully local, no cloud, no telemetry, MIT |
Comparison is by typical product form; hyseed capabilities and numbers come from the project README.
Three steps into your workflow
No separate Python runtime: the VS Code extension bundles hyseed.exe.
Install the extension
Install hyseed from the VS Code Marketplace. Windows x64, VS Code 1.101+.
code --install-extension limccn.kbmcp-vscodeOpen the admin console
Run hyseed: Open admin console from the Command Palette. The extension starts a local server bound to 127.0.0.1.
Add or search knowledge
Create a document, import Markdown, or ask your MCP client to search. Every change is versioned.
hyseed search "sqlite fts5"The first run creates a local vault; the default data root is %USERPROFILE%\.local\share\hyseed.
Usable today, with the boundaries stated
hyseed is in Beta; the VS Code extension's current public release is 0.1.1. The core storage and API have been stable since 0.5.0, so it is reasonable to start with a small team. Semantic search requires the optional sqlite-vec extension; lexical, fuzzy and RRF hybrid search work by default.
Public release
Marketplace 0.1.1
Platform
Windows x64
License
MIT
Data
Local SQLite · no telemetry
CLI distribution
Bundled in the extension; PyPI pending
Bring team knowledge into your AI workflow
Install hyseed from the VS Code Marketplace and get a local, traceable, collaborative knowledge base in minutes.
Windows x64 · no separate runtime install · MIT